title
stringlengths 14
154
| paper_url
stringlengths 42
42
| authors
listlengths 1
21
| type
stringclasses 3
values | abstract
stringlengths 413
2.52k
| keywords
stringlengths 4
397
| TL;DR
stringlengths 5
250
⌀ | submission_number
int64 2
14.3k
| arxiv_id
stringlengths 10
10
⌀ | embedding
listlengths 384
384
| github_url
stringlengths 0
126
| github_stars
int64 0
55k
| num_models
int64 0
82
| num_datasets
int64 0
7
| num_spaces
int64 0
100
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
HELMET: How to Evaluate Long-context Models Effectively and Thoroughly
|
https://openreview.net/forum?id=293V3bJbmE
|
[
"Howard Yen",
"Tianyu Gao",
"Minmin Hou",
"Ke Ding",
"Daniel Fleischer",
"Peter Izsak",
"Moshe Wasserblat",
"Danqi Chen"
] |
Poster
|
Many benchmarks exist for evaluating long-context language models (LCLMs), yet developers often rely on synthetic tasks such as needle-in-a-haystack (NIAH) or an arbitrary subset of tasks. However, it remains unclear whether these benchmarks reflect the diverse downstream applications of LCLMs, and such inconsistencies further complicate model comparison. We investigate the underlying reasons behind these practices and find that existing benchmarks often provide noisy signals due to limited coverage of applications, insufficient context lengths, unreliable metrics, and incompatibility with base models. In this work, we introduce HELMET (How to Evaluate Long-context Models Effectively and Thoroughly), a comprehensive benchmark encompassing seven diverse, application-centric categories. We also address several issues in previous benchmarks by adding controllable lengths up to 128K tokens, model-based evaluation for reliable metrics, and few-shot prompting for robustly evaluating base models. Consequently, we demonstrate that HELMET offers more reliable and consistent rankings of frontier LCLMs. Through a comprehensive study of 59 LCLMs, we find that (1) synthetic tasks like NIAH do not reliably predict downstream performance; (2) the diverse categories in HELMET exhibit distinct trends and low correlations with each other; and (3) while most LCLMs achieve perfect NIAH scores, open-source models significantly lag behind closed ones when tasks require full-context reasoning or following complex instructions---the gap widens as length increases. Finally, we recommend using our RAG tasks for fast model development, as they are easy to run and better predict other downstream performance; ultimately, we advocate for a holistic evaluation across diverse tasks.
|
long-context language models, benchmarking
| null | 12,024 |
2410.02694
|
[
-0.036936331540346146,
-0.0749112069606781,
0.024740343913435936,
0.011985985562205315,
0.11203360557556152,
-0.0415901355445385,
-0.011810850352048874,
-0.007271584123373032,
-0.006890006363391876,
-0.04173323139548302,
-0.09292416274547577,
-0.14879196882247925,
0.05579070374369621,
0.019035451114177704,
0.0491090752184391,
-0.05120355263352394,
0.08101650327444077,
0.02583071030676365,
-0.0694153904914856,
-0.02869843877851963,
-0.010060550644993782,
0.06352651864290237,
0.02310202829539776,
0.008197149261832237,
0.03395647183060646,
-0.02443462237715721,
-0.07038199156522751,
-0.08300904184579849,
0.006147921551018953,
0.030245834961533546,
-0.0010653937933966517,
0.07247718423604965,
0.014909479767084122,
0.055010102689266205,
-0.03777081519365311,
0.04462535306811333,
-0.020108576864004135,
-0.027199044823646545,
0.0066032106988132,
-0.03496139124035835,
-0.06290421634912491,
-0.013340950943529606,
0.04280294105410576,
0.03976545110344887,
0.0645502507686615,
-0.055045224726200104,
0.014314360916614532,
-0.03816217556595802,
-0.025010546669363976,
-0.005002837628126144,
-0.07792104780673981,
-0.06150578334927559,
-0.0017295634606853127,
0.011015811003744602,
-0.010782627388834953,
0.022077729925513268,
0.010866955853998661,
0.053297899663448334,
0.011495323851704597,
0.03435725346207619,
-0.06287633627653122,
-0.08383872359991074,
-0.10581525415182114,
0.02584044635295868,
-0.04409611225128174,
-0.013036050833761692,
-0.004659673199057579,
0.006191634573042393,
-0.011132767423987389,
0.024647070094943047,
-0.06847066432237625,
0.05455535650253296,
-0.038654498755931854,
0.08825018256902695,
0.037141963839530945,
0.03377874940633774,
0.03612535819411278,
-0.01842847838997841,
0.035818420350551605,
-0.12538646161556244,
0.010124041698873043,
0.010895375162363052,
0.026024265214800835,
0.0009379524854011834,
0.1452196091413498,
0.02095167711377144,
0.06890372186899185,
-0.0005922227283008397,
0.018987195566296577,
0.025575056672096252,
-0.0009950381936505437,
-0.10397140681743622,
0.08525513112545013,
-0.009621607139706612,
0.06056731566786766,
0.03787434473633766,
0.054532602429389954,
-0.04296857863664627,
-0.033608052879571915,
0.09891091287136078,
-0.02023608237504959,
0.025593828409910202,
-0.03230499103665352,
-0.09074926376342773,
-0.012167319655418396,
0.007265094667673111,
0.02592943049967289,
0.04383526369929314,
-0.03994332253932953,
-0.041810616850852966,
0.012013321742415428,
0.0615137554705143,
-0.00812207255512476,
-0.012056002393364906,
0.05058646947145462,
-0.04143548756837845,
0.03472689539194107,
-0.10089962184429169,
0.035990383476018906,
0.11715991050004959,
-0.02279486134648323,
0.03228042274713516,
0.005511632189154625,
-0.023818567395210266,
-0.03940916806459427,
0.024023374542593956,
-0.039201006293296814,
3.594749886890867e-33,
0.03883488476276398,
0.06458407640457153,
0.0244330782443285,
0.06434599310159683,
0.018550576642155647,
-0.017424583435058594,
-0.015087808482348919,
0.047705408185720444,
-0.029336269944906235,
0.011229309253394604,
0.05798366665840149,
0.12163732945919037,
-0.0746689960360527,
0.02117992378771305,
0.09521286934614182,
0.028323877602815628,
-0.04520990699529648,
0.07229261100292206,
-0.09240175783634186,
-0.015887383371591568,
0.04925978556275368,
-0.0816512405872345,
-0.034269966185092926,
-0.0490504652261734,
0.07925453782081604,
0.05457654595375061,
0.03949776291847229,
-0.07159316539764404,
-0.029544973745942116,
0.05249352008104324,
-0.11055266857147217,
-0.018748382106423378,
0.031481191515922546,
0.08018584549427032,
0.024290457367897034,
-0.036027323454618454,
-0.021393217146396637,
-0.04562615230679512,
0.001109324162825942,
-0.045606013387441635,
-0.06467876583337784,
0.01831868104636669,
-0.010150939226150513,
-0.003859704127535224,
-0.06441951543092728,
-0.09723792225122452,
-0.03654113784432411,
-0.01055235043168068,
-0.026323210448026657,
-0.021617304533720016,
0.11060693860054016,
0.06290408968925476,
-0.008845040574669838,
0.008197234943509102,
-0.030803220346570015,
-0.01631677709519863,
0.0264813881367445,
-0.017251674085855484,
0.003959828056395054,
0.12135372310876846,
0.04021547734737396,
0.015244822017848492,
-0.03995656222105026,
0.07477101683616638,
0.04679787904024124,
0.033302679657936096,
-0.054907772690057755,
0.0644940659403801,
0.03192434832453728,
-0.03319421783089638,
-0.04462774842977524,
-0.011583574116230011,
0.014301212504506111,
0.016942784190177917,
0.04356914758682251,
-0.03622864559292793,
0.014459606260061264,
-0.0783318281173706,
-0.030450420454144478,
0.018763771280646324,
0.038142189383506775,
0.04216207191348076,
-0.0014513294445350766,
-0.04772014543414116,
-0.01518578827381134,
-0.11327473819255829,
-0.014985863119363785,
-0.06721867620944977,
-0.09188379347324371,
-0.07071062177419662,
-0.01159295067191124,
-0.012182669714093208,
-0.004025912377983332,
-0.041361864656209946,
-0.03728634491562843,
-5.0555796956699255e-33,
-0.005621300544589758,
-0.05821850895881653,
-0.03242398798465729,
0.056586749851703644,
-0.045656342059373856,
-0.10750188678503036,
0.028766032308340073,
0.008862346410751343,
-0.04895629733800888,
-0.05033646523952484,
-0.029675479978322983,
-0.055157072842121124,
0.05588124319911003,
0.04984385520219803,
0.0754762664437294,
-0.050486642867326736,
-0.015579105354845524,
-0.057864222675561905,
0.007438736502081156,
0.06453995406627655,
0.055376771837472916,
0.046420324593782425,
-0.09191502630710602,
0.008052694611251354,
-0.05163691192865372,
-0.022197358310222626,
-0.15593110024929047,
0.010035007260739803,
0.02647472731769085,
-0.024559589102864265,
-0.07327187061309814,
0.03108595311641693,
-0.011124685406684875,
-0.050890274345874786,
-0.03834740072488785,
0.026964835822582245,
0.019391844049096107,
-0.016513507813215256,
-0.03385591879487038,
0.09252873808145523,
0.11355520784854889,
-0.033692799508571625,
-0.029756689444184303,
-0.02103325165808201,
0.016400275751948357,
0.04065129905939102,
-0.09799547493457794,
-0.002645876258611679,
0.02205914631485939,
-0.03877075016498566,
-0.006624686066061258,
0.028149161487817764,
-0.03995649144053459,
0.011134836822748184,
-0.012174731120467186,
-0.054442983120679855,
-0.07972802966833115,
-0.07627096027135849,
0.025135038420557976,
-0.05942514166235924,
-0.049653444439172745,
0.01638135313987732,
0.025626564398407936,
0.023226480931043625,
0.06805470585823059,
-0.04509901627898216,
0.011613310314714909,
-0.014468907378613949,
-0.014475954696536064,
-0.04353311285376549,
-0.009625046513974667,
-0.03731704130768776,
-0.0839775949716568,
0.02310793288052082,
-0.07116525620222092,
0.03687068074941635,
0.02765486016869545,
-0.10328804701566696,
0.023050278425216675,
-0.014268630184233189,
0.023102283477783203,
-0.004434617236256599,
0.05455757677555084,
0.0019374999683350325,
0.031556159257888794,
0.11884275823831558,
0.019296567887067795,
0.07475510984659195,
0.012158501893281937,
0.030479110777378082,
-0.01936756819486618,
0.01926228031516075,
-0.04776936024427414,
0.0939742848277092,
-0.03689953312277794,
-4.7336026653965746e-8,
-0.037070535123348236,
0.09792006015777588,
-0.027049873024225235,
0.023508500307798386,
0.003350890474393964,
0.030931809917092323,
-0.0595942921936512,
0.06539987772703171,
0.05446314439177513,
0.05116535350680351,
-0.01795589178800583,
-0.008050830103456974,
-0.06986789405345917,
-0.005665527191013098,
-0.006403084844350815,
0.06335926800966263,
-0.004532753024250269,
0.05704297497868538,
-0.020284662023186684,
-0.011293557472527027,
0.04985138401389122,
0.122101329267025,
-0.010332954116165638,
-0.0021074824035167694,
-0.006756457965821028,
-0.07831856608390808,
-0.015901312232017517,
0.13623079657554626,
0.10363728553056717,
-0.030254172161221504,
-0.0125200767070055,
0.04266560077667236,
-0.06057988107204437,
-0.07128498703241348,
-0.041657403111457825,
0.053111981600522995,
0.027573030441999435,
-0.02418295294046402,
0.028956236317753792,
0.019877927377820015,
0.08301801234483719,
0.05679936707019806,
-0.0808897614479065,
0.033014051616191864,
0.025265183299779892,
0.033561453223228455,
0.018771206960082054,
-0.06649620831012726,
-0.018919067457318306,
-0.07050596922636032,
0.008562147617340088,
-0.03418973460793495,
-0.05881381034851074,
0.0406133197247982,
0.04212912917137146,
0.06192021444439888,
-0.043891675770282745,
-0.07134126126766205,
0.0002790137368720025,
0.04369405284523964,
-0.01829654537141323,
0.025085121393203735,
-0.015783941373229027,
0.07959452271461487
] |
https://github.com/princeton-nlp/helmet
| 140 | 0 | 0 | 0 |
Uncovering Latent Memories in Large Language Models
|
https://openreview.net/forum?id=KSBx6FBZpE
|
[
"Sunny Duan",
"Mikail Khona",
"Abhiram Iyer",
"Rylan Schaeffer",
"Ila R Fiete"
] |
Poster
|
Frontier AI systems are making transformative impacts across society, but such benefits are not without costs: models trained on web-scale datasets containing personal and private data raise profound concerns about data privacy and security. Language models are trained on extensive corpora including potentially sensitive or proprietary information, and the risk of data leakage, where the model response reveals pieces of such information, remains inadequately understood. Prior work has investigated that sequence complexity and the number of repetitions are the primary drivers of memorization. In this work, we examine the most vulnerable class of data: highly complex sequences that are presented only once during training. These sequences often contain the most sensitive information and pose considerable risk if memorized. By analyzing the progression of memorization for these sequences throughout training, we uncover a striking observation: many memorized sequences persist in the model's memory, exhibiting resistance to catastrophic forgetting even after just one encounter. Surprisingly, these sequences may not appear memorized immediately after their first exposure but can later be “uncovered” during training, even in the absence of subsequent exposures - a phenomenon we call "latent memorization." Latent memorization presents a serious challenge for data privacy, as sequences that seem hidden at the final checkpoint of a model may still be easily recoverable. We demonstrate how these hidden sequences can be revealed through random weight perturbations, and we introduce a diagnostic test based on cross-entropy loss to accurately identify latent memorized sequences.
|
Large Language Models, Memorization, Empirical Study, Data Leakage, Privacy, LLMs, Dynamics, Interpretability, Mechanistic
|
We study "latent memorization" in AI models, showing that complex sequences can persist and be revealed later, even after a single encounter, and study how these latent memories can be recovered.
| 12,019 | null |
[
-0.04571056365966797,
-0.09113352000713348,
-0.04505452886223793,
0.06455110758543015,
0.07851256430149078,
0.10443975776433945,
-0.013829268515110016,
-0.04070805758237839,
0.09165024012327194,
-0.032368253916502,
0.008104881271719933,
0.0114928362891078,
0.05100022628903389,
-0.05347669497132301,
-0.019355259835720062,
-0.04566535726189613,
-0.028107475489377975,
0.030622228980064392,
-0.07513947039842606,
-0.09294092655181885,
0.026028655469417572,
0.07071027159690857,
0.057221438735723495,
0.03323550522327423,
0.009857160970568657,
-0.01099125761538744,
0.0412132702767849,
-0.05687737464904785,
-0.005783953238278627,
-0.038253895938396454,
0.05809012055397034,
0.042914073914289474,
0.01075349748134613,
0.07939387112855911,
-0.016991963610053062,
-0.018633918836712837,
-0.08835621178150177,
-0.02381836622953415,
-0.0038415889721363783,
-0.031714361160993576,
0.013556627556681633,
-0.03601368889212608,
-0.008253812789916992,
-0.0010404062923043966,
0.018365750089287758,
-0.02479349635541439,
0.02451178804039955,
-0.04344791918992996,
-0.004322534892708063,
-0.041573796421289444,
-0.0878356322646141,
0.0014001294039189816,
-0.0033791325986385345,
0.04217529296875,
-0.06809946894645691,
-0.07627097517251968,
0.03839594125747681,
0.0532991997897625,
-0.07385240495204926,
0.03999974578619003,
-0.061226993799209595,
-0.12352968752384186,
-0.02359919250011444,
-0.06028912588953972,
-0.039609652012586594,
0.020517095923423767,
0.03469133749604225,
0.04042151942849159,
0.004970296751707792,
0.06613338738679886,
-0.01052584033459425,
0.08965043723583221,
-0.07741270214319229,
0.03522549569606781,
0.04463769495487213,
0.053876206278800964,
-0.01053231954574585,
-0.0740414708852768,
0.0212863702327013,
-0.07753494381904602,
-0.013987833634018898,
-0.01381500344723463,
0.0757855623960495,
0.0425751693546772,
0.06266354769468307,
0.017803940922021866,
0.06531178951263428,
-0.010062755085527897,
0.027474237605929375,
-0.019087471067905426,
-0.006342073902487755,
-0.10251870006322861,
0.03552582487463951,
0.03127770125865936,
0.04157894477248192,
0.02951170690357685,
0.020734483376145363,
0.0055646286346018314,
-0.027469830587506294,
0.007088355720043182,
-0.010170431807637215,
0.07443390786647797,
0.0025456598959863186,
0.05669049918651581,
0.002229649806395173,
-0.0007831486873328686,
0.07566718012094498,
0.0031052776612341404,
0.027932707220315933,
0.007681392133235931,
0.02622312121093273,
0.04729108884930611,
-0.03604445978999138,
0.04295070096850395,
0.0604216493666172,
-0.08979051560163498,
-0.009130985476076603,
-0.042807143181562424,
-0.017375992611050606,
0.1192934587597847,
0.034536924213171005,
0.06619961559772491,
0.035530321300029755,
0.02967410534620285,
-0.0513324998319149,
-0.06143005192279816,
-0.049585260450839996,
3.364357402233514e-33,
0.03415634483098984,
0.015198292210698128,
-0.03090582601726055,
0.02663313038647175,
0.018183471634984016,
-0.02457728423178196,
-0.04181239381432533,
0.037754982709884644,
-0.012865563854575157,
0.04899157956242561,
0.02488137036561966,
0.07748604565858841,
-0.022768178954720497,
0.03596380352973938,
0.0033244439400732517,
0.06868277490139008,
-0.037345364689826965,
0.0506608821451664,
0.04887239262461662,
-0.019617855548858643,
0.10840781778097153,
-0.0035408204421401024,
0.04317271336913109,
-0.05156819149851799,
0.042213473469018936,
-0.03984533250331879,
0.04963643476366997,
-0.03939492255449295,
0.009329253807663918,
0.026792554184794426,
-0.1288832426071167,
0.04701147973537445,
0.03938485309481621,
0.031243756413459778,
0.061808113008737564,
0.012140674516558647,
0.042190372943878174,
-0.07111898809671402,
0.05957494676113129,
-0.019347207620739937,
0.006387518718838692,
0.04451590031385422,
0.029408810660243034,
0.0014441945822909474,
-0.08047989010810852,
-0.05922958627343178,
0.02930058166384697,
-0.03231292590498924,
-0.03211015835404396,
-0.04350069910287857,
-0.02475600130856037,
0.08447971940040588,
-0.04739778861403465,
-0.043423715978860855,
-0.07140593975782394,
0.0018714990001171827,
-0.01011597365140915,
-0.06788544356822968,
-0.033425383269786835,
0.04866131767630577,
-0.024978671222925186,
0.058996960520744324,
0.04455922171473503,
0.05115894228219986,
-0.006381714250892401,
-0.014551595784723759,
-0.048309896141290665,
0.056550685316324234,
0.08441036194562912,
-0.02585821971297264,
-0.01955224573612213,
-0.04182012751698494,
-0.07407835125923157,
-0.05878380686044693,
0.0395081602036953,
-0.02352217771112919,
0.09980377554893494,
-0.08401019126176834,
-0.012114745564758778,
-0.02069481834769249,
0.038442645221948624,
-0.09667044132947922,
0.03657963499426842,
0.021799681708216667,
-0.1278708428144455,
-0.045552000403404236,
0.04699482023715973,
-0.025670170783996582,
-0.042447324842214584,
0.04905756562948227,
-0.02258552610874176,
-0.052425816655159,
0.02583245001733303,
-0.018969442695379257,
-0.06227045878767967,
-5.440056714517033e-33,
-0.0996016412973404,
-0.06505291908979416,
-0.06054415926337242,
0.016632890328764915,
0.00954902172088623,
-0.07480836659669876,
-0.011819325387477875,
0.07178080826997757,
-0.04264219477772713,
-0.08662954717874527,
-0.0276181660592556,
-0.008978581987321377,
0.05886637419462204,
-0.00044751563109457493,
0.04412363842129707,
0.0018628875259310007,
0.06650390475988388,
0.08699730783700943,
-0.020875981077551842,
-0.017965899780392647,
0.024852503091096878,
0.05030530318617821,
-0.07800473272800446,
-0.0012105826754122972,
-0.016471954062581062,
0.007523606065660715,
-0.08122170716524124,
-0.005871060770004988,
0.035128023475408554,
-0.00408931216225028,
-0.031824562698602676,
0.03401581943035126,
-0.024353312328457832,
-0.03827104717493057,
-0.09691036492586136,
0.016422415152192116,
0.07180918753147125,
-0.006327023264020681,
-0.040213603526353836,
0.06090345233678818,
0.08770084381103516,
0.002666742540895939,
-0.10624964535236359,
-0.010982106439769268,
-0.022460779175162315,
-0.11476162821054459,
-0.08635632693767548,
0.05631478503346443,
0.05120975151658058,
0.03751207888126373,
0.01970393769443035,
0.027933502569794655,
-0.010599184781312943,
-0.024897007271647453,
-0.0503600537776947,
-0.0420573391020298,
0.04174317419528961,
-0.035213883966207504,
0.04392775893211365,
0.055583249777555466,
-0.06110932677984238,
0.04697895050048828,
-0.041416581720113754,
-0.07451991736888885,
0.04325944930315018,
-0.0856252908706665,
-0.03759528324007988,
0.02676316536962986,
0.016987869516015053,
-0.025350112468004227,
0.08342625200748444,
-0.05692531540989876,
-0.04053901508450508,
0.031082332134246826,
-0.05330214276909828,
-0.009781643748283386,
-0.04588789865374565,
-0.0028508782852441072,
-0.05144549533724785,
0.012935327365994453,
0.06598497927188873,
0.00704864040017128,
0.005148791708052158,
0.004716447554528713,
0.08124756067991257,
0.050664328038692474,
0.06122295558452606,
-0.003698561806231737,
-0.062392350286245346,
0.05157744511961937,
0.0010508089326322079,
-0.0571431964635849,
-0.1184549331665039,
0.061148300766944885,
-0.10546490550041199,
-4.3938154448142086e-8,
-0.04340178519487381,
0.00965744536370039,
0.027944117784500122,
0.03798452392220497,
0.10065300762653351,
-0.0904957503080368,
-0.06371670961380005,
0.08178123086690903,
-0.058335840702056885,
-0.06141868606209755,
0.0937592163681984,
0.031982190907001495,
-0.019469361752271652,
-0.023732516914606094,
0.0061496528796851635,
0.09855017811059952,
-0.0055480157025158405,
-0.05652124062180519,
-0.008222324773669243,
-0.028695737943053246,
0.0007763747125864029,
-0.03249518573284149,
-0.01980767957866192,
0.0029527281876653433,
0.025569578632712364,
0.015310470946133137,
0.0028599025681614876,
0.0982147678732872,
0.053612738847732544,
0.09400167316198349,
-0.0033430042676627636,
-0.0457603894174099,
-0.029586181044578552,
-0.030225742608308792,
-0.015355397015810013,
0.0838475301861763,
0.03169388696551323,
-0.08337768912315369,
0.00443642633035779,
-0.04459850490093231,
0.0395636223256588,
0.018097206950187683,
0.04576236009597778,
0.06517849117517471,
0.043349336832761765,
0.015332654118537903,
-0.04285300150513649,
-0.10543876886367798,
0.0629439428448677,
-0.08892737329006195,
-0.0487859733402729,
0.017114587128162384,
0.06101421266794205,
0.14563611149787903,
0.09562598168849945,
0.03848561644554138,
0.03368264064192772,
-0.01855040341615677,
0.03164423629641533,
0.04389423877000809,
0.058519016951322556,
-0.004451611544936895,
-0.047226034104824066,
-0.01182074286043644
] | 0 | 0 | 0 | 0 |
|
Episodic Memories Generation and Evaluation Benchmark for Large Language Models
|
https://openreview.net/forum?id=6ycX677p2l
|
[
"Alexis Huet",
"Zied Ben Houidi",
"Dario Rossi"
] |
Poster
|
Episodic memory -- the ability to recall specific events grounded in time and space -- is a cornerstone of human cognition, enabling not only coherent storytelling, but also planning and decision-making. Despite their remarkable capabilities, Large Language Models (LLMs) lack a robust mechanism for episodic memory: we argue that integrating episodic memory capabilities into LLM is essential for advancing AI towards human-like cognition, increasing their potential to reason consistently and ground their output in real-world episodic events, hence avoiding confabulations. To address this challenge, we introduce a comprehensive framework to model and evaluate LLM episodic memory capabilities. Drawing inspiration from cognitive science, we develop a structured approach to represent episodic events, encapsulating temporal and spatial contexts, involved entities, and detailed descriptions. We synthesize a unique episodic memory benchmark, free from contamination, and release open source code and datasets to assess LLM performance across various recall and episodic reasoning tasks. Our evaluation of state-of-the-art models, including GPT-4 and Claude variants, Llama 3.1, and o1-mini, reveals that even the most advanced LLMs struggle with episodic memory tasks, particularly when dealing with multiple related events or complex spatio-temporal relationships -- even in contexts as short as 10k-100k tokens.
|
Episodic Memory Modeling, Large Language Models, Synthetic Benchmark Generation, Cue-based Retrieval, Temporal-Spatial Reasoning, Long-context Understanding, Human-inspired AI
|
A new benchmark for testing episodic memory in AI language models, using synthetic stories to evaluate recall of specific events, their contexts, and chronological relationships.
| 12,017 |
2501.13121
|
[
0.031084472313523293,
-0.05517053231596947,
-0.010838622227311134,
0.04297369718551636,
0.08613758534193039,
0.03779163956642151,
-0.060835495591163635,
-0.01494267862290144,
0.05269889906048775,
-0.049420613795518875,
-0.06392963230609894,
-0.031220415607094765,
0.0021438540425151587,
0.062430016696453094,
0.028136935085058212,
-0.003161356784403324,
0.04136340320110321,
0.03364952653646469,
-0.04987427964806557,
-0.06427623331546783,
0.11690175533294678,
0.034623533487319946,
0.09108230471611023,
0.021284714341163635,
0.05269346386194229,
0.014937294647097588,
0.052609413862228394,
-0.04591568186879158,
0.014849413186311722,
0.019961409270763397,
0.10932841897010803,
0.009104320779442787,
0.0003907058562617749,
0.04614144563674927,
-0.013072365894913673,
0.06357451528310776,
-0.08211967349052429,
0.03394068777561188,
-0.02273613214492798,
-0.06591781973838806,
-0.00532860541716218,
0.0026780525222420692,
0.04244866222143173,
0.017842624336481094,
0.07318606227636337,
-0.04697570204734802,
-0.006874011363834143,
-0.06480705738067627,
-0.04665250703692436,
0.06278663873672485,
-0.0791790634393692,
0.04185253009200096,
-0.03071417286992073,
-0.01353355310857296,
-0.008325221948325634,
0.08832642436027527,
0.054925624281167984,
0.03476128727197647,
-0.04379243776202202,
-0.08537213504314423,
-0.06256348639726639,
-0.1595853716135025,
-0.031207207590341568,
0.00003183698208886199,
-0.06989573687314987,
-0.015020795166492462,
0.024115927517414093,
0.07829031348228455,
-0.026471637189388275,
0.010349087417125702,
-0.04506414011120796,
0.06934364885091782,
0.0008904758142307401,
0.057072870433330536,
-0.003779714461416006,
0.10875534266233444,
-0.0038660906720906496,
-0.037426192313432693,
0.05763305723667145,
-0.05242571607232094,
-0.0035801383201032877,
0.026464374735951424,
0.022431254386901855,
0.01944122277200222,
0.0877656489610672,
0.018230509012937546,
0.07854832708835602,
0.0845102071762085,
-0.004995707422494888,
0.03146177902817726,
-0.024542896077036858,
-0.13296851515769958,
0.1112644225358963,
0.008450629189610481,
0.09720780700445175,
0.07465558499097824,
-0.020795511081814766,
-0.08049393445253372,
-0.022107291966676712,
0.038118425756692886,
0.007050185464322567,
0.07489417493343353,
-0.052547387778759,
-0.009398812428116798,
0.03580503910779953,
-0.0350320003926754,
0.04847153648734093,
0.003411495592445135,
-0.04009352624416351,
-0.02546556293964386,
0.058915719389915466,
-0.022872701287269592,
0.08164262771606445,
0.04514683410525322,
0.07281115651130676,
-0.02390117198228836,
0.020399821922183037,
-0.041149165481328964,
0.04502304643392563,
0.079155832529068,
0.011904578655958176,
0.019158367067575455,
0.019688067957758904,
0.03577093780040741,
-0.03875359147787094,
-0.05188918858766556,
0.034630440175533295,
-8.50642813552702e-34,
0.02935403771698475,
0.0026343592908233404,
-0.011340844444930553,
0.08171501755714417,
0.11220330744981766,
0.01025120448321104,
-0.05116940662264824,
0.04265691712498665,
-0.016442036256194115,
0.008956843987107277,
-0.03207966685295105,
0.003682867856696248,
-0.06937284022569656,
0.07123751193284988,
0.03655890002846718,
0.02951967343688011,
-0.09630930423736572,
0.08130348473787308,
0.024788441136479378,
-0.03015873208642006,
0.0019270804477855563,
-0.0034082578495144844,
0.021679382771253586,
-0.06872498989105225,
0.030375566333532333,
-0.001330088241957128,
0.051161203533411026,
-0.06988903135061264,
-0.029873881489038467,
-0.00728236511349678,
-0.0930371955037117,
0.013243302702903748,
-0.01978900469839573,
0.025688322260975838,
0.02417643368244171,
-0.03937942162156105,
-0.035020917654037476,
-0.08460909873247147,
0.013648313470184803,
-0.001277049072086811,
0.004314671736210585,
0.0014292218256741762,
-0.028192570433020592,
-0.00936009082943201,
-0.04842021316289902,
-0.10684718936681747,
0.03822780027985573,
0.01780194230377674,
-0.021325092762708664,
-0.0026545864529907703,
0.059968240559101105,
0.04473406821489334,
-0.06476372480392456,
-0.05076232925057411,
-0.05901401862502098,
-0.01139682624489069,
-0.010600821115076542,
-0.04843388497829437,
0.02808617427945137,
0.06398320943117142,
0.048822976648807526,
0.08857058733701706,
0.042503878474235535,
0.05207470431923866,
0.035869304090738297,
0.0394197516143322,
-0.03152273967862129,
0.13400401175022125,
0.06961328536272049,
0.005770361050963402,
0.026317667216062546,
-0.03918110206723213,
0.023473156616091728,
-0.03853144496679306,
-0.014260311610996723,
0.0017861584201455116,
0.09501583129167557,
-0.08421071618795395,
-0.05854533985257149,
0.010917908512055874,
0.012146687135100365,
-0.07986283302307129,
-0.0628233402967453,
-0.028830107301473618,
-0.042586490511894226,
-0.055958133190870285,
0.053578946739435196,
-0.04613746330142021,
-0.009576800279319286,
-0.03578772395849228,
-0.026084179058670998,
-0.01980806700885296,
0.03169672563672066,
0.015326565131545067,
-0.0404895581305027,
-1.3651432599663042e-34,
-0.04474586248397827,
-0.061744339764118195,
-0.022030450403690338,
-0.01569470949470997,
0.03351084887981415,
-0.13151012361049652,
-0.06509517133235931,
0.04523492604494095,
-0.024932963773608208,
-0.06712237000465393,
-0.07937130331993103,
-0.08421403169631958,
0.06469197571277618,
-0.01038601715117693,
0.05253466218709946,
-0.05539468675851822,
-0.002420827280730009,
-0.036514896899461746,
-0.023911776021122932,
0.0717606171965599,
0.05907020345330238,
0.03837167099118233,
-0.09731132537126541,
-0.008277347311377525,
-0.000004878025720245205,
0.06416203081607819,
0.053939346224069595,
-0.04686712101101875,
-0.024041946977376938,
-0.009654292836785316,
-0.04312217980623245,
0.02962324768304825,
-0.03478773683309555,
0.021097669377923012,
-0.08341460675001144,
0.0600774921476841,
0.10008960217237473,
-0.06740086525678635,
-0.09486653655767441,
0.010025430470705032,
0.06357032805681229,
-0.05127039924263954,
-0.04783032089471817,
-0.021983088925480843,
-0.06873536109924316,
0.025982649996876717,
-0.0632704421877861,
0.023318946361541748,
0.07910318672657013,
0.023135893046855927,
-0.036146167665719986,
0.0317046195268631,
-0.06446633487939835,
-0.010855351574718952,
-0.024816295132040977,
-0.1449386179447174,
-0.05176403746008873,
-0.09427083283662796,
0.008621294051408768,
-0.01002742350101471,
-0.07353385537862778,
0.01716044172644615,
0.005478220991790295,
-0.04349231719970703,
0.03203849121928215,
-0.036218129098415375,
0.005954430438578129,
0.004434040281921625,
-0.08196497708559036,
0.014853832311928272,
0.04893714562058449,
-0.04633183032274246,
-0.059639282524585724,
-0.012538043782114983,
0.026098765432834625,
0.06639910489320755,
-0.014997378922998905,
-0.047497138381004333,
-0.029456688091158867,
-0.07249315083026886,
-0.010283364914357662,
-0.010283084586262703,
-0.008796913549304008,
0.11724750697612762,
0.024882536381483078,
0.05430125072598457,
-0.017662495374679565,
0.1289665251970291,
0.04071899875998497,
0.06908106058835983,
-0.04778902977705002,
0.017954625189304352,
-0.026009704917669296,
0.08089400082826614,
-0.04308897629380226,
-4.0466087369850356e-8,
-0.05323513597249985,
0.0007726186304353178,
0.005776145961135626,
0.05709771811962128,
0.0066614942625164986,
-0.06121007725596428,
-0.08206730335950851,
0.06227355822920799,
0.018586909398436546,
-0.04791651666164398,
0.03227410838007927,
-0.007573225069791079,
-0.06592075526714325,
0.012385336682200432,
0.013328250497579575,
0.013053511269390583,
0.025957345962524414,
-0.0034394715912640095,
0.04785808548331261,
-0.03217065706849098,
0.013116369023919106,
0.029050709679722786,
0.03989842161536217,
-0.028884615749120712,
-0.041243117302656174,
-0.04401494935154915,
-0.05238555744290352,
0.1104164719581604,
0.021736416965723038,
0.042131420224905014,
0.030758239328861237,
0.024694716557860374,
-0.025852929800748825,
0.011383351869881153,
0.03215213492512703,
-0.001668497221544385,
0.021273639053106308,
-0.010518267750740051,
0.007242028601467609,
-0.029762811958789825,
0.0572998933494091,
0.00875046756118536,
0.03426907956600189,
0.055965423583984375,
0.02331608161330223,
-0.0032978535164147615,
-0.05293872952461243,
-0.07868625968694687,
0.0001706212351564318,
-0.031256213784217834,
-0.07778428494930267,
0.025656769052147865,
0.012103154323995113,
0.07752878218889236,
0.08787662535905838,
0.07279875874519348,
-0.028919557109475136,
-0.029357334598898888,
0.04651332274079323,
-0.010983382351696491,
0.06919380277395248,
0.08825326710939407,
-0.0756768137216568,
-0.018305767327547073
] |
https://github.com/ahstat/episodic-memory-benchmark
| 41 | 0 | 0 | 0 |
Continual Slow-and-Fast Adaptation of Latent Neural Dynamics (CoSFan): Meta-Learning What-How & When to Adapt
|
https://openreview.net/forum?id=Dl3MsjaIdp
|
[
"Ryan Missel",
"Linwei Wang"
] |
Poster
|
An increasing interest in learning to forecast for time-series of high-dimensional observations is the ability to adapt to systems with diverse underlying dynamics. Access to observations that define a stationary distribution of these systems is often unattainable, as the underlying dynamics may change over time. Naively training or retraining models at each shift may lead to catastrophic forgetting about previously-seen systems. We present a new continual meta-learning (CML) framework to realize continual slow-and fast adaptation of latent dynamics (CoSFan). We leverage a feed-forward meta-model to infer *what* the current system is and *how* to adapt a latent dynamics function to it, enabling *fast adaptation* to specific dynamics. We then develop novel strategies to automatically detect *when* a shift of data distribution occurs, with which to identify its underlying dynamics and its relation with previously-seen dynamics. In combination with fixed-memory experience replay mechanisms, this enables continual *slow update* of the *what-how* meta-model. Empirical studies demonstrated that both the meta- and continual-learning component was critical for learning to forecast across non-stationary distributions of diverse dynamics systems, and the feed-forward meta-model combined with task-aware/-relational continual learning strategies significantly outperformed existing CML alternatives.
|
continual meta-learning, latent dynamics forecasting, time-series
|
In this work, we present a new continual meta-learning (CML) framework to realize continual slow-and fast adaptation of latent dynamics (CoSFan) by meta-learning what-how & when to adapt.
| 12,012 | null |
[
-0.02962992712855339,
-0.10565917938947678,
0.006872063502669334,
0.10809125006198883,
0.0757436603307724,
0.03484542667865753,
-0.03493047505617142,
-0.07222127914428711,
0.07638179510831833,
-0.03714537248015404,
-0.0883558988571167,
0.013314686715602875,
-0.03047879971563816,
0.011348545551300049,
-0.08384089171886444,
-0.0585361011326313,
0.002797227120026946,
0.03938516974449158,
-0.03453058749437332,
-0.05042356252670288,
-0.003857197007164359,
0.019365794956684113,
0.016291096806526184,
0.0803990364074707,
-0.015228558331727982,
0.032379601150751114,
-0.04135392606258392,
0.008270971477031708,
0.003577508730813861,
-0.04509101063013077,
0.03199146315455437,
0.06700357049703598,
-0.1273471862077713,
-0.022302530705928802,
-0.08331108838319778,
0.04324744641780853,
-0.0663948655128479,
0.07585028558969498,
-0.0028636958450078964,
-0.00981536041945219,
0.04320459067821503,
-0.00005201211024541408,
-0.03640141710639,
-0.0017124206060543656,
-0.003333711763843894,
-0.052017346024513245,
0.01698680780827999,
-0.08254803717136383,
0.006406255532056093,
0.01388331688940525,
-0.08283857256174088,
0.01317776832729578,
-0.04218652471899986,
0.06418044865131378,
0.09515241533517838,
0.07178992033004761,
0.038355227559804916,
0.03675884008407593,
-0.048390038311481476,
-0.010473047383129597,
-0.011772383004426956,
-0.05339612066745758,
0.010896865278482437,
-0.013615270145237446,
-0.07732969522476196,
0.08621565252542496,
-0.019426418468356133,
0.08166360855102539,
0.032880254089832306,
0.003398242872208357,
-0.0347755067050457,
0.11000645160675049,
-0.06575725972652435,
0.011048411950469017,
0.041838981211185455,
0.07359287142753601,
-0.001545096980407834,
-0.052945178002119064,
0.058825816959142685,
-0.00371152488514781,
0.005437890533357859,
0.07049189507961273,
-0.008324531838297844,
-0.05689501389861107,
0.10058501362800598,
-0.015238451771438122,
0.06319913268089294,
0.03954891860485077,
0.049914028495550156,
-0.035253968089818954,
0.03238779306411743,
-0.01670258678495884,
0.028679372742772102,
-0.004118255339562893,
0.011518449522554874,
0.04777970165014267,
0.04258451610803604,
-0.06363637000322342,
0.013839401304721832,
0.07516499608755112,
-0.008128700777888298,
0.06531061232089996,
-0.03260604292154312,
0.10461711883544922,
-0.0002891314506996423,
-0.09160079807043076,
0.020570719614624977,
0.10302633792161942,
0.020323704928159714,
-0.011125253513455391,
0.021163199096918106,
0.021723726764321327,
0.024942755699157715,
0.01157216913998127,
0.015180197544395924,
-0.008991985581815243,
-0.001819443074055016,
-0.0123674925416708,
-0.06164172664284706,
0.10005174577236176,
0.051258739084005356,
0.006487314589321613,
-0.008096720091998577,
-0.020334575325250626,
0.08558019995689392,
0.01150792557746172,
-0.08377144485712051,
-8.107071378579418e-34,
-0.004158113617449999,
0.026229197159409523,
0.05984365940093994,
-0.0036418060772120953,
0.014624780975282192,
-0.06129715219140053,
0.002282531699165702,
0.016549456864595413,
0.05053291097283363,
-0.04850127175450325,
0.02484198845922947,
0.09520065784454346,
-0.07919015735387802,
0.07158993929624557,
0.002500233706086874,
-0.03946636989712715,
-0.08948485553264618,
0.016715798527002335,
0.038160260766744614,
-0.04405719414353371,
0.03818719834089279,
0.02056957222521305,
-0.023447126150131226,
-0.0876532644033432,
0.015335490927100182,
0.0070059290155768394,
0.04520449787378311,
0.04562721773982048,
-0.027154039591550827,
0.037329476326704025,
-0.05419537052512169,
0.05342372506856918,
-0.022583043202757835,
0.0003070531238336116,
-0.03426320478320122,
-0.09781399369239807,
0.011871042661368847,
-0.05519566312432289,
0.018127845600247383,
-0.08511582762002945,
0.032490890473127365,
-0.0007413914427161217,
-0.08235904574394226,
0.002555616432800889,
-0.061579011380672455,
-0.13996754586696625,
0.058423951268196106,
0.029980091378092766,
-0.06292339414358139,
-0.05601954087615013,
0.024555345997214317,
0.040703218430280685,
-0.0660509467124939,
-0.0860416516661644,
-0.08229643851518631,
0.019660184159874916,
-0.01968006044626236,
-0.03631393983960152,
-0.01523379236459732,
0.05102432146668434,
0.048591259866952896,
0.004992694593966007,
0.004910551011562347,
0.07552876323461533,
0.02226930670440197,
0.04387335851788521,
-0.0013936248142272234,
0.06841021031141281,
0.005242324434220791,
0.002547621028497815,
0.0384879894554615,
-0.015723291784524918,
-0.034139782190322876,
-0.07573681324720383,
0.08631225675344467,
-0.07958631962537766,
0.08335527032613754,
-0.09063674509525299,
-0.04082614555954933,
0.03099500574171543,
0.040542107075452805,
-0.08920596539974213,
-0.03704271465539932,
0.04052809253334999,
-0.010152594186365604,
-0.039171166718006134,
0.04588339850306511,
-0.08274286985397339,
-0.036909300833940506,
-0.03492926061153412,
-0.049155231565237045,
-0.06091531366109848,
0.06572611629962921,
0.0014491890324279666,
-0.021201608702540398,
-1.4390728777132766e-33,
-0.012744714505970478,
-0.06116195023059845,
-0.012838169001042843,
0.034918639808893204,
-0.012706190347671509,
0.004593368619680405,
-0.08280668407678604,
0.045827895402908325,
-0.06040821596980095,
-0.10952543467283249,
-0.013380567543208599,
-0.06916417926549911,
0.003046911209821701,
0.045161791145801544,
-0.022468550130724907,
-0.01734585501253605,
0.034362051635980606,
-0.048154741525650024,
0.010882895439863205,
0.011814335361123085,
-0.016821913421154022,
0.009658697061240673,
-0.044975169003009796,
-0.0012247114209458232,
-0.00023446683189831674,
0.03110813908278942,
0.04850335419178009,
0.09504622220993042,
-0.03155296668410301,
-0.023610329255461693,
-0.05160680413246155,
-0.013780390843749046,
0.007752326317131519,
-0.06515081971883774,
-0.046414539217948914,
0.11159486323595047,
0.029044628143310547,
-0.039073482155799866,
-0.04308639466762543,
0.0948399007320404,
0.08523182570934296,
-0.05899694189429283,
0.013433191925287247,
-0.024942105636000633,
0.001036991598084569,
-0.046521928161382675,
-0.06186715513467789,
0.10045772790908813,
0.08224649727344513,
0.06513439863920212,
0.0118568679317832,
0.037897273898124695,
-0.11453709006309509,
-0.059948667883872986,
-0.03147099167108536,
-0.007641898933798075,
0.0318029522895813,
-0.01801237091422081,
0.03464888781309128,
0.04178142175078392,
-0.03859233856201172,
-0.04938575252890587,
0.00004489703133003786,
-0.06970631331205368,
0.025048119947314262,
-0.007211669813841581,
0.014130971394479275,
-0.07867960631847382,
0.07342149317264557,
0.013696513138711452,
0.11937617510557175,
0.06394065916538239,
-0.05850078910589218,
-0.0482124425470829,
0.06727009266614914,
-0.004439891315996647,
-0.05074920505285263,
-0.047682490199804306,
-0.03692229092121124,
-0.02948296256363392,
-0.003833142342045903,
0.03362175077199936,
0.051255371421575546,
0.01972523145377636,
-0.061889130622148514,
0.08083202689886093,
-0.022905811667442322,
0.10231783241033554,
0.01235722005367279,
-0.005736235994845629,
-0.054698556661605835,
-0.03263135254383087,
-0.07037869095802307,
-0.00170027872081846,
-0.07737883180379868,
-4.033208611531336e-8,
-0.08350399881601334,
0.07118378579616547,
0.07317929714918137,
0.0430259071290493,
0.08995940536260605,
-0.0027233704458922148,
0.036741483956575394,
-0.003627002239227295,
0.002988097257912159,
-0.0768764317035675,
0.1344185471534729,
-0.020076822489500046,
0.04927844926714897,
0.00044596154475584626,
0.006094984244555235,
0.07988148182630539,
0.037676237523555756,
-0.024974403902888298,
-0.03645884618163109,
-0.055470675230026245,
0.007996795699000359,
0.029686743393540382,
0.05073738843202591,
-0.029518622905015945,
0.022332720458507538,
-0.06538501381874084,
-0.018809938803315163,
0.12977123260498047,
0.06854893267154694,
-0.007428513839840889,
-0.006464540027081966,
-0.014925516210496426,
0.0010942347580567002,
-0.007540679071098566,
-0.059834741055965424,
0.05402473732829094,
0.07319515943527222,
-0.06698868423700333,
-0.01244348380714655,
-0.024903152137994766,
-0.003336801892146468,
0.04340420663356781,
-0.02329903654754162,
0.016549767926335335,
-0.03759283944964409,
-0.011348026804625988,
-0.023690201342105865,
-0.0743294432759285,
0.031544022262096405,
-0.04021186754107475,
0.011106127873063087,
-0.020685110241174698,
0.015368836000561714,
0.09580422937870026,
0.08488161861896515,
0.023413654416799545,
-0.014092721976339817,
-0.046209536492824554,
0.004561730660498142,
0.026582781225442886,
0.012555682100355625,
0.023985743522644043,
-0.06300397217273712,
0.017090201377868652
] | 0 | 0 | 0 | 0 |
|
R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference
|
https://openreview.net/forum?id=9VMW4iXfKt
|
[
"Zhenyu Zhang",
"Zechun Liu",
"Yuandong Tian",
"Harshit Khaitan",
"Zhangyang Wang",
"Steven Li"
] |
Poster
|
Large Language Models (LLMs), while demonstrating remarkable capabilities across various applications, present significant challenges during inference due to their substantial model size, especially when deployed on edge devices. Activation sparsity offers a promising solution to reduce computation and memory movement, enabling more efficient inference, particularly for small-batch on-device applications. However, current approaches face limitations with non-ReLU activation function, which are foundational to most advanced LLMs, or require heavy continual training. Additionally, the difficulty in predicting active channels and limited achievable sparsity ratios constrain the effectiveness of activation sparsity-based methods. In this paper, we introduce R-Sparse, a training-free activation sparsity approach capable of achieving high sparsity levels in advanced LLMs. We conducted two preliminary investigations into how different components contribute to the output within a single linear layer and found two key observations: (i) the non-sparse components of the input function can be regarded as a few bias terms, and (ii) The full computation can be effectively approximated by an appropriate combination of input channels and weight singular values. Building on this, we replace the linear layers in LLMs with a rank-aware sparse inference method that leverages the sparsity of input channels and singular value components, eliminating the need for active channel prediction like the output sparsity based approaches. Experiments on Llama-2/3 and Mistral models across ten diverse tasks demonstrate that R-Sparse achieves comparable performance at 50\% model-level sparsity, resulting in a significant 43\% end-to-end efficient improvements with customized kernels.
|
Large Language Model; Efficient Inference; Activation Sparsity
| null | 12,008 | null |
[
0.002520184265449643,
-0.12440090626478195,
0.007541181985288858,
0.07081171870231628,
0.08802522718906403,
0.03658224642276764,
-0.022778823971748352,
-0.007434018887579441,
-0.017217963933944702,
-0.06312812119722366,
-0.022485537454485893,
-0.02929658815264702,
0.003951616119593382,
0.012509899213910103,
0.02261473797261715,
0.019473709166049957,
0.0715121403336525,
0.07627307623624802,
-0.08128238469362259,
-0.09638656675815582,
0.012289183214306831,
-0.021395517513155937,
0.017071273177862167,
0.010581258684396744,
0.11963779479265213,
-0.07870856672525406,
-0.0439564473927021,
0.013641170226037502,
-0.00392123544588685,
-0.03775971010327339,
0.07343790680170059,
0.05289342254400253,
0.01177092082798481,
0.004023067653179169,
-0.0041470942087471485,
0.011958256363868713,
-0.04674345999956131,
0.01868502050638199,
0.00007284445746336132,
-0.03377104923129082,
0.028932660818099976,
-0.019153840839862823,
-0.003255479736253619,
0.05857523903250694,
0.03657886013388634,
-0.03237899765372276,
0.006066243164241314,
-0.058463215827941895,
0.020266294479370117,
0.03499453514814377,
0.022779179736971855,
0.012632797472178936,
-0.009257908910512924,
0.04620169848203659,
-0.07825404405593872,
-0.05180245265364647,
-0.004428232088685036,
0.040940821170806885,
-0.04615307226777077,
0.005000689532607794,
-0.07044839859008789,
-0.13233932852745056,
0.016280943527817726,
0.012296143919229507,
-0.05381322279572487,
-0.01582067459821701,
0.06413774937391281,
0.02858780138194561,
-0.010653022676706314,
0.054680757224559784,
-0.005312083754688501,
0.018743475899100304,
-0.05404715612530708,
0.042502082884311676,
0.03311165049672127,
-0.029504483565688133,
0.05231374129652977,
-0.02335546165704727,
0.07853704690933228,
-0.0016055607702583075,
0.03520284220576286,
-0.002083699218928814,
-0.03311769664287567,
-0.053302861750125885,
0.09554989635944366,
0.004194499924778938,
0.03088770993053913,
0.005064179189503193,
-0.0003191698051523417,
0.03448082506656647,
-0.016784684732556343,
-0.03893925994634628,
-0.013257955200970173,
-0.0019447748782113194,
-0.003862216603010893,
0.0013836342841386795,
-0.018747609108686447,
-0.07688423246145248,
-0.09498006105422974,
0.11993501335382462,
0.002609887858852744,
-0.01172112487256527,
-0.03304620832204819,
-0.007433271035552025,
-0.008651962503790855,
0.0049704513512551785,
0.020447954535484314,
0.09498237818479538,
0.007418051827698946,
-0.08574362844228745,
0.026073036715388298,
0.05568096414208412,
0.061927519738674164,
0.0923105925321579,
0.010834138840436935,
-0.022128049284219742,
0.10649257898330688,
-0.10485612601041794,
0.04866785556077957,
0.0013441558694466949,
-0.04912343621253967,
-0.013006896711885929,
-0.019360365346074104,
-0.03621849790215492,
-0.002486306009814143,
-0.009440833702683449,
-0.04608241841197014,
3.5381380789552655e-33,
0.07728885859251022,
0.10343579947948456,
-0.028410717844963074,
0.0005682319751940668,
0.01636713556945324,
-0.0070337168872356415,
0.10015071928501129,
0.05917257070541382,
-0.038671575486660004,
0.004832518752664328,
-0.11490084230899811,
-0.026359261944890022,
-0.022159339860081673,
0.06763243675231934,
0.004555048421025276,
0.03757580742239952,
0.023441331461071968,
0.029897518455982208,
0.007696426007896662,
-0.08674955368041992,
0.01332723069936037,
-0.00409671850502491,
-0.00949928443878889,
-0.06580868363380432,
0.01759464666247368,
-0.0168625321239233,
0.07143117487430573,
-0.07693660259246826,
-0.07986295968294144,
0.0011395657202228904,
-0.11925875395536423,
0.015872981399297714,
0.07006125152111053,
0.021087048575282097,
0.07511328905820847,
0.017168089747428894,
0.02812577597796917,
-0.02190428599715233,
0.08401576429605484,
-0.05482632294297218,
-0.03258634731173515,
0.007239453494548798,
0.01826135814189911,
-0.06356208771467209,
-0.13654758036136627,
-0.020024022087454796,
-0.009459166787564754,
0.012535651214420795,
-0.03589668497443199,
0.018870750442147255,
0.07181081175804138,
-0.00930069386959076,
-0.027223635464906693,
0.024897634983062744,
0.03589104115962982,
-0.09013616293668747,
-0.03682463616132736,
0.05010220408439636,
0.08586994558572769,
0.061882395297288895,
-0.03926558420062065,
0.024247152730822563,
-0.03349882736802101,
0.04886311665177345,
0.05209842696785927,
0.023324282839894295,
-0.02505033276975155,
0.011822117492556572,
0.03550896421074867,
-0.031618643552064896,
0.01739598996937275,
0.009764027781784534,
0.005222556181252003,
-0.06433099508285522,
0.05305797606706619,
0.00866380613297224,
0.07656347006559372,
-0.1442025750875473,
0.047166548669338226,
0.05863040313124657,
0.02770751528441906,
-0.03288149833679199,
0.014817068353295326,
-0.06427744776010513,
-0.08267753571271896,
-0.03746270388364792,
0.031972069293260574,
-0.0682605430483818,
0.025235667824745178,
-0.03622213006019592,
-0.0522104874253273,
-0.030560879036784172,
-0.02294539287686348,
-0.07396173477172852,
-0.01928240805864334,
-1.956633856830714e-33,
-0.05255081504583359,
0.07508604973554611,
-0.0037736166268587112,
0.03942909464240074,
-0.04915134608745575,
-0.05360719934105873,
0.0067785256542265415,
0.06271123886108398,
-0.07080499827861786,
-0.09107878804206848,
-0.03914918377995491,
-0.06339932233095169,
0.026733718812465668,
-0.01120623666793108,
0.09511122852563858,
-0.009284314699470997,
-0.11047939956188202,
0.02521720714867115,
-0.004635037388652563,
0.04294140264391899,
-0.03024100884795189,
0.09217680245637894,
-0.056280069053173065,
0.07605720311403275,
-0.038006603717803955,
0.026188433170318604,
-0.048808325082063675,
0.10430815815925598,
0.025154557079076767,
-0.07258113473653793,
-0.020377863198518753,
0.020174412056803703,
-0.05414849892258644,
-0.04775970056653023,
-0.007413018494844437,
0.10501140356063843,
0.07773936539888382,
0.06665906310081482,
-0.015885693952441216,
0.020124610513448715,
0.08815794438123703,
-0.024109255522489548,
-0.05972969904541969,
-0.011227737180888653,
0.026390716433525085,
-0.023870151489973068,
-0.0902707502245903,
0.0019550621509552,
0.02609783224761486,
-0.001639653230085969,
0.0003198539197910577,
0.030423913151025772,
0.0016263641882687807,
0.022728336974978447,
-0.01624196022748947,
-0.0721106305718422,
-0.03714032843708992,
-0.04014269635081291,
0.062122706323862076,
-0.03522089123725891,
-0.0017205419717356563,
-0.04289090260863304,
-0.037253327667713165,
-0.11098186671733856,
0.014094062149524689,
-0.03528061509132385,
0.043485186994075775,
0.025326773524284363,
-0.01769077777862549,
0.04833128675818443,
0.04561606049537659,
-0.020477889105677605,
0.02671870030462742,
0.009717191569507122,
-0.020544400438666344,
0.005802270490676165,
-0.03178003057837486,
-0.08640790730714798,
-0.06292315572500229,
0.02094578929245472,
0.06398194283246994,
0.10105043649673462,
-0.02466840296983719,
0.026839157566428185,
0.08343703299760818,
0.15778252482414246,
0.02707085572183132,
0.05602027103304863,
0.04627212509512901,
0.017857490107417107,
-0.03796817734837532,
0.04320485517382622,
0.026921499520540237,
0.045244161039590836,
0.015525496564805508,
-4.351084825771068e-8,
0.002779455855488777,
0.03817729651927948,
-0.012325405143201351,
0.007500280626118183,
-0.009047088213264942,
-0.054296962916851044,
0.0601569302380085,
0.03802680969238281,
0.014755398966372013,
-0.03819691389799118,
0.056722242385149,
-0.022448426112532616,
-0.0459304042160511,
-0.07249006628990173,
0.044045768678188324,
0.10982000082731247,
-0.006904542446136475,
0.04208756610751152,
-0.0069975960068404675,
-0.04235660284757614,
0.015605135820806026,
0.08419626206159592,
0.011789117939770222,
-0.00571979396045208,
0.044245343655347824,
-0.05593939870595932,
0.03491687774658203,
0.14495231211185455,
0.11065990477800369,
-0.033153656870126724,
-0.016705557703971863,
-0.02842419221997261,
0.02931949682533741,
-0.026230517774820328,
0.01278984360396862,
0.09724993258714676,
0.06774748861789703,
-0.013129855506122112,
-0.04813355579972267,
-0.07936161011457443,
-0.09671605378389359,
-0.017324667423963547,
-0.02393527887761593,
0.013035163283348083,
-0.04163803905248642,
-0.03495281562209129,
-0.05398871377110481,
-0.1339058130979538,
0.04132701829075813,
-0.05828329920768738,
0.0387304425239563,
0.0014861886156722903,
-0.052300747483968735,
0.06800412386655807,
0.006751765497028828,
0.030865907669067383,
-0.00819004699587822,
0.02995866909623146,
0.023216642439365387,
0.041133176535367966,
0.05243409052491188,
0.018269484862685204,
-0.08268603682518005,
0.003918435424566269
] | 0 | 0 | 0 | 0 |
|
Query-based Knowledge Transfer for Heterogeneous Learning Environments
|
https://openreview.net/forum?id=XKv29sMyjF
|
[
"Norah Alballa",
"Wenxuan Zhang",
"Ziquan Liu",
"Ahmed M. Abdelmoniem",
"Mohamed Elhoseiny",
"Marco Canini"
] |
Poster
|
Decentralized collaborative learning under data heterogeneity and privacy constraints has rapidly advanced. However, existing solutions like federated learning, ensembles, and transfer learning, often fail to adequately serve the unique needs of clients, especially when local data representation is limited.
To address this issue, we propose a novel framework called Query-based Knowledge Transfer (QKT) that enables tailored knowledge acquisition to fulfill specific client needs without direct data exchange.
It employs a data-free masking strategy to facilitate the communication-efficient query-focused knowledge transformation while refining task-specific parameters to mitigate knowledge interference and forgetting. Our experiments, conducted on both standard and clinical benchmarks, show that QKT significantly outperforms existing collaborative learning methods by an average of 20.91% points in single-class query settings and an average of 14.32% points in multi-class query scenarios.
Further analysis and ablation studies reveal that QKT effectively balances the learning of new and existing knowledge, showing strong potential for its application in decentralized learning.
|
Collaborative Learning, Knowledge Distillation, Query-based Knowledge Transfer.
|
Query-based Knowledge Transfer
| 12,001 | null |
[
-0.027352675795555115,
-0.014729516580700874,
0.0033519186545163393,
0.11262012273073196,
-0.02335331402719021,
0.005297002382576466,
0.025291671976447105,
-0.01831340231001377,
-0.003938162699341774,
-0.010117225348949432,
-0.028085215017199516,
0.01980282925069332,
0.06543589383363724,
0.008505338802933693,
0.00919516384601593,
0.03636154532432556,
0.04915138706564903,
-0.005080713890492916,
-0.047225967049598694,
-0.008166616782546043,
-0.013958368450403214,
0.03611072525382042,
0.012170189060270786,
0.03649840131402016,
0.04423314705491066,
-0.009244932793080807,
0.022057445719838142,
-0.09544633328914642,
0.012975890189409256,
-0.004270137287676334,
-0.012434142641723156,
0.029237322509288788,
-0.08124309778213501,
0.04112342745065689,
-0.08204232156276703,
0.028751840814948082,
-0.046457502990961075,
0.08662068843841553,
-0.0752522274851799,
-0.03177114203572273,
-0.04595330357551575,
-0.002296748571097851,
-0.037497010082006454,
0.0718744620680809,
-0.006844832096248865,
0.032264456152915955,
-0.020441625267267227,
0.045411672443151474,
0.01442770380526781,
0.02798713557422161,
-0.11917043477296829,
-0.08847305923700333,
-0.09106841683387756,
0.08695334196090698,
0.026465343311429024,
0.04951312392950058,
-0.05230821669101715,
-0.03388119116425514,
-0.09144958853721619,
-0.01267192605882883,
-0.11131293326616287,
-0.07455316931009293,
0.025603359565138817,
0.04255342483520508,
-0.05333778262138367,
-0.032379426062107086,
0.0454300157725811,
0.08307226747274399,
0.035731907933950424,
-0.0725044459104538,
-0.07741352170705795,
0.050606198608875275,
-0.036842230707407,
-0.011072400957345963,
0.0683947503566742,
0.018973132595419884,
-0.06548576802015305,
-0.01681964471936226,
0.033378519117832184,
-0.009768097661435604,
0.014609315432608128,
0.08724545687437057,
-0.024320611730217934,
-0.002613421995192766,
0.11755552142858505,
-0.05039341375231743,
0.07868281751871109,
0.003145144321024418,
-0.06590689718723297,
-0.10300443321466446,
0.0031529907137155533,
0.06607311964035034,
-0.04506130516529083,
-0.031008698046207428,
-0.010924830101430416,
-0.020188575610518456,
0.05056540668010712,
-0.018803097307682037,
0.02028028294444084,
0.06990503519773483,
-0.022101789712905884,
0.09557020664215088,
-0.04648468643426895,
-0.020833110436797142,
-0.08109576255083084,
-0.011860938742756844,
0.05003952980041504,
0.005338453222066164,
0.08219899237155914,
-0.02277449145913124,
-0.017704946920275688,
0.0508754625916481,
-0.027745896950364113,
-0.046519361436367035,
-0.04709704965353012,
0.033286385238170624,
-0.002155567519366741,
0.02157410979270935,
-0.024269724264740944,
0.10048332810401917,
-0.036870326846838,
0.017043564468622208,
0.016683783382177353,
0.016722002997994423,
0.027573203667998314,
0.03002968430519104,
-0.1033046692609787,
1.2144634579199849e-33,
0.1122848317027092,
0.06806807219982147,
0.046154167503118515,
-0.018078595399856567,
-0.011904408223927021,
-0.0076104686595499516,
0.042746346443891525,
0.08164255321025848,
-0.05941741541028023,
-0.028184808790683746,
0.08245421200990677,
0.03504102677106857,
0.07314596325159073,
0.020345624536275864,
0.04736846312880516,
0.027876093983650208,
-0.08285605907440186,
0.06721417605876923,
0.040289998054504395,
-0.037363771349191666,
0.059858713299036026,
-0.04160210117697716,
-0.021453484892845154,
-0.010189403779804707,
-0.02439907193183899,
0.04958905652165413,
-0.006300572771579027,
0.09148760139942169,
0.10831481218338013,
-0.014009416103363037,
-0.09820186346769333,
-0.04656282812356949,
0.009514681063592434,
0.01949637569487095,
0.004441861528903246,
0.0021996619179844856,
0.013715849258005619,
-0.09736286848783493,
0.026533959433436394,
-0.008324269205331802,
-0.036745116114616394,
-0.01245517935603857,
0.03404797986149788,
-0.03519408032298088,
-0.05427457392215729,
-0.0730024054646492,
-0.025815164670348167,
-0.06058688089251518,
-0.038992445915937424,
-0.0856708437204361,
-0.03463348373770714,
-0.05309532955288887,
-0.030636971816420555,
-0.01248080376535654,
0.011321415193378925,
0.031722281128168106,
0.005180935841053724,
0.02071469835937023,
-0.05692391097545624,
0.04087122902274132,
0.01819595880806446,
0.01810430735349655,
0.012352300807833672,
0.025585496798157692,
0.006963843479752541,
0.009746274910867214,
0.0011066346196457744,
0.02976107969880104,
0.12678512930870056,
-0.09165776520967484,
-0.033331602811813354,
0.024726970121264458,
-0.02050233818590641,
0.00877612829208374,
0.04633527249097824,
-0.04026232659816742,
-0.019373958930373192,
-0.07634946703910828,
0.027438372373580933,
0.0033327373676002026,
0.05879999324679375,
-0.04953201115131378,
-0.050638675689697266,
0.019402211531996727,
-0.07949165999889374,
-0.013231186196208,
0.006908736657351255,
-0.02905896306037903,
-0.07806851714849472,
0.08322854340076447,
-0.033830419182777405,
0.0330323688685894,
-0.03627370297908783,
0.02362922951579094,
0.053061842918395996,
-1.4483198856569209e-33,
-0.007562323473393917,
-0.035093020647764206,
-0.044471994042396545,
0.06733623892068863,
0.0738258883357048,
-0.022524410858750343,
-0.021458378061652184,
-0.03756803646683693,
-0.025884084403514862,
-0.0347989946603775,
-0.01916102133691311,
0.0032150635961443186,
0.04207437485456467,
-0.04589270055294037,
-0.04375440627336502,
0.08853612095117569,
-0.06676915287971497,
0.016897644847631454,
-0.047610942274332047,
0.047527991235256195,
-0.07902694493532181,
0.03847064822912216,
-0.03477287292480469,
0.07320115715265274,
0.04483981803059578,
0.025085026398301125,
-0.04932944104075432,
0.07506690919399261,
0.05779789015650749,
0.030794521793723106,
-0.031621117144823074,
-0.1138162836432457,
-0.03195231407880783,
-0.1114279180765152,
-0.07353653758764267,
0.04700321704149246,
0.004343883134424686,
0.0016847423976287246,
-0.07434116303920746,
0.07493053376674652,
0.01713191531598568,
-0.047454990446567535,
-0.10168120265007019,
-0.01669912040233612,
0.017560498788952827,
-0.023675762116909027,
-0.0828908309340477,
-0.015566766262054443,
0.04617329314351082,
0.030973326414823532,
0.08541526645421982,
0.057068921625614166,
-0.07476437091827393,
-0.050074346363544464,
-0.018517572432756424,
-0.017579957842826843,
0.019022131338715553,
0.008696911856532097,
0.012495007365942001,
0.01044198963791132,
0.01213381439447403,
-0.022592034190893173,
-0.01782728172838688,
0.03179473802447319,
0.06400366872549057,
-0.009413044899702072,
0.03237757459282875,
0.09905374050140381,
-0.02763327769935131,
0.07783941179513931,
-0.049667518585920334,
0.009074443951249123,
0.01230224221944809,
-0.10184181481599808,
0.11652282625436783,
0.042888153344392776,
0.013015500269830227,
-0.04049975425004959,
-0.11804413795471191,
0.05626346915960312,
0.053046416491270065,
-0.05333699658513069,
0.03840044140815735,
0.09659866988658905,
0.04542839527130127,
0.040556736290454865,
0.09031215310096741,
0.06504343450069427,
-0.03229020908474922,
-0.04106626287102699,
-0.027452871203422546,
-0.017904607579112053,
-0.06904521584510803,
0.0037045502103865147,
-0.01500620972365141,
-4.326400571130762e-8,
-0.04315878823399544,
0.06984415650367737,
-0.01654977723956108,
0.035575371235609055,
0.04115680977702141,
-0.011377123184502125,
-0.09298166632652283,
0.09949009865522385,
-0.071705661714077,
0.04209225997328758,
-0.01655728742480278,
0.02345818281173706,
0.07816465944051743,
-0.02257937751710415,
0.06547586619853973,
0.03678448125720024,
0.03568819537758827,
0.0019619078375399113,
-0.013328623957931995,
-0.06013130769133568,
-0.031624797731637955,
-0.04570598900318146,
-0.036723844707012177,
-0.01076000276952982,
0.044441696256399155,
-0.03561348840594292,
0.06425565481185913,
0.09324898570775986,
0.10008177906274796,
0.0317179299890995,
-0.08916721493005753,
-0.05843926966190338,
0.03306717053055763,
-0.016644904389977455,
0.0718197301030159,
0.028791556134819984,
0.012052732519805431,
-0.04002310335636139,
-0.0599108524620533,
-0.03874675929546356,
0.02445068582892418,
0.07347826659679413,
-0.02868608571588993,
0.06638786196708679,
0.01726691797375679,
-0.007067643105983734,
-0.012666002847254276,
-0.04226575419306755,
0.05945204198360443,
0.019467655569314957,
0.04303300380706787,
-0.04024200513958931,
0.03310529887676239,
0.03125903755426407,
0.11262549459934235,
0.03576912358403206,
-0.006258358247578144,
0.002122283913195133,
0.07004761695861816,
0.011109715327620506,
0.04620383679866791,
-0.02816089428961277,
-0.07855946570634842,
0.08834244310855865
] | 0 | 0 | 0 | 0 |
|
Language Models Are Implicitly Continuous
|
https://openreview.net/forum?id=SMK0f8JoKF
|
[
"Samuele Marro",
"Davide Evangelista",
"X. Angelo Huang",
"Emanuele La Malfa",
"Michele Lombardi",
"Michael J. Wooldridge"
] |
Poster
|
Language is typically modelled with discrete sequences. However, the most successful approaches to language modelling, namely neural networks, are continuous and smooth function approximators.
In this work, we show that Transformer-based language models implicitly learn to represent sentences as continuous-time functions defined over a continuous input space.
This phenomenon occurs in most state-of-the-art Large Language Models (LLMs), including Llama2, Llama3, Phi3, Gemma, Gemma2, and Mistral, and suggests that LLMs reason about language in ways that fundamentally differ from humans.
Our work formally extends Transformers to capture the nuances of time and space continuity in both input and output space.
Our results challenge the traditional interpretation of how LLMs understand language, with several linguistic and engineering implications.
|
llm, continuity, spatiotemporal transformers, linguistics
|
LLMs implicitly behave like continuous models, even when trained in a discrete fashion.
| 11,994 | null |
[
-0.12370007485151291,
-0.05344853922724724,
0.05996743217110634,
-0.007169273216277361,
-0.024307215586304665,
-0.0011797036277130246,
-0.03598606586456299,
-0.004147286061197519,
0.11245816200971603,
-0.05688798427581787,
-0.007084183860570192,
-0.022399522364139557,
-0.0007189199677668512,
0.06230451539158821,
0.0020004434045404196,
0.0016425854992121458,
-0.006297057494521141,
0.07886971533298492,
-0.0729779303073883,
-0.10296978056430817,
0.0688994750380516,
0.09163841605186462,
-0.009409807622432709,
0.019717037677764893,
0.01804264262318611,
0.08718954771757126,
0.022516587749123573,
0.004855865612626076,
0.014351326040923595,
0.013039465062320232,
-0.027031386271119118,
-0.011190890334546566,
-0.03865165263414383,
0.034855037927627563,
-0.07000067830085754,
0.0746387392282486,
-0.042613327503204346,
-0.01118427887558937,
-0.03375086933374405,
-0.03918462246656418,
0.01348655391484499,
-0.00806022435426712,
0.057092584669589996,
-0.013657043687999249,
0.09086588770151138,
-0.049743153154850006,
0.018309343606233597,
-0.054255515336990356,
-0.10374387353658676,
0.06013442203402519,
-0.08102866262197495,
-0.02545156516134739,
0.028132230043411255,
0.05511119216680527,
-0.02468382939696312,
-0.003688941476866603,
0.022723179310560226,
0.052357953041791916,
-0.016757208853960037,
-0.09790089726448059,
-0.04763343185186386,
-0.03043203242123127,
0.0028136104810982943,
0.015149371698498726,
-0.01945572905242443,
0.009792895056307316,
-0.011055435985326767,
0.09303531050682068,
-0.05021418258547783,
0.07027345150709152,
-0.048042524605989456,
0.06223759055137634,
-0.03588932752609253,
0.040673818439245224,
-0.012338205240666866,
-0.03531010076403618,
0.12299435585737228,
0.009877489879727364,
0.09145505726337433,
-0.04016933590173721,
0.031551212072372437,
-0.031930550932884216,
0.030076272785663605,
0.033854708075523376,
0.022788042202591896,
0.036267634481191635,
-0.058841172605752945,
0.07911238819360733,
-0.011010684072971344,
-0.02193276211619377,
-0.07539398968219757,
-0.057615190744400024,
0.005796109791845083,
-0.009111435152590275,
0.005976396147161722,
0.014144770801067352,
-0.015338970348238945,
-0.030570892617106438,
-0.004299785941839218,
0.11710931360721588,
-0.027686309069395065,
0.03702990338206291,
0.05834376439452171,
-0.07717856764793396,
-0.02068324387073517,
-0.07954619079828262,
-0.02044695056974888,
0.010180287063121796,
0.03909654915332794,
-0.05464937537908554,
-0.013255544006824493,
0.00011370801803423092,
-0.021357735618948936,
0.0322117917239666,
0.04061862826347351,
-0.052615970373153687,
0.024044882506132126,
-0.0789472684264183,
0.02475688047707081,
0.03951543942093849,
-0.01583990827202797,
0.06325890123844147,
-0.037809696048498154,
0.0687059760093689,
0.04269913211464882,
-0.003132400568574667,
0.012480949983000755,
-2.1280301167302193e-33,
-0.009472431614995003,
-0.06488262861967087,
-0.006756027694791555,
0.10929185897111893,
0.050684574991464615,
-0.03176461160182953,
0.050478991121053696,
0.032618917524814606,
0.01859772764146328,
0.02098812349140644,
-0.05073394253849983,
0.08539663255214691,
-0.018397320061922073,
0.049997273832559586,
0.01777390018105507,
0.05584050342440605,
0.04829114302992821,
0.010506111197173595,
-0.005075204651802778,
-0.0832623690366745,
0.06207874044775963,
0.024359017610549927,
0.009175868704915047,
-0.050533343106508255,
-0.05226965993642807,
0.04763924330472946,
0.0414162240922451,
-0.029240421950817108,
-0.04216484725475311,
-0.015926627442240715,
-0.1182418093085289,
0.04658014327287674,
-0.012221712619066238,
-0.036727603524923325,
0.09439973533153534,
-0.025946956127882004,
0.0057944864965975285,
-0.03801066055893898,
0.008786141872406006,
-0.050063688308000565,
-0.02014828287065029,
0.018550051376223564,
-0.0000130511361930985,
0.03924204781651497,
-0.0014439185615628958,
-0.02092757634818554,
0.009646989405155182,
0.008009363897144794,
-0.04652360454201698,
-0.0179920457303524,
0.03272377699613571,
0.03735283389687538,
-0.049419015645980835,
0.011652419343590736,
0.06689948588609695,
0.010934657417237759,
-0.0017914942000061274,
0.038449566811323166,
-0.01619919389486313,
0.0494040846824646,
-0.036415357142686844,
0.04805978015065193,
0.054109763354063034,
-0.011584503576159477,
0.05659085139632225,
-0.037494875490665436,
-0.05890105664730072,
0.08357379585504532,
0.07055236399173737,
-0.005427094642072916,
0.015355953015387058,
0.04500781372189522,
-0.07512587308883667,
-0.004984584171324968,
-0.00934070348739624,
-0.030984805896878242,
0.05478997528553009,
-0.09184455871582031,
-0.0303820688277483,
0.083460733294487,
0.041707951575517654,
-0.07651981711387634,
0.03986406698822975,
-0.06055997684597969,
0.0009828233160078526,
-0.06360124796628952,
0.09500275552272797,
-0.03663739934563637,
0.010384056717157364,
-0.08517564088106155,
-0.061676736921072006,
0.01088043674826622,
0.030676251277327538,
-0.01418424304574728,
0.001506467000581324,
-8.620869857419263e-34,
-0.09462704509496689,
0.04924936592578888,
-0.10857238620519638,
0.05272888392210007,
-0.06804882735013962,
-0.10147587209939957,
-0.011021236889064312,
0.08796550333499908,
0.015471741557121277,
-0.003518681740388274,
-0.011988835409283638,
-0.06717315316200256,
0.08531267940998077,
-0.04240758717060089,
0.0511312372982502,
-0.046481192111968994,
0.04205397889018059,
-0.07486319541931152,
0.044196829199790955,
0.020803220570087433,
0.044739432632923126,
0.036742765456438065,
-0.16757996380329132,
-0.010878399945795536,
-0.11207787692546844,
0.04461420327425003,
-0.062871553003788,
0.06928126513957977,
-0.028236929327249527,
0.0486968457698822,
-0.02051418647170067,
-0.08073879778385162,
-0.03421953320503235,
-0.048897817730903625,
-0.06929535418748856,
-0.06788244843482971,
-0.01874227076768875,
0.00854866486042738,
0.0011396767804399133,
-0.013265717774629593,
0.05616750940680504,
-0.07708289474248886,
-0.044719334691762924,
0.026879625394940376,
-0.0307551808655262,
0.011346100829541683,
-0.03789996728301048,
-0.05778433382511139,
-0.044851548969745636,
0.03824644163250923,
0.056858792901039124,
-0.011545154266059399,
-0.04859190061688423,
-0.04651873931288719,
-0.056843142956495285,
-0.06541606783866882,
-0.014210492372512817,
-0.08699184656143188,
0.043094150722026825,
0.0000677511197864078,
-0.01982821337878704,
-0.060734495520591736,
0.09562212228775024,
-0.07374326139688492,
-0.042630575597286224,
-0.0050114477053284645,
-0.01773846335709095,
-0.01267944648861885,
0.057547468692064285,
-0.014360749162733555,
0.0610676072537899,
-0.03120524063706398,
-0.07498794794082642,
0.04339761659502983,
-0.0522654689848423,
-0.058182165026664734,
0.017840854823589325,
-0.07575490325689316,
0.0472714938223362,
-0.06435097754001617,
0.038373492658138275,
0.014704933390021324,
0.06697776168584824,
0.027050163596868515,
0.029414281249046326,
0.034502219408750534,
0.010066061280667782,
0.062017302960157394,
0.06932850927114487,
0.002476342488080263,
0.055147089064121246,
0.05543844774365425,
-0.05183984711766243,
0.041760705411434174,
-0.09801114350557327,
-4.040831669271938e-8,
-0.12020862847566605,
0.05703170970082283,
-0.03308038040995598,
0.04021415114402771,
0.02445341646671295,
-0.03341713547706604,
0.02432679943740368,
0.04137035831809044,
0.03930925577878952,
-0.01960618607699871,
0.05756641924381256,
0.012999470345675945,
0.0009181052446365356,
-0.11425209045410156,
0.013816664926707745,
0.12666301429271698,
0.07033929228782654,
-0.021556127816438675,
-0.004200576804578304,
-0.012983093038201332,
-0.009793544188141823,
0.11766921728849411,
-0.041186846792697906,
-0.009709778241813183,
0.03684494271874428,
-0.003403156530112028,
-0.057662706822156906,
0.11488792300224304,
0.04655468836426735,
0.004514607600867748,
-0.0057360888458788395,
0.06919731199741364,
-0.02010522410273552,
-0.015699107199907303,
-0.0361812599003315,
0.09708902984857559,
0.06847076117992401,
-0.06438449025154114,
0.021955454722046852,
-0.05033530667424202,
0.008924592286348343,
0.04143187776207924,
-0.07522125542163849,
0.04020629823207855,
0.0009026790503412485,
-0.039050500839948654,
0.037620414048433304,
-0.10944108664989471,
-0.007276187185198069,
-0.01090477965772152,
-0.005579643417149782,
0.014602535404264927,
-0.04523583874106407,
0.025872496888041496,
0.06135476753115654,
-0.010192175395786762,
-0.0137894656509161,
-0.08424411714076996,
-0.02926011197268963,
0.09511227905750275,
-0.014672146178781986,
0.06667445600032806,
-0.00922707375138998,
0.017634006217122078
] | 0 | 0 | 0 | 0 |
|
Understanding Warmup-Stable-Decay Learning Rates: A River Valley Loss Landscape View
|
https://openreview.net/forum?id=m51BgoqvbP
|
[
"Kaiyue Wen",
"Zhiyuan Li",
"Jason S. Wang",
"David Leo Wright Hall",
"Percy Liang",
"Tengyu Ma"
] |
Poster
|
Training language models currently requires pre-determining a fixed compute budget because the typical cosine learning rate schedule depends on the total number of steps. In contrast, the Warmup-Stable-Decay (WSD) schedule uses a constant learning rate to produce a main branch of iterates that can in principle continue indefinitely without a pre-specified compute budget. Then, given any compute budget, one can branch out from the main branch at a proper time with a rapidly decaying learning rate to produce a strong model. Empirically, WSD generates an intriguing, non-traditional loss curve: the loss remains elevated during the stable phase but sharply declines during the decay phase. Towards explaining this phenomenon, we conjecture that pretraining loss exhibits a river valley landscape, which resembles a deep valley with a river at its bottom. Under this assumption, we show that during the stable phase, the iterate undergoes large oscillations due to the high learning rate, yet it progresses swiftly along the river. During the decay phase, the rapidly dropping learning rate minimizes the iterate’s oscillations, moving it closer to the river and revealing true optimization progress. Therefore, the sustained high learning rate phase and fast decaying phase are responsible for progress in the river and the mountain directions, respectively, and are both critical. Our analysis predicts phenomenons consistent with empirical observations and shows that this landscape can naturally emerge from pretraining on a simple bi-gram dataset. Inspired by the theory, we introduce WSD-S, a variant of WSD that reuses previous checkpoints’ decay phases and keeps only one main branch, where we resume from a decayed checkpoint. WSD-S empirically outperforms WSD and Cyclic-Cosine in obtaining multiple pretrained language model checkpoints across various compute budgets in a single run for parameters scaling from 0.1B to 1.2B.
|
pretraining, language model, learning rate, loss landscape, manifold
|
We develop a river valley landscape analogy to pretraining loss landscape to explain the nontraditional loss curve of WSD learning rate and develop a variant of WSD based on the theory.
| 11,991 | null |
[
-0.028656501322984695,
-0.017027679830789566,
0.03700365498661995,
0.06802379339933395,
0.04500274732708931,
-0.07444456964731216,
-0.06821952760219574,
-0.041438955813646317,
0.07319530099630356,
0.000444583420176059,
-0.06625152379274368,
-0.004287100397050381,
-0.035901620984077454,
-0.031965285539627075,
-0.04047706723213196,
0.011233773082494736,
-0.03425782546401024,
0.027314744889736176,
-0.06626702100038528,
-0.06314639002084732,
0.016542043536901474,
0.037385500967502594,
-0.0031535960733890533,
0.041034262627363205,
0.09003329277038574,
0.005731559358537197,
-0.016666624695062637,
0.013563686981797218,
0.05214505270123482,
-0.002092665294185281,
0.027306441217660904,
0.04179896041750908,
-0.009654134511947632,
-0.03976253420114517,
0.02480061911046505,
0.11958536505699158,
-0.002338673919439316,
-0.04414350911974907,
0.0038743566256016493,
0.04982944205403328,
-0.012474000453948975,
0.08296568691730499,
0.046747319400310516,
0.03149857744574547,
0.052727073431015015,
0.023867297917604446,
0.021002953872084618,
-0.05296429619193077,
-0.048698507249355316,
-0.017104050144553185,
-0.009424339979887009,
-0.08747155964374542,
-0.006262634880840778,
0.07231659442186356,
-0.08934997767210007,
0.054729804396629333,
0.019306901842355728,
0.04985370859503746,
-0.05331343039870262,
-0.0009293581824749708,
-0.06851720809936523,
-0.09034286439418793,
-0.07261413335800171,
-0.008615982718765736,
-0.008296975865960121,
-0.00014010084851179272,
0.04276144877076149,
0.046979885548353195,
-0.005835251417011023,
0.023252280429005623,
0.00013049649714957923,
0.0653311088681221,
-0.04971407726407051,
-0.02307330071926117,
0.03406849876046181,
-0.00590286310762167,
0.08116137981414795,
-0.0019937115721404552,
0.0562969334423542,
-0.0927601084113121,
0.040571052581071854,
-0.03030005469918251,
0.03156200423836708,
-0.11815740168094635,
0.06753214448690414,
0.004643390420824289,
0.08156514167785645,
-0.007020909339189529,
0.08199433237314224,
-0.012242124415934086,
-0.06121589615941048,
-0.0009684823453426361,
-0.014365512877702713,
-0.06259537488222122,
0.029221946373581886,
0.08237063884735107,
0.02806512638926506,
-0.0036531498190015554,
-0.04508537799119949,
0.07591727375984192,
-0.02218654938042164,
0.02145538106560707,
0.07422610372304916,
0.01102184783667326,
0.05588280037045479,
-0.039651960134506226,
0.022985180839896202,
0.09230486303567886,
0.07368336617946625,
-0.01726388745009899,
0.039060208946466446,
0.03439056873321533,
0.0501846969127655,
0.06915798783302307,
0.05134986340999603,
0.06530556082725525,
-0.03579551726579666,
-0.03536498546600342,
-0.031737592071294785,
0.08705497533082962,
-0.04718399792909622,
0.011782988905906677,
-0.07317854464054108,
0.0545906126499176,
0.013599292375147343,
-0.03381531685590744,
-0.007153193932026625,
8.115744322836579e-34,
0.04798073321580887,
-0.018693532794713974,
-0.024885565042495728,
-0.06339588016271591,
0.008569872938096523,
-0.0992317795753479,
0.002632714109495282,
0.0009362661512568593,
-0.02593420073390007,
-0.005012999754399061,
-0.03333807736635208,
-0.012660899199545383,
-0.05623241141438484,
0.07740715891122818,
0.03403569757938385,
-0.08852681517601013,
-0.039318207651376724,
0.001772510353475809,
0.010519670322537422,
-0.052090488374233246,
0.07363639026880264,
-0.06425896286964417,
-0.023892246186733246,
-0.138175368309021,
-0.038184668868780136,
0.016567016020417213,
0.04902580752968788,
0.023038608953356743,
-0.08307477831840515,
0.03257688507437706,
-0.06703384965658188,
-0.028077764436602592,
-0.035106297582387924,
-0.0050066797994077206,
0.045926857739686966,
-0.06542321294546127,
0.024302596226334572,
0.028163248673081398,
0.06038110330700874,
-0.005021595861762762,
0.006982431281358004,
0.01637217029929161,
0.031822655349969864,
0.043995991349220276,
-0.012438149191439152,
-0.10208147764205933,
0.11995742470026016,
-0.009699967689812183,
-0.122976154088974,
-0.011284883134067059,
-0.010257614776492119,
-0.037887170910835266,
-0.07380478829145432,
0.028710542246699333,
-0.03326381370425224,
0.04973194748163223,
0.08990325778722763,
-0.024910971522331238,
-0.00461203046143055,
0.07776936888694763,
0.021275293081998825,
0.025795524939894676,
-0.016190579161047935,
0.06893555074930191,
0.0001569198939250782,
0.059204570949077606,
0.011505624279379845,
0.05970040708780289,
0.06263362616300583,
-0.03372730314731598,
-0.03942637890577316,
-0.01612236350774765,
-0.018496178090572357,
-0.05650510638952255,
0.08416272699832916,
-0.02095232531428337,
0.09458150714635849,
-0.0626407340168953,
0.009332011453807354,
0.04440847039222717,
0.010508675128221512,
-0.06676752865314484,
-0.025276750326156616,
-0.042277127504348755,
-0.021637186408042908,
-0.0698978379368782,
0.07389618456363678,
-0.15895645320415497,
-0.03500860929489136,
-0.019253462553024292,
-0.0742194801568985,
-0.012405169196426868,
0.07244967669248581,
-0.005258789751678705,
0.04627428576350212,
-1.7284891377773242e-33,
-0.01983834244310856,
-0.01639421470463276,
-0.061294157058000565,
0.0960700586438179,
-0.07130929827690125,
-0.013725122436881065,
0.007653862237930298,
0.006683524698019028,
-0.060721833258867264,
-0.08914119005203247,
-0.023391613736748695,
0.0023728888481855392,
-0.04967813566327095,
0.035575300455093384,
0.07382278144359589,
-0.05509110167622566,
0.08344951272010803,
0.04452693089842796,
0.11219460517168045,
0.033642321825027466,
-0.00742728728801012,
0.04931020736694336,
-0.13739344477653503,
-0.05223685875535011,
0.007270632311701775,
0.033000875264406204,
-0.07739126682281494,
0.05097683519124985,
0.03961631655693054,
0.03199860453605652,
-0.05775110796093941,
-0.023871148005127907,
0.0014728083042427897,
-0.03713051974773407,
-0.10703463107347488,
0.10437935590744019,
0.021483099088072777,
-0.040598541498184204,
-0.008395138196647167,
0.06287340074777603,
0.14883968234062195,
-0.03377550095319748,
0.00236983853392303,
-0.016319382935762405,
0.0008722121128812432,
-0.009628703817725182,
-0.06708948314189911,
0.06998473405838013,
-0.048787571489810944,
0.07144243270158768,
0.04617924243211746,
0.0320519395172596,
-0.08757228404283524,
0.07296255975961685,
-0.03903963789343834,
-0.043944135308265686,
-0.010163497179746628,
0.005929794628173113,
-0.03414855897426605,
-0.032872315496206284,
-0.08083681017160416,
-0.007482117973268032,
-0.030489858239889145,
0.009459941647946835,
0.048595212399959564,
-0.029624350368976593,
-0.035097718238830566,
0.008920961059629917,
0.03921502083539963,
0.018647948279976845,
0.03806475177407265,
0.03481905907392502,
-0.025487106293439865,
-0.037037890404462814,
-0.03204717859625816,
-0.003200835781171918,
-0.05203913152217865,
-0.033995334059000015,
-0.04275394231081009,
-0.028527135029435158,
-0.02977662719786167,
0.09435156732797623,
-0.01798289082944393,
0.050428055226802826,
0.05124709755182266,
-0.01754867471754551,
-0.03273443505167961,
0.013632900081574917,
0.0221958439797163,
0.00980424415320158,
0.055362049490213394,
-0.047416336834430695,
-0.056862760335206985,
0.03639478236436844,
-0.061916135251522064,
-4.2821135082249384e-8,
-0.07502646744251251,
0.05136740952730179,
-0.03584803640842438,
0.0368683747947216,
0.15708576142787933,
-0.06216973438858986,
0.03528046980500221,
0.045069899410009384,
-0.018646307289600372,
-0.03217511624097824,
0.09472419321537018,
0.014486478641629219,
-0.01658439449965954,
-0.040808893740177155,
-0.01988428458571434,
0.07390464842319489,
-0.01239690836519003,
-0.0339798741042614,
-0.04011233150959015,
-0.06649820506572723,
-0.009768986143171787,
-0.022523874416947365,
0.038815394043922424,
0.006500844843685627,
-0.023012543097138405,
-0.060801032930612564,
-0.0005675280699506402,
0.07799806445837021,
0.06191108748316765,
-0.029346374794840813,
-0.03768015280365944,
0.05977267026901245,
-0.012102981097996235,
-0.003203927306458354,
0.03213096410036087,
0.07949448376893997,
-0.009476585313677788,
-0.020265739411115646,
0.045358456671237946,
0.031488291919231415,
-0.028291404247283936,
0.07554284483194351,
-0.05654402822256088,
0.00023024696565698832,
0.0656072273850441,
0.026384660974144936,
0.02785615622997284,
-0.06947048753499985,
0.03949190676212311,
-0.025468379259109497,
-0.00010708004992920905,
0.019476355984807014,
0.0015069302171468735,
0.05865130200982094,
0.0822167620062828,
-0.03118315152823925,
-0.048972051590681076,
-0.06467769294977188,
-0.07342690229415894,
0.09865593910217285,
-0.0008692737319506705,
-0.005085885524749756,
-0.0480327308177948,
-0.0236224252730608
] | 0 | 0 | 0 | 0 |
|
ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning
|
https://openreview.net/forum?id=kSdWcw5mkp
|
[
"Ruchika Chavhan",
"Da Li",
"Timothy Hospedales"
] |
Poster
|
While large-scale text-to-image diffusion models have demonstrated impressive image-generation capabilities, there are significant concerns about their potential misuse for generating unsafe content, violating copyright, and perpetuating societal biases. Recently, the text-to-image generation community has begun addressing these concerns by editing or unlearning undesired concepts from pre-trained models. However, these methods often involve data-intensive and inefficient fine-tuning or utilize various forms of token remapping, rendering them susceptible to adversarial jailbreaks. In this paper, we present a simple and effective training-free approach, ConceptPrune, wherein we first identify critical regions within pre-trained models responsible for generating undesirable concepts, thereby facilitating straightforward concept unlearning via weight pruning. Experiments across a range of concepts including artistic styles, nudity, and object erasure demonstrate that target concepts can be efficiently erased by pruning a tiny fraction, approximately 0.12% of total weights, enabling multi-concept erasure and robustness against various white-box and black-box adversarial attacks.
|
diffusion models, concept editing, pruning
|
Training free Concept editing via neuron pruning
| 11,989 |
2405.19237
|
[
-0.0011192768579348922,
-0.04658222943544388,
0.010905967094004154,
0.09953135251998901,
0.07148373126983643,
-0.017758628353476524,
-0.07258349657058716,
-0.057405900210142136,
0.025214334949851036,
-0.0031045940704643726,
-0.01521170511841774,
-0.01175736915320158,
0.037136245518922806,
-0.026560191065073013,
-0.039489973336458206,
-0.02882445603609085,
0.07439612597227097,
0.12547823786735535,
-0.1469666063785553,
-0.018919972702860832,
0.026052378118038177,
-0.016446080058813095,
0.04446331411600113,
-0.021782396361231804,
0.07693259418010712,
0.025543341413140297,
-0.019107414409518242,
-0.026722794398665428,
0.060646750032901764,
-0.07108493149280548,
0.011484052985906601,
-0.05476660281419754,
-0.033713798969984055,
0.06355410069227219,
-0.07602758705615997,
0.06313508003950119,
-0.05805283412337303,
0.029109252616763115,
0.05280713737010956,
-0.05622289329767227,
0.01285229530185461,
0.00234780996106565,
-0.0621500201523304,
0.049560774117708206,
0.0737316906452179,
0.002896711230278015,
-0.028390254825353622,
-0.0010188771411776543,
-0.03263729810714722,
-0.024428993463516235,
0.01597505435347557,
-0.027145812287926674,
-0.08046197891235352,
0.08154202997684479,
-0.04853184521198273,
-0.012211054563522339,
0.03516906499862671,
0.029238181188702583,
0.005116719752550125,
-0.026877528056502342,
-0.05638604983687401,
-0.0011564751621335745,
-0.02424716390669346,
-0.028446821495890617,
0.010228821076452732,
-0.06364519149065018,
0.04801637679338455,
0.08199514448642731,
0.0015619740588590503,
0.03430697321891785,
0.15858210623264313,
0.047417666763067245,
-0.037789445370435715,
0.07485270500183105,
0.04127923771739006,
0.05449281632900238,
-0.008140819147229195,
-0.014833826571702957,
-0.042031656950712204,
-0.09129209816455841,
0.06933819502592087,
-0.030659224838018417,
0.06748101860284805,
0.03718189150094986,
0.043195076286792755,
0.04407314956188202,
-0.024373134598135948,
0.05356685444712639,
0.061147987842559814,
-0.024435285478830338,
-0.07448825985193253,
0.008618218824267387,
0.0775400847196579,
-0.01701381243765354,
0.04814108833670616,
-0.027754947543144226,
-0.04963766410946846,
-0.0555143877863884,
-0.0857371836900711,
0.09976501017808914,
-0.038626402616500854,
0.06538672000169754,
-0.008084192872047424,
-0.06231633201241493,
0.08991006761789322,
-0.02939670719206333,
0.10375238955020905,
0.013788441196084023,
-0.009829295799136162,
0.029105449095368385,
0.03896550089120865,
0.056495774537324905,
-0.06413879245519638,
-0.003992479294538498,
0.10726574808359146,
-0.08207318186759949,
0.0024931433144956827,
-0.009440939873456955,
0.005570403765887022,
0.07290122658014297,
0.021945839747786522,
0.05765931308269501,
-0.05402006581425667,
0.008151724934577942,
-0.01740565337240696,
-0.09086386114358902,
-0.10788565129041672,
1.4693788669519022e-33,
0.014495865441858768,
0.01923605427145958,
0.02826724201440811,
0.04134437069296837,
0.11574070155620575,
0.0361739844083786,
-0.020403359085321426,
0.03740282356739044,
-0.02513996884226799,
-0.03806662559509277,
-0.025844700634479523,
-0.010488944128155708,
-0.06774573028087616,
0.0852750763297081,
0.07218743115663528,
-0.015442951582372189,
-0.04900924488902092,
-0.0022149160504341125,
0.02357540838420391,
0.02531273476779461,
-0.017282871529459953,
-0.019922686740756035,
-0.040255218744277954,
-0.035323403775691986,
-0.08571618795394897,
0.013807784765958786,
0.05031111463904381,
-0.08860796689987183,
0.014771266840398312,
0.03536631166934967,
-0.14647473394870758,
0.028785904869437218,
0.06667128205299377,
0.007248867768794298,
0.013121450319886208,
-0.044108860194683075,
0.0053289467468857765,
-0.04357712343335152,
0.03642057627439499,
0.030640006065368652,
-0.06471827626228333,
0.013918774202466011,
-0.04697997123003006,
-0.06268840283155441,
-0.044869810342788696,
0.03209631145000458,
0.058468010276556015,
0.03662635758519173,
-0.05898210406303406,
-0.019070683047175407,
0.04120723903179169,
0.04486847296357155,
-0.002376047894358635,
-0.03419362008571625,
-0.10504098236560822,
-0.04573174938559532,
-0.003965891897678375,
-0.07560400664806366,
0.00962028093636036,
-0.03597644343972206,
0.028947580605745316,
0.11683283746242523,
-0.02863180823624134,
0.10371853411197662,
0.016460251063108444,
-0.00008528002217644826,
0.002902837935835123,
0.04483221098780632,
-0.004836725536733866,
-0.01167806051671505,
-0.10124810039997101,
0.023615457117557526,
-0.08644158393144608,
-0.056688256561756134,
-0.06897521018981934,
-0.005164425354450941,
0.09863273799419403,
-0.0029256485868245363,
-0.00033919996349141,
-0.02855796553194523,
-0.028156043961644173,
-0.08715181052684784,
0.06962846964597702,
-0.04425371438264847,
-0.09516924619674683,
-0.055396635085344315,
0.0689592957496643,
-0.029297444969415665,
-0.0022142911329865456,
0.013673491775989532,
0.0037276693619787693,
-0.019135799258947372,
-0.05405847355723381,
-0.003290500957518816,
0.002660301048308611,
-1.032955650716004e-33,
-0.009049653075635433,
0.004584011621773243,
-0.09460236877202988,
0.06750030070543289,
-0.02015943080186844,
-0.008662736974656582,
-0.03905283659696579,
0.0764467865228653,
-0.005989087745547295,
-0.0645507201552391,
-0.07563230395317078,
-0.02845940738916397,
-0.00977381318807602,
-0.05992954224348068,
-0.02493629790842533,
-0.09272903949022293,
0.0041827186942100525,
0.02612336166203022,
-0.043268706649541855,
0.04389926791191101,
0.025553317740559578,
0.07197660952806473,
-0.08088874071836472,
0.0562889389693737,
-0.03818785771727562,
0.02312025986611843,
-0.0342838317155838,
0.029736988246440887,
0.00356254237703979,
0.010518002323806286,
-0.046261802315711975,
-0.01603853888809681,
0.03344201669096947,
-0.04512944445014,
-0.09885157644748688,
0.06457473337650299,
0.06736510246992111,
0.036330774426460266,
0.010979149490594864,
0.08879809081554413,
0.054603077471256256,
-0.0318509005010128,
-0.08954337984323502,
-0.004775157198309898,
-0.03864884376525879,
-0.10045205056667328,
-0.07032248377799988,
0.0028015782590955496,
0.053615130484104156,
0.023933537304401398,
0.002166085410863161,
-0.015880722552537918,
-0.0721852108836174,
-0.012163442559540272,
-0.02061895653605461,
-0.0629800483584404,
-0.026171335950493813,
-0.00960324052721262,
0.07743269950151443,
0.03290762007236481,
-0.033214375376701355,
-0.046896371990442276,
-0.03198251128196716,
-0.09295579046010971,
-0.007506806403398514,
-0.08645625412464142,
-0.0291123166680336,
0.08002141118049622,
0.0019421656616032124,
0.023032641038298607,
-0.03784458711743355,
0.019461022689938545,
0.010143987834453583,
-0.012648962438106537,
-0.055171579122543335,
-0.05459275469183922,
0.02799288183450699,
-0.0024800754617899656,
0.028345361351966858,
-0.07662180066108704,
0.05030373856425285,
-0.0258202962577343,
0.0016628806479275227,
0.03874479979276657,
0.10657063871622086,
0.06930605322122574,
-0.008938084356486797,
0.07299666851758957,
0.016908692196011543,
0.07018857449293137,
-0.010767482221126556,
0.001607000594958663,
0.011630181223154068,
0.08529651910066605,
-0.07747818529605865,
-3.723252461895754e-8,
-0.028198465704917908,
0.017806431278586388,
0.019149692729115486,
0.11307211965322495,
0.04555346816778183,
-0.005773442331701517,
-0.03385542333126068,
0.06105400621891022,
-0.008803915232419968,
-0.023375017568469048,
0.02731527015566826,
0.012159346602857113,
-0.036821749061346054,
-0.02992107905447483,
-0.06760457158088684,
0.07832519710063934,
-0.025113001465797424,
0.03714970871806145,
0.0057029761373996735,
-0.006254185922443867,
-0.01529077161103487,
0.016921808943152428,
-0.008463643491268158,
-0.05427928268909454,
0.02473808452486992,
-0.05724190175533295,
-0.023295534774661064,
0.04311390221118927,
0.03157010301947594,
0.05600031465291977,
-0.028546778485178947,
-0.0011422319803386927,
0.01740732416510582,
0.015294017270207405,
0.01685480587184429,
0.08165251463651657,
0.029305147007107735,
0.05834060534834862,
-0.042954228818416595,
-0.028487160801887512,
0.05953216552734375,
0.06048900634050369,
0.07051413506269455,
0.010466784238815308,
0.0033219088800251484,
-0.030301949009299278,
0.04723132774233818,
-0.03868195414543152,
0.09930993616580963,
-0.013117571361362934,
-0.016041412949562073,
0.038453150540590286,
0.006231548730283976,
0.07267057150602341,
0.013485291041433811,
0.006696786731481552,
0.008180258795619011,
0.06610993295907974,
-0.01705407351255417,
0.11239051073789597,
0.11997759342193604,
-0.013029857538640499,
-0.007805262226611376,
0.019437825307250023
] |
https://github.com/ruchikachavhan/concept-prune
| 18 | 0 | 0 | 0 |
Q-SFT: Q-Learning for Language Models via Supervised Fine-Tuning
|
https://openreview.net/forum?id=v4MTnPiYXY
|
[
"Joey Hong",
"Anca Dragan",
"Sergey Levine"
] |
Poster
|
Value-based reinforcement learning (RL) can in principle learn effective policies for a wide range of multi-turn problems, from games to dialogue to robotic control, including via offline RL from static previously collected datasets. However, despite the widespread use of policy gradient methods to train large language models for single turn tasks (e.g., question answering), value-based methods for multi-turn RL in an off-policy or offline setting have proven particularly challenging to scale to the setting of large language models. This setting requires effectively leveraging pretraining, scaling to large architectures with billions of parameters, and training on large datasets, all of which represent major challenges for current value-based RL methods. In this work, we propose a novel offline RL algorithm that addresses these drawbacks, casting Q-learning as a modified supervised fine-tuning (SFT) problem where the probabilities of tokens directly translate to Q-values. In this way we obtain an algorithm that smoothly transitions from maximizing the likelihood of the data during pretraining to learning a near-optimal Q-function during finetuning. Our algorithm has strong theoretical foundations, enjoying performance bounds similar to state-of-the-art Q-learning methods, while in practice utilizing an objective that closely resembles SFT. Because of this, our approach can enjoy the full benefits of the pretraining of language models, without the need to reinitialize any weights before RL finetuning, and without the need to initialize new heads for predicting values or advantages. Empirically, we evaluate our method on both pretrained LLMs and VLMs, on a variety of tasks including both natural language dialogue and robotic manipulation and navigation from images.
|
offline reinforcement learning, language models, dialogue, robotics
|
We present a new offline RL algorithm specifically to fine-tune pretrained LLMs and VLMs better.
| 11,981 | null |
[
-0.033578045666217804,
-0.05204983800649643,
0.044532403349876404,
0.024193650111556053,
0.01789618656039238,
0.08436167240142822,
0.048435959964990616,
0.03689749911427498,
0.00820520892739296,
0.006595428101718426,
-0.07698249816894531,
-0.05109398066997528,
-0.009569636546075344,
0.055508654564619064,
0.061849504709243774,
0.04986746236681938,
0.07032547891139984,
0.04442385211586952,
-0.1264140009880066,
-0.040570151060819626,
0.024990690872073174,
0.05158698558807373,
0.09317111223936081,
0.0030186674557626247,
0.006144537590444088,
-0.0005240305908955634,
0.012431907467544079,
-0.035676755011081696,
-0.0003862402809318155,
0.004450005013495684,
-0.02057107351720333,
0.011560946702957153,
-0.05183093622326851,
-0.029304297640919685,
-0.06528845429420471,
0.09555219858884811,
-0.09604485332965851,
-0.0201988834887743,
-0.025579171255230904,
-0.025206830352544785,
-0.033861543983221054,
0.007626283913850784,
0.024247338995337486,
0.025217382237315178,
0.11017677932977676,
0.006959185469895601,
-0.03309173509478569,
-0.024515265598893166,
0.02318580076098442,
-0.00018604322394821793,
-0.06398390978574753,
-0.04986780509352684,
0.021162347868084908,
0.07016799598932266,
-0.004108359571546316,
0.04044570401310921,
0.05767858773469925,
0.015015429817140102,
0.0790666863322258,
-0.0704772025346756,
-0.0651688352227211,
-0.16274750232696533,
-0.02371491864323616,
-0.027695240452885628,
-0.11745952069759369,
-0.05425431951880455,
-0.018180863931775093,
0.014554007910192013,
-0.0025603065732866526,
0.040074434131383896,
-0.06279011815786362,
0.07131725549697876,
0.03676726669073105,
0.005538383033126593,
0.01340793538838625,
-0.016507208347320557,
-0.018001267686486244,
-0.017020493745803833,
0.007808719761669636,
-0.05034533515572548,
0.008240506052970886,
-0.05541902408003807,
0.02309322915971279,
0.04672962799668312,
0.13655662536621094,
-0.0018510681111365557,
0.011988667771220207,
0.06295675039291382,
0.08921366184949875,
0.05517371743917465,
-0.03326963260769844,
-0.0036057420074939728,
0.04931129887700081,
-0.03141012042760849,
-0.027819490060210228,
0.01892685331404209,
-0.06819715350866318,
-0.1551199108362198,
-0.01597939431667328,
0.07035928219556808,
0.015156416222453117,
0.1053871214389801,
-0.002647094428539276,
0.015076006762683392,
-0.024561090394854546,
-0.010159014724195004,
0.01684216409921646,
0.05592324584722519,
0.022510338574647903,
-0.0841856449842453,
0.025493977591395378,
0.026367340236902237,
0.05221473425626755,
0.03719182312488556,
0.03322415426373482,
0.0314653217792511,
0.013980815187096596,
-0.0402197502553463,
-0.06908800452947617,
0.07029827684164047,
-0.01699463278055191,
-0.06602830439805984,
-0.01900602877140045,
0.012437282130122185,
-0.004417167045176029,
-0.020494572818279266,
-0.03385525941848755,
-8.072867248035374e-34,
0.09165187925100327,
0.017728306353092194,
0.06010011211037636,
0.013218632899224758,
0.030261969193816185,
-0.0075187296606600285,
0.05280653014779091,
0.048004355281591415,
-0.03193369135260582,
-0.019840309396386147,
0.03011498600244522,
-0.02266916260123253,
-0.0677151158452034,
0.004692955873906612,
0.004851642530411482,
-0.002957323333248496,
-0.06441721320152283,
-0.0011117919348180294,
-0.030021997168660164,
-0.021144863218069077,
0.09985287487506866,
-0.0352589450776577,
0.009292027913033962,
-0.016214603558182716,
0.07329715043306351,
-0.0034785328898578882,
0.06825432181358337,
-0.06368647515773773,
-0.012081107124686241,
-0.02050609141588211,
-0.09843961149454117,
-0.05505013465881348,
-0.010291381739079952,
0.07699030637741089,
0.009689204394817352,
-0.00673047685995698,
-0.0971074178814888,
-0.005702255759388208,
-0.0329054519534111,
0.01984412781894207,
-0.0067647830583155155,
-0.01676601730287075,
0.05241706594824791,
-0.01664031483232975,
-0.052406780421733856,
-0.09715073555707932,
0.009731575846672058,
-0.01648712530732155,
-0.04255157709121704,
-0.0781334862112999,
0.00598275987431407,
0.06254947930574417,
-0.04784579947590828,
-0.03338072821497917,
-0.02360658347606659,
-0.02183130942285061,
0.013734403997659683,
0.05980230122804642,
-0.10050533711910248,
0.020945142954587936,
0.07931812852621078,
0.01046871580183506,
0.008073254488408566,
0.11699427664279938,
-0.008678608573973179,
0.07042239606380463,
0.01603294163942337,
0.08251448720693588,
0.0640297681093216,
0.004544774070382118,
0.007005200255662203,
0.020831510424613953,
0.04343534633517265,
-0.04909197613596916,
0.0495903305709362,
0.0538187250494957,
0.022348197177052498,
-0.1030275970697403,
0.05653485283255577,
0.016587311401963234,
0.012595991604030132,
-0.015745945274829865,
-0.0487554632127285,
-0.03703306242823601,
0.04427768662571907,
-0.1303528994321823,
0.017862550914287567,
-0.10044560581445694,
-0.005339795257896185,
-0.014813801273703575,
-0.12169907987117767,
-0.037540849298238754,
-0.06123591214418411,
0.024851450696587563,
0.0010036269668489695,
4.491402054429982e-35,
0.010135378688573837,
-0.07446999102830887,
-0.0017020622035488486,
0.08221087604761124,
-0.06326901167631149,
-0.0327327623963356,
-0.03135301172733307,
0.006499307230114937,
0.0525045171380043,
-0.03506778925657272,
-0.0822816714644432,
-0.009427644312381744,
0.037961166352033615,
-0.021839480847120285,
0.0541839562356472,
-0.03734786808490753,
-0.023085344582796097,
0.012011924758553505,
0.06743543595075607,
0.073053278028965,
-0.015502615831792355,
0.04785355553030968,
-0.06593559682369232,
0.07059630006551743,
-0.0478789359331131,
0.013254839926958084,
-0.01914036273956299,
0.03702019900083542,
0.007123277522623539,
0.06786992400884628,
-0.08368335664272308,
-0.06201975792646408,
0.023304155096411705,
-0.025008752942085266,
-0.08325796574354172,
0.06558013707399368,
0.08332391828298569,
-0.038215991109609604,
-0.05920553207397461,
0.14615550637245178,
0.08194081485271454,
-0.08241403102874756,
-0.10971441864967346,
0.031379975378513336,
-0.006624032277613878,
-0.03547719866037369,
-0.08455978333950043,
0.0033357294742017984,
-0.011592389084398746,
0.002338557504117489,
-0.022007912397384644,
0.03255258500576019,
-0.04644588753581047,
-0.000617302895989269,
-0.08105598390102386,
-0.06401138007640839,
-0.0754549577832222,
-0.0038083246909081936,
0.03070935420691967,
0.008716384880244732,
-0.04894290864467621,
-0.0112049151211977,
0.007136255502700806,
0.011999507434666157,
0.009811869822442532,
-0.019075293093919754,
0.008797590620815754,
0.03462989628314972,
0.02936447784304619,
-0.026951100677251816,
-0.022771678864955902,
-0.01937561295926571,
0.07190586626529694,
-0.0031406194902956486,
-0.018431413918733597,
-0.021916253492236137,
0.005923214368522167,
-0.06973137706518173,
0.022923016920685768,
-0.027899308130145073,
0.08502621948719025,
0.007692027371376753,
0.03804131597280502,
0.11188488453626633,
0.0656277984380722,
0.04406076297163963,
0.0390421487390995,
0.12527289986610413,
0.05094519630074501,
-0.003743317211046815,
0.0016603500116616488,
0.02688264101743698,
0.04520095884799957,
0.04217180982232094,
-0.07113593071699142,
-3.649688906648407e-8,
-0.060266654938459396,
0.021155713126063347,
0.06011227145791054,
0.09369619935750961,
0.05187418311834335,
-0.01763678342103958,
-0.02196415513753891,
0.05227839946746826,
-0.04722719267010689,
0.024948863312602043,
-0.009151153266429901,
0.06156627833843231,
-0.007009660359472036,
-0.1005597785115242,
-0.05233472213149071,
0.08213305473327637,
-0.00013467983808368444,
-0.00743600819259882,
0.02077035792171955,
0.0071442145854234695,
0.06297105550765991,
0.005762259475886822,
0.013504445552825928,
-0.018400344997644424,
-0.014202121645212173,
-0.042837679386138916,
-0.012201189994812012,
0.0635136291384697,
0.07728419452905655,
-0.02685428224503994,
-0.00452202744781971,
-0.06519494950771332,
0.04774708300828934,
-0.007542407140135765,
0.024629240855574608,
0.03734409436583519,
0.035947006195783615,
-0.08194561302661896,
0.0424748994410038,
-0.03413823992013931,
0.01927197352051735,
0.10241372138261795,
-0.0525234118103981,
0.00241778208874166,
-0.01720644347369671,
-0.0014887836296111345,
-0.11803314089775085,
-0.08121369779109955,
0.04066799581050873,
-0.07243536412715912,
0.007192652206867933,
0.0019008596427738667,
0.005413141567260027,
0.06642603129148483,
0.07332046329975128,
0.03932487219572067,
0.009324520826339722,
-0.01651955023407936,
0.036007825285196304,
0.09610341489315033,
-0.03823937848210335,
-0.018361609429121017,
-0.03945951163768768,
0.02036023885011673
] | 0 | 0 | 0 | 0 |
|
MM-EMBED: UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS
|
https://openreview.net/forum?id=i45NQb2iKO
|
[
"Sheng-Chieh Lin",
"Chankyu Lee",
"Mohammad Shoeybi",
"Jimmy Lin",
"Bryan Catanzaro",
"Wei Ping"
] |
Poster
|
State-of-the-art retrieval models typically address a straightforward search scenario, in which retrieval tasks are fixed (e.g., finding a passage to answer a specific question) and only a single modality is supported for both queries and retrieved results. This paper introduces techniques for advancing information retrieval with multimodal large language models (MLLMs), enabling a broader search scenario, termed universal multimodal retrieval, where multiple modalities and diverse retrieval tasks are accommodated. To this end, we first study fine-tuning an MLLM as a bi-encoder retriever on 10 datasets with 16 retrieval tasks. Our empirical results show that the fine-tuned MLLM retriever is capable of understanding challenging queries, composed of both text and image, but it underperforms compared to a smaller CLIP retriever in cross-modal retrieval tasks due to the modality bias exhibited by MLLMs. To address the issue, we propose modality-aware hard negative mining to mitigate the modality bias exhibited by MLLM retrievers. Second, we propose continuously fine-tuning the universal multimodal retriever to enhance its text retrieval capability while preserving multimodal retrieval capability. As a result, our model, MM-Embed, achieves state-of-the-art performance on the multimodal retrieval benchmark M-BEIR, which spans multiple domains and tasks, while also surpassing the state-of-the-art text retrieval model, NV-Embed-v1, on the MTEB retrieval benchmark. Finally, we explore prompting the off-the-shelf MLLMs as zero-shot rerankers to refine the ranking of the candidates from the multimodal retriever. We find that, through prompt-and-reranking, MLLMs can further improve multimodal retrieval when the user queries (e.g., text-image composed queries) are more complex and challenging to understand. These findings also pave the way for advancing universal multimodal retrieval in the future. We release the model weights at: https://huggingface.co/nvidia/MM-Embed.
|
multimodal, embedding model, retriever, LLM
| null | 11,980 | null |
[
-0.005496556404978037,
-0.15007704496383667,
0.04822215065360069,
0.013691713102161884,
0.09527796506881714,
-0.03399258852005005,
-0.02344644069671631,
0.007552712224423885,
0.033016420900821686,
-0.06407225877046585,
0.027353927493095398,
-0.06440453976392746,
0.04376320540904999,
0.0653788223862648,
-0.020855415612459183,
0.0576610192656517,
0.09106111526489258,
0.0680355355143547,
-0.12842930853366852,
-0.009767071343958378,
0.04614553228020668,
0.011976542882621288,
0.05405246838927269,
-0.067843958735466,
0.008779756724834442,
0.02654947340488434,
-0.052609823644161224,
-0.049909818917512894,
0.04601609706878662,
-0.03174608573317528,
-0.004385588224977255,
0.05578869581222534,
-0.005637664347887039,
0.09285526722669601,
-0.024794716387987137,
0.029608868062496185,
-0.10162317752838135,
0.026868324726819992,
-0.0002557830302976072,
-0.0618150420486927,
-0.01915564201772213,
0.016217589378356934,
0.013003120198845863,
-0.0275561660528183,
0.03567037358880043,
-0.019183186814188957,
-0.058501869440078735,
-0.05781148746609688,
0.07835663855075836,
0.024249911308288574,
-0.05235976725816727,
0.012204053811728954,
-0.012875697575509548,
0.07816998660564423,
0.011740271002054214,
-0.06887193769216537,
0.02052890695631504,
0.010291649959981441,
-0.007798319682478905,
-0.0471823588013649,
-0.002212578896433115,
-0.10219485312700272,
-0.022064456716179848,
-0.039565373212099075,
0.038939766585826874,
-0.053953129798173904,
0.023756401613354683,
-0.03882961720228195,
-0.005638653878122568,
-0.029981136322021484,
0.0002493805659469217,
-0.03225883096456528,
-0.05169760808348656,
0.032998234033584595,
0.011387782171368599,
0.08866902440786362,
0.04035310819745064,
-0.023296868428587914,
0.03239043802022934,
-0.03929885849356651,
0.03428947553038597,
0.000614772317931056,
0.07518871128559113,
0.003667264012619853,
0.15706638991832733,
-0.03733513504266739,
0.006295070052146912,
0.00786706991493702,
0.03942057490348816,
0.042198482900857925,
-0.06445301324129105,
-0.028044233098626137,
0.01036788523197174,
-0.021858932450413704,
0.04405957832932472,
0.009973514825105667,
0.05294907093048096,
-0.023290056735277176,
0.001971921417862177,
0.056960511952638626,
0.04869302734732628,
0.09244199842214584,
-0.01323944702744484,
-0.009257668629288673,
-0.012866360135376453,
-0.013987436890602112,
0.07470373809337616,
0.05215463414788246,
0.04740867763757706,
-0.10250522196292877,
0.008067592978477478,
0.048507172614336014,
0.08537307381629944,
-0.00475720688700676,
0.03616347163915634,
-0.057984184473752975,
0.05917076766490936,
0.02180289290845394,
0.030193408951163292,
-0.010076640173792839,
-0.05339984595775604,
0.04227818176150322,
-0.036961913108825684,
-0.06991088390350342,
0.04805140197277069,
0.053962383419275284,
-0.03711631894111633,
-8.216402455110468e-34,
0.042529333382844925,
0.0020303307101130486,
0.03126426041126251,
0.07407672703266144,
0.03870999068021774,
-0.030733076855540276,
-0.05280013754963875,
0.12438251078128815,
-0.08435504138469696,
-0.02240334264934063,
-0.031352099031209946,
0.03733525052666664,
-0.040863897651433945,
0.07249840348958969,
-0.026944328099489212,
0.025075534358620644,
-0.05592336133122444,
-0.07443787157535553,
0.020666515454649925,
0.041879113763570786,
0.020746318623423576,
-0.04654800519347191,
0.009774633683264256,
0.006325585767626762,
0.02349352277815342,
-0.01676121912896633,
0.0923488661646843,
-0.11228078603744507,
-0.0800238773226738,
0.004942350555211306,
-0.1831015795469284,
-0.008331802673637867,
0.020371148362755775,
-0.06421760469675064,
0.05430314689874649,
-0.034564655274152756,
-0.012017273344099522,
0.028509587049484253,
-0.04139098525047302,
-0.02183026261627674,
-0.03214948624372482,
-0.042309775948524475,
-0.03429490327835083,
-0.0743316039443016,
-0.08551790565252304,
-0.05859500914812088,
0.0050581698305904865,
0.017415836453437805,
-0.052075959742069244,
0.003966003656387329,
0.09073054045438766,
0.02061077021062374,
-0.05009899660944939,
-0.08740974217653275,
0.05373237282037735,
-0.06826884299516678,
0.014821408316493034,
0.049178894609212875,
0.061582326889038086,
0.020740486681461334,
-0.0028462717309594154,
0.016916224732995033,
0.05254241079092026,
0.046476565301418304,
0.07579771429300308,
0.007469573523849249,
-0.03414532542228699,
0.04704892635345459,
0.04566646367311478,
0.04446046054363251,
-0.00643231812864542,
0.034057747572660446,
0.06075518578290939,
-0.05942120775580406,
-0.006244969554245472,
0.021417269483208656,
-0.003980738576501608,
-0.09747687727212906,
0.09231150895357132,
0.03812585771083832,
0.02996479533612728,
-0.023531051352620125,
0.02437056601047516,
-0.09065728634595871,
-0.05509958416223526,
-0.05365659296512604,
0.04479636996984482,
-0.1421290636062622,
0.00014075481158215553,
-0.03239596262574196,
-0.03444647416472435,
0.041765857487916946,
-0.04362759366631508,
-0.03055313415825367,
0.0322754792869091,
1.0928891485476398e-33,
0.0623827688395977,
-0.03846404328942299,
-0.003323123324662447,
0.006953093688935041,
0.009426913224160671,
-0.01708752103149891,
0.02876472845673561,
0.033247265964746475,
-0.04280498996376991,
-0.040569745004177094,
0.02624349296092987,
-0.04082748666405678,
0.032641950994729996,
-0.03520847484469414,
0.028187323361635208,
0.027483398094773293,
0.00457565113902092,
0.029334308579564095,
-0.04375080019235611,
0.04148826375603676,
-0.04645556956529617,
0.06165263056755066,
-0.07911994308233261,
0.07661671936511993,
-0.07718026638031006,
0.09917034208774567,
-0.08138139545917511,
-0.03981902450323105,
-0.07432331889867783,
-0.02820374257862568,
0.0027456192765384912,
-0.0016275455709546804,
0.002344349632039666,
-0.04173121228814125,
-0.015947703272104263,
0.08144329488277435,
0.031083954498171806,
-0.013563239015638828,
-0.03860989958047867,
0.0994541272521019,
0.08015350997447968,
0.0054466682486236095,
-0.009985196404159069,
0.025223128497600555,
0.014598583802580833,
-0.10648363083600998,
-0.09676533937454224,
-0.01299744751304388,
0.08488612622022629,
0.03160139545798302,
0.041140515357255936,
-0.06383025646209717,
-0.05020739138126373,
-0.0025969529524445534,
0.019729435443878174,
-0.06730407476425171,
-0.06308295577764511,
-0.027094407007098198,
0.018142983317375183,
0.0832429975271225,
-0.04461631178855896,
-0.09011927992105484,
-0.0561375692486763,
-0.05673060566186905,
-0.04337407648563385,
-0.02568151243031025,
0.03287578374147415,
-0.042393021285533905,
-0.029980333521962166,
-0.03739383816719055,
-0.009653706103563309,
-0.057589780539274216,
0.030217725783586502,
-0.007364442106336355,
0.057588011026382446,
0.038320597261190414,
0.06502620130777359,
0.03747233375906944,
-0.0025494899600744247,
-0.04728124290704727,
-0.043045561760663986,
-0.04611845687031746,
-0.015460921451449394,
0.08422789722681046,
0.036184828728437424,
0.06277337670326233,
-0.02536609396338463,
0.05241336300969124,
0.018682850524783134,
-0.05049942061305046,
-0.00346292438916862,
-0.0161321759223938,
0.020668640732765198,
0.06133725494146347,
0.1090160384774208,
-3.27927409671247e-8,
-0.07977288961410522,
0.03798282518982887,
-0.06929289549589157,
-0.013551383279263973,
-0.001685276860371232,
0.012498444877564907,
-0.054879236966371536,
0.07496646046638489,
0.004198132548481226,
0.031714897602796555,
0.10857165604829788,
-0.04978388547897339,
-0.07689465582370758,
0.01806623861193657,
0.025040922686457634,
0.029683828353881836,
0.021364286541938782,
0.020156461745500565,
0.03139088675379753,
-0.02841547504067421,
0.058731142431497574,
0.007907555438578129,
0.04643356800079346,
-0.031534262001514435,
-0.013277432881295681,
-0.023473110049962997,
-0.032965075224637985,
0.08064761012792587,
0.05575675144791603,
-0.0231840368360281,
0.001685102703049779,
-0.00924854725599289,
-0.003496674122288823,
-0.07313045114278793,
-0.002781308488920331,
-0.010612846352159977,
-0.014819065108895302,
-0.0610632598400116,
-0.0007517734775319695,
-0.013624381273984909,
0.09713137894868851,
0.05362976714968681,
-0.05349693074822426,
0.027973057702183723,
-0.011890862137079239,
0.012936389073729515,
0.03593127429485321,
-0.03341221064329147,
-0.005093961022794247,
0.02640213444828987,
0.010825087316334248,
-0.0734778419137001,
0.08701019734144211,
0.14279960095882416,
0.053858496248722076,
-0.009274683892726898,
0.02561376616358757,
0.04576677083969116,
0.10275351256132126,
0.014212840236723423,
0.052100371569395065,
0.011794314719736576,
-0.1027541384100914,
-0.008359829895198345
] | 0 | 0 | 0 | 0 |
|
RNNs are not Transformers (Yet): The Key Bottleneck on In-Context Retrieval
|
https://openreview.net/forum?id=h3wbI8Uk1Z
|
[
"Kaiyue Wen",
"Xingyu Dang",
"Kaifeng Lyu"
] |
Poster
|
This paper investigates the gap in representation powers of Transformers and Recurrent Neural Networks (RNNs), which are more memory efficient than Transformers. We aim to understand whether RNNs can match the performance of Transformers, particularly when enhanced with Chain-of-Thought (CoT) prompting. Our theoretical analysis reveals that CoT improves RNNs but is insufficient to close the gap with Transformers. A key bottleneck lies in the inability of RNNs to perfectly retrieve information from the context, even with CoT:
for several tasks that explicitly or implicitly require this capability, such as associative recall and determining if a graph is a tree, we prove that RNNs are not expressive enough to solve the tasks while Transformers can solve them with ease.
Conversely, we prove that adopting techniques to enhance the in-context retrieval capability of RNNs, including Retrieval-Augmented Generation (RAG) and adding a single Transformer layer, can elevate RNNs to be capable of solving all polynomial-time solvable problems with CoT, hence closing the representation gap with Transformers. We validate our theory on synthetic and natural language experiments.
|
rnn, cot, representation theory
|
We identify a representation gap between Transformers and RNNs (w/o CoT) due to the incapability to retrieve in context and show by adopting techniques to enhance this capability can bridge the gap.
| 11,974 |
2402.18510
|
[
-0.12105108052492142,
-0.04380546510219574,
0.008833413943648338,
0.053303442895412445,
0.05939759686589241,
0.023097770288586617,
0.0004929081769660115,
0.04417433962225914,
0.07451950758695602,
-0.09712892025709152,
-0.046964358538389206,
0.015009326860308647,
0.042857810854911804,
0.05702602118253708,
0.06281901150941849,
0.04892626404762268,
0.04446716606616974,
0.12207826226949692,
-0.051240526139736176,
-0.042813558131456375,
0.02992224134504795,
0.03775698319077492,
0.04536357522010803,
-0.024212747812271118,
0.03144489973783493,
-0.008343610912561417,
-0.043679993599653244,
-0.05882587656378746,
0.07419826835393906,
0.061812639236450195,
0.06367640197277069,
0.07066959887742996,
-0.09915844351053238,
0.06488652527332306,
-0.06739643961191177,
0.11735501140356064,
-0.11883589625358582,
0.04941978305578232,
0.01991059072315693,
-0.08983896672725677,
-0.04432281479239464,
0.033463556319475174,
-0.0047591510228812695,
0.06200743839144707,
0.10112900286912918,
-0.04404599592089653,
0.011690332554280758,
0.004645438864827156,
-0.012216758914291859,
-0.005177950486540794,
-0.09370481967926025,
0.029096122831106186,
-0.04415888339281082,
0.08752530813217163,
0.000009087585567613132,
0.0601036511361599,
0.044308390468358994,
-0.006173678208142519,
-0.0541982464492321,
-0.05478056147694588,
-0.06700912863016129,
-0.09838586300611496,
-0.026358401402831078,
-0.07129395008087158,
-0.08631671965122223,
-0.01958291605114937,
0.07941948622465134,
0.066382996737957,
-0.018736714497208595,
0.008783589117228985,
0.0018971969839185476,
0.05659058690071106,
-0.025859832763671875,
0.024998128414154053,
0.07469324767589569,
0.07909779995679855,
0.017592793330550194,
-0.0934734120965004,
0.0044217901304364204,
-0.0030440669506788254,
0.013390209525823593,
0.03165922686457634,
0.04902791976928711,
0.029929082840681076,
0.08222416043281555,
0.02577863074839115,
0.01695793867111206,
0.0119102718308568,
-0.04355927184224129,
0.050845615565776825,
-0.038981176912784576,
-0.07475604861974716,
0.0706692487001419,
-0.057649437338113785,
0.07173175364732742,
0.06598738580942154,
0.012381887063384056,
-0.027509810402989388,
-0.024263037368655205,
0.07113069295883179,
0.010912910103797913,
0.053204745054244995,
-0.030831987038254738,
-0.0903579518198967,
-0.09277418255805969,
0.015370472334325314,
0.028836797922849655,
0.003719872096553445,
0.0014412145828828216,
-0.05329468473792076,
0.021407995373010635,
0.03669266775250435,
0.01950988918542862,
-0.015571249648928642,
-0.017005041241645813,
-0.036366790533065796,
0.015505223535001278,
-0.09836040437221527,
0.00027931033400818706,
0.07524572312831879,
-0.01966877654194832,
0.09101162105798721,
-0.012520119547843933,
0.06254084408283234,
0.006509371101856232,
-0.039491523057222366,
0.02155674807727337,
4.879897656835719e-34,
-0.002229072619229555,
0.018669672310352325,
0.058729298412799835,
0.09774158149957657,
0.09588319063186646,
0.028691189363598824,
0.020667674019932747,
0.05944939702749252,
-0.04794923961162567,
0.009059329517185688,
-0.05739110708236694,
0.08846791088581085,
-0.05597136542201042,
0.00959670078009367,
0.02100287564098835,
-0.010392696596682072,
-0.061678748577833176,
-0.024591129273176193,
-0.04109886661171913,
-0.009814654476940632,
-0.017773747444152832,
-0.015318506397306919,
0.04827311262488365,
-0.0019546214025467634,
-0.02408076822757721,
-0.030101215466856956,
0.05332574620842934,
-0.03045993484556675,
0.009818471036851406,
-0.020023444667458534,
-0.10669385641813278,
0.025300493463873863,
-0.0029777490999549627,
0.04847114533185959,
0.009728791192173958,
-0.07846325635910034,
0.003088536439463496,
-0.03170868009328842,
-0.004553793929517269,
0.007879281416535378,
0.013868877664208412,
0.06833327561616898,
-0.015396826900541782,
0.05938372388482094,
-0.07794550806283951,
-0.07188452035188675,
0.03944546356797218,
0.021952450275421143,
0.014775230549275875,
0.024425635114312172,
0.04243996739387512,
0.04787778854370117,
-0.08289077877998352,
-0.058340124785900116,
0.11186481267213821,
-0.05000118538737297,
0.0005119700217619538,
0.09014597535133362,
0.054520465433597565,
0.0694294273853302,
0.02797076292335987,
0.030168941244482994,
0.05550654977560043,
0.06746038794517517,
0.030529553070664406,
0.02585325762629509,
-0.045497775077819824,
0.06062771752476692,
0.08315976709127426,
-0.02857825718820095,
-0.0032595437951385975,
0.022901052609086037,
-0.02568330243229866,
-0.021606585010886192,
0.04025985673069954,
0.002719445386901498,
0.003498499048873782,
-0.10556963086128235,
-0.033026278018951416,
0.006177239120006561,
0.016860414296388626,
-0.0704018697142601,
0.01818002387881279,
0.0005395836778916419,
-0.06010454520583153,
-0.0699286088347435,
0.051447462290525436,
-0.08345796167850494,
0.012118273414671421,
-0.05148981884121895,
-0.04414607584476471,
-0.004270312376320362,
-0.042876485735177994,
-0.023024292662739754,
0.005571709014475346,
-9.316199561514608e-34,
0.004547211341559887,
-0.025660963729023933,
-0.03715018928050995,
0.036909252405166626,
-0.0036033540964126587,
-0.09910735487937927,
-0.007123264018446207,
-0.037808794528245926,
-0.0511355996131897,
-0.013163676485419273,
-0.049767643213272095,
-0.06998088955879211,
0.06253446638584137,
-0.0032144994474947453,
0.052531417459249496,
-0.016979584470391273,
-0.02555399388074875,
-0.05422719568014145,
-0.04399522393941879,
0.05748753622174263,
0.006517451722174883,
0.10440950095653534,
-0.13184575736522675,
-0.00632608775049448,
-0.08657410740852356,
0.020909463986754417,
-0.08425357937812805,
0.016755057498812675,
-0.001254148781299591,
-0.02769785188138485,
-0.035264596343040466,
-0.030617976561188698,
0.004869993310421705,
-0.005283982027322054,
-0.03644539788365364,
0.037197086960077286,
0.06361003965139389,
-0.030019214376807213,
-0.07566060870885849,
0.03361984342336655,
0.0767286866903305,
-0.06291613727807999,
-0.018095415085554123,
0.02606169879436493,
-0.059711866080760956,
-0.007791634649038315,
-0.10069209337234497,
-0.005530282389372587,
0.017808841541409492,
0.06154331564903259,
0.006290307734161615,
0.010396203957498074,
-0.08793552219867706,
-0.06327968835830688,
0.0017062549013644457,
-0.10684467852115631,
-0.056769974529743195,
-0.03521488979458809,
0.06571171432733536,
-0.008458196185529232,
0.00720709515735507,
-0.02306295931339264,
0.006355614867061377,
-0.12645161151885986,
-0.003916345536708832,
-0.04793301969766617,
-0.002824092283844948,
-0.04698813706636429,
0.004390618298202753,
0.0667911022901535,
0.053357698023319244,
0.03391776233911514,
0.001072191633284092,
-0.03965885192155838,
0.04372791200876236,
0.03990592435002327,
0.02826670929789543,
-0.07199166715145111,
-0.03429500386118889,
-0.10502944886684418,
-0.059908948838710785,
-0.02250116877257824,
-0.007176184561103582,
0.028180016204714775,
0.0034074843861162663,
0.07967465370893478,
0.035011690109968185,
0.09719029068946838,
0.05029553920030594,
0.006017228122800589,
-0.014428592287003994,
0.034525204449892044,
-0.03868716210126877,
0.08158373832702637,
0.005019668489694595,
-4.5207780630107663e-8,
-0.08681373298168182,
0.02484794519841671,
-0.012783198617398739,
0.028304480016231537,
0.03430918976664543,
-0.021231412887573242,
0.026225587353110313,
0.1355976015329361,
-0.02224864810705185,
0.03943447396159172,
0.07415727525949478,
-0.04905751347541809,
-0.03988314047455788,
-0.056005291640758514,
0.04510337486863136,
0.03350352123379707,
0.017411915585398674,
-0.013191708363592625,
0.004205831792205572,
-0.0569644421339035,
-0.00041765597416087985,
0.08978923410177231,
-0.03198368847370148,
0.01229169312864542,
-0.024294162169098854,
-0.06096235290169716,
-0.006758900824934244,
0.06319180876016617,
0.041814666241407394,
-0.02461576648056507,
0.04798369109630585,
-0.026785356923937798,
0.006140653509646654,
-0.09685740619897842,
-0.029260234907269478,
0.050378527492284775,
0.05717630684375763,
-0.006170992273837328,
-0.011220073327422142,
-0.07065852731466293,
0.07524023205041885,
0.030880548059940338,
-0.01941573992371559,
0.03593624383211136,
-0.04867754131555557,
0.0077181123197078705,
-0.0067504714243113995,
-0.09220664203166962,
0.04134213179349899,
-0.008839105255901814,
0.03832842409610748,
-0.07456784695386887,
-0.04398898407816887,
0.048795245587825775,
0.04510285332798958,
0.029708916321396828,
0.0031510915141552687,
-0.05494789034128189,
-0.0020539972465485334,
0.05102992802858353,
-0.0430850051343441,
0.07539433985948563,
-0.02544422820210457,
0.026235906407237053
] |
https://github.com/dangxingyu/rnn-icrag
| 27 | 0 | 0 | 0 |
Regressing the Relative Future: Efficient Policy Optimization for Multi-turn RLHF
|
https://openreview.net/forum?id=cVyELMpMRS
|
[
"Zhaolin Gao",
"Wenhao Zhan",
"Jonathan Daniel Chang",
"Gokul Swamy",
"Kianté Brantley",
"Jason D. Lee",
"Wen Sun"
] |
Poster
|
Large Language Models (LLMs) have achieved remarkable success at tasks like summarization that involve a single turn of interaction. However, they can still struggle with multi-turn tasks like dialogue that require long-term planning. Previous works on multi-turn dialogue extend single-turn reinforcement learning from human feedback (RLHF) methods to the multi-turn setting by treating all prior dialogue turns as a long context. Such approaches suffer from covariate shift: the conversations in the training set have previous turns generated by some reference policy, which means that low training error may not necessarily correspond to good performance when the learner is actually in the conversation loop. In response, we introduce REgressing the RELative FUture (REFUEL), an efficient policy optimization approach designed to address multi-turn RLHF in LLMs. REFUEL employs a single model to estimate $Q$-values and trains on self-generated data, addressing the covariate shift issue. REFUEL frames the multi-turn RLHF problem as a sequence of regression tasks on iteratively collected datasets, enabling ease of implementation. Theoretically, we prove that REFUEL can match the performance of any policy covered by the training set. Empirically, we evaluate our algorithm by using Llama-3.1-70B-it to simulate a user in conversation with our model. REFUEL consistently outperforms state-of-the-art methods such as DPO and REBEL across various settings. Furthermore, despite having only 8 billion parameters, Llama-3-8B-it fine-tuned with REFUEL outperforms Llama-3.1-70B-it on long multi-turn dialogues. Implementation of REFUEL can be found at https://github.com/ZhaolinGao/REFUEL/, and models trained by REFUEL can be found at https://huggingface.co/Cornell-AGI.
|
Reinforcement Learning, Reinforcement Learning from Human Feedback
| null | 11,971 |
2410.04612
|
[
-0.041861921548843384,
-0.06632369756698608,
0.04099986329674721,
0.01890508458018303,
0.04425514489412308,
0.06896866112947464,
-0.002881658961996436,
0.003793523646891117,
0.05738487094640732,
-0.006508934777230024,
-0.08184535801410675,
-0.0038912559393793344,
-0.02687370404601097,
0.04783109202980995,
0.01871403679251671,
0.040903545916080475,
0.054172273725271225,
0.005641618277877569,
-0.03773898631334305,
-0.05385277420282364,
0.03699605539441109,
0.020019056275486946,
0.09951701015233994,
0.03739848732948303,
0.002284371294081211,
-0.0063508958555758,
-0.05411256477236748,
-0.022241991013288498,
-0.021919911727309227,
-0.010588482953608036,
0.037258539348840714,
0.08051956444978714,
-0.02781841717660427,
-0.008965586312115192,
-0.1336066573858261,
0.11904218047857285,
-0.07171478122472763,
0.034204717725515366,
0.015448986552655697,
0.037900060415267944,
-0.06823524087667465,
0.022496940568089485,
0.06614476442337036,
0.026719704270362854,
0.12038207799196243,
-0.03829321637749672,
-0.05805560201406479,
-0.0026912223547697067,
0.039017222821712494,
0.016871025785803795,
-0.029781116172671318,
-0.0009181837667711079,
0.03730548545718193,
0.06655347347259521,
0.011750586330890656,
0.052245479077100754,
0.025968141853809357,
0.07294825464487076,
0.027432257309556007,
-0.0571167878806591,
-0.05937959998846054,
-0.17065978050231934,
-0.035935722291469574,
-0.06280135363340378,
-0.11744534224271774,
-0.027175571769475937,
-0.026180336251854897,
0.03649821877479553,
-0.0514601469039917,
0.08446399122476578,
-0.05280151590704918,
0.02820418030023575,
-0.024073027074337006,
-0.03826897591352463,
0.019374847412109375,
0.012178001925349236,
0.02431967295706272,
-0.004927300848066807,
0.04202733561396599,
-0.0602772980928421,
-0.004790981765836477,
-0.037409380078315735,
0.0685424655675888,
-0.0566585436463356,
0.08173386007547379,
-0.012591552920639515,
-0.005579235963523388,
0.06112072616815567,
0.07284971326589584,
-0.02881314605474472,
-0.020981205627322197,
-0.028076641261577606,
0.04394398257136345,
0.00980229489505291,
-0.00983229000121355,
0.04537813365459442,
-0.04942621663212776,
-0.07789406925439835,
-0.035372294485569,
0.06299085915088654,
0.04284799471497536,
0.08376958966255188,
0.0195976160466671,
-0.031103597953915596,
-0.04519260302186012,
-0.0331139899790287,
0.004344450309872627,
0.06851796805858612,
-0.031581707298755646,
-0.08703497797250748,
0.010996945202350616,
0.05019737035036087,
0.039812490344047546,
0.06118573620915413,
0.016293257474899292,
0.030288705602288246,
-0.0017941039986908436,
-0.06121319159865379,
-0.053532760590314865,
0.02458270825445652,
-0.026197288185358047,
-0.05708673596382141,
0.04461336508393288,
0.05098865553736687,
0.005777048412710428,
-0.03934596851468086,
0.0005984940798953176,
3.527843687177939e-33,
0.06761598587036133,
0.0599503368139267,
0.035605911165475845,
0.03254599869251251,
0.016423864290118217,
0.006205504294484854,
0.04662657156586647,
0.05212673172354698,
-0.025022076442837715,
-0.06458152830600739,
0.06983285397291183,
0.00016942174988798797,
-0.043538425117731094,
0.014484385028481483,
-0.010051820427179337,
-0.0018499556463211775,
-0.06859274953603745,
0.05379995331168175,
0.017272736877202988,
-0.039431553333997726,
0.06409157067537308,
0.010449595749378204,
0.00007395254942821339,
-0.05854913219809532,
0.06184355169534683,
0.030841626226902008,
0.09806781262159348,
-0.05050097778439522,
-0.04891769215464592,
0.0035856044851243496,
-0.04748346284031868,
-0.0026363502256572247,
-0.03417686000466347,
0.04701590910553932,
0.0035753955598920584,
-0.01902601309120655,
-0.10190348327159882,
-0.03153758868575096,
0.018040237948298454,
0.03179919347167015,
-0.026956766843795776,
0.0005063776625320315,
0.049984414130449295,
-0.014486484229564667,
-0.08217968046665192,
-0.08225344121456146,
-0.0036473595537245274,
0.025777209550142288,
-0.07921634614467621,
-0.01781788468360901,
-0.03627582639455795,
0.05678016319870949,
-0.08996595442295074,
-0.03753549978137016,
-0.05715940147638321,
-0.00356313562951982,
0.016737302765250206,
0.01246327068656683,
-0.013622887432575226,
0.033023446798324585,
0.01660979725420475,
0.06399799138307571,
-0.034121498465538025,
0.10036642849445343,
0.03360794484615326,
0.04190536588430405,
-0.04951721057295799,
0.026516031473875046,
0.11414997279644012,
-0.014649159274995327,
0.014915443025529385,
0.00033248335239477456,
0.02374443970620632,
-0.01842292957007885,
0.09255284070968628,
-0.0011638300493359566,
0.05525343120098114,
-0.07851464301347733,
0.02943561226129532,
-0.0006664099055342376,
-0.0037076384760439396,
0.000007353043656621594,
0.010876055806875229,
-0.08066876232624054,
0.0713975578546524,
-0.08123231679201126,
0.04646943509578705,
-0.07415760308504105,
-0.006144870538264513,
0.01325759943574667,
-0.056269846856594086,
-0.012973995879292488,
-0.055369433015584946,
0.05127498880028725,
0.0306126419454813,
-2.6198541980493797e-33,
0.0663028359413147,
-0.019470609724521637,
-0.04005523771047592,
0.05436215177178383,
0.020876547321677208,
0.005008410196751356,
0.027467411011457443,
-0.008265150710940361,
0.026169097051024437,
-0.08386286348104477,
-0.07866477966308594,
-0.015802407637238503,
0.0598302036523819,
0.029050208628177643,
0.05786239728331566,
-0.08087994158267975,
0.036222491413354874,
0.010751848109066486,
0.06689268350601196,
0.06718097627162933,
-0.01594788022339344,
0.030853234231472015,
-0.09219301491975784,
0.06700342893600464,
-0.059118740260601044,
0.003774122567847371,
0.026437895372509956,
0.07152687013149261,
-0.03694239258766174,
0.012326279655098915,
-0.04818405210971832,
-0.07423294335603714,
-0.020770054310560226,
-0.021667499095201492,
-0.05975053459405899,
0.049940358847379684,
0.0003802878491114825,
-0.007179481443017721,
-0.05617164447903633,
0.10860968381166458,
0.1013515442609787,
-0.08895134180784225,
-0.08939153701066971,
0.013688209466636181,
-0.044357556849718094,
-0.04332869499921799,
-0.12750644981861115,
0.013375415466725826,
-0.017987757921218872,
-0.00197763298638165,
-0.046690575778484344,
0.014263597317039967,
-0.05750659108161926,
-0.007971539162099361,
-0.05153267830610275,
-0.09092726558446884,
-0.002783311763778329,
-0.0332406647503376,
0.04614797607064247,
0.002922580810263753,
-0.05603857338428497,
0.028065374121069908,
0.03640588000416756,
-0.04841993749141693,
0.08431525528430939,
-0.06074294447898865,
-0.015462853945791721,
-0.04438522458076477,
0.06888765096664429,
-0.006212406791746616,
0.07870963960886002,
-0.00649112556129694,
0.048359569162130356,
-0.012512585148215294,
0.042832594364881516,
-0.00311127002350986,
-0.03172782436013222,
-0.08812694251537323,
-0.053458768874406815,
-0.013667786493897438,
0.013719340786337852,
-0.002408255124464631,
0.06996667385101318,
0.07632316648960114,
0.018003184348344803,
0.050418250262737274,
0.04846109077334404,
0.1373763233423233,
0.028445161879062653,
-0.024974534288048744,
-0.023304734379053116,
-0.020992593839764595,
0.014323699288070202,
0.04053838551044464,
-0.0011866290587931871,
-4.402964037808488e-8,
-0.09770908206701279,
0.0295413825660944,
0.07100223749876022,
0.07313476502895355,
0.02972150593996048,
-0.0519891232252121,
0.04264124855399132,
0.03669336065649986,
-0.038262948393821716,
0.016403308138251305,
0.020747264847159386,
0.08463631570339203,
0.04806435480713844,
-0.034615326672792435,
-0.054813213646411896,
0.06715428829193115,
-0.037125762552022934,
-0.013577158562839031,
-0.02947636879980564,
0.007853245362639427,
0.11179401725530624,
-0.029246505349874496,
0.009271515533328056,
-0.03359178826212883,
0.04144161939620972,
-0.042830973863601685,
0.0030707139521837234,
0.09450063109397888,
0.04664164036512375,
-0.050636399537324905,
-0.013517123647034168,
-0.05122090131044388,
0.012340865097939968,
0.011292103677988052,
-0.015623328275978565,
-0.0010918235639110208,
0.08863411098718643,
-0.016358550637960434,
0.07370986044406891,
-0.04246886819601059,
0.0013099618954584002,
0.0579521544277668,
-0.09865003079175949,
0.03396265208721161,
0.0021079464349895716,
0.0005508490721695125,
-0.11113061010837555,
-0.13743238151073456,
0.01762566529214382,
-0.09608184546232224,
0.027837837114930153,
-0.06047029048204422,
0.01228476595133543,
0.0583379827439785,
0.11379215866327286,
0.05283663049340248,
-0.03395383059978485,
-0.03989535942673683,
0.050718724727630615,
0.08672401309013367,
-0.04019566625356674,
0.017628375440835953,
-0.0392875038087368,
0.004575397819280624
] |
https://github.com/zhaolingao/refuel
| 16 | 0 | 0 | 0 |
Diffusing States and Matching Scores: A New Framework for Imitation Learning
|
https://openreview.net/forum?id=kWRKNDU6uN
|
[
"Runzhe Wu",
"Yiding Chen",
"Gokul Swamy",
"Kianté Brantley",
"Wen Sun"
] |
Poster
|
Adversarial Imitation Learning is traditionally framed as a two-player zero-sum game between a learner and an adversarially chosen cost function, and can therefore be thought of as the sequential generalization of a Generative Adversarial Network (GAN). However, in recent years, diffusion models have emerged as a non-adversarial alternative to GANs that merely require training a score function via regression, yet produce generations of higher quality. In response, we investigate how to lift insights from diffusion modeling to the sequential setting. We propose diffusing states and performing *score-matching* along diffused states to measure the discrepancy between the expert's and learner's states. Thus, our approach only requires training score functions to predict noises via standard regression, making it significantly easier and more stable to train than adversarial methods. Theoretically, we prove first- and second-order instance-dependent bounds with linear scaling in the horizon, proving that our approach avoids the compounding errors that stymie offline approaches to imitation learning. Empirically, we show our approach outperforms both GAN-style imitation learning baselines and discriminator-free imitation learning baselines across various continuous control problems, including complex tasks like controlling humanoids to walk, sit, crawl, and navigate through obstacles.
|
imitation learning, diffusion model, score matching
| null | 11,966 |
2410.13855
|
[
-0.11917614936828613,
-0.11003411561250687,
0.058747436851263046,
0.05761835351586342,
0.03452412411570549,
0.012245053425431252,
-0.03908815234899521,
-0.07283154129981995,
0.01941135711967945,
-0.03581460192799568,
-0.03785204514861107,
-0.03526488319039345,
-0.0014382745139300823,
0.0605255551636219,
-0.06898406147956848,
-0.009087543934583664,
0.09473731368780136,
0.055547043681144714,
-0.027038557454943657,
-0.025265444070100784,
-0.02704799920320511,
0.024043768644332886,
-0.03639387711882591,
0.02000119350850582,
0.05244272202253342,
0.015227360650897026,
0.029773131012916565,
-0.06365252286195755,
-0.012846380472183228,
-0.035172559320926666,
0.021008595824241638,
-0.01241952646523714,
-0.006070693954825401,
0.015323574654757977,
-0.10999616980552673,
0.052994415163993835,
-0.0803365707397461,
-0.023219933733344078,
0.053153254091739655,
-0.06628303229808807,
0.05309617146849632,
-0.023476190865039825,
-0.00795189943164587,
-0.016172384843230247,
0.05713856592774391,
0.005058540962636471,
0.004361123777925968,
0.007321909070014954,
0.022708186879754066,
-0.012639062479138374,
0.009898957796394825,
0.016963202506303787,
-0.022370290011167526,
0.04290161654353142,
-0.0004887178656645119,
0.028358835726976395,
0.12090876698493958,
0.058505311608314514,
-0.017053421586751938,
0.006511176470667124,
-0.06301771849393845,
-0.06300912797451019,
-0.03280890733003616,
-0.015870850533246994,
-0.0072600157000124454,
-0.07214510440826416,
-0.016743533313274384,
0.0736938938498497,
-0.10302048176527023,
0.11600380390882492,
0.015390890650451183,
0.07160165905952454,
-0.028229957446455956,
0.008649478666484356,
0.07370994985103607,
0.046303171664476395,
0.04012022912502289,
-0.052365466952323914,
0.008114141412079334,
-0.08377059549093246,
0.018028859049081802,
-0.027147050946950912,
0.03474961966276169,
0.06159653142094612,
0.047146767377853394,
0.00008819194044917822,
-0.02484157308936119,
-0.023278728127479553,
0.13055115938186646,
-0.017649302259087563,
-0.08884374797344208,
0.056021638214588165,
-0.08107385039329529,
0.02028474770486355,
0.09236878156661987,
0.03423840552568436,
0.013750645332038403,
0.005984807852655649,
0.009762638248503208,
0.10427204519510269,
-0.03862245753407478,
0.09034683555364609,
-0.022905969992280006,
-0.05544309318065643,
0.024372609332203865,
-0.02654104307293892,
0.06838582456111908,
0.022554153576493263,
0.03233189880847931,
-0.04201488196849823,
0.09064369648694992,
0.011494318954646587,
0.07194949686527252,
0.028162816539406776,
0.036972776055336,
-0.0887761190533638,
-0.00238754041492939,
0.021086907014250755,
-0.031080663204193115,
0.0856858640909195,
-0.018409548327326775,
0.03828511759638786,
0.0070394491776824,
0.01769641786813736,
-0.017675558105111122,
-0.015719356015324593,
-0.0358850434422493,
1.0488915888578117e-33,
0.051054250448942184,
-0.04168970137834549,
0.05373692139983177,
0.027565954253077507,
0.047655556350946426,
-0.013616382144391537,
-0.019429882988333702,
-0.032630275934934616,
-0.06508082896471024,
-0.04737124592065811,
-0.014623629860579967,
0.04979569837450981,
-0.12348008155822754,
0.09303367882966995,
0.06562407314777374,
0.027493592351675034,
-0.02137601189315319,
0.002494238084182143,
0.0372026227414608,
0.014891812577843666,
0.07251564413309097,
-0.06694228202104568,
-0.07654405385255814,
-0.028315788134932518,
0.033344801515340805,
-0.02292647212743759,
-0.02429128810763359,
-0.05076441541314125,
0.034146420657634735,
0.0020805331878364086,
-0.09729284048080444,
0.00046962915803305805,
-0.018695367500185966,
0.022793447598814964,
0.04876428097486496,
0.015085055492818356,
-0.03127491846680641,
-0.0659569725394249,
-0.06915547698736191,
-0.06042465940117836,
0.02883097343146801,
0.011220493353903294,
-0.03127849102020264,
-0.01224918756633997,
-0.11801684647798538,
-0.0791638195514679,
0.011811438947916031,
0.02830681763589382,
-0.05683213472366333,
-0.007590443827211857,
0.021598421037197113,
0.03851136192679405,
-0.04820801690220833,
-0.16927392780780792,
-0.09659320116043091,
-0.03767480328679085,
0.03451400250196457,
-0.028128596022725105,
0.012881840579211712,
0.0025007922668009996,
0.007944982498884201,
0.024773480370640755,
-0.025260794907808304,
0.0519573912024498,
-0.01944030076265335,
0.04511898383498192,
-0.044903457164764404,
0.04753122851252556,
0.04598570987582207,
-0.009899819269776344,
-0.02891191653907299,
0.02546769566833973,
-0.06134616583585739,
-0.06847617775201797,
0.10011152923107147,
-0.06982506811618805,
0.037095807492733,
-0.05493875592947006,
0.020658502355217934,
-0.007672895677387714,
-0.03245711699128151,
-0.04293275251984596,
-0.05860910564661026,
-0.09341014921665192,
-0.032885175198316574,
-0.014040528796613216,
-0.0011588666820898652,
-0.035422295331954956,
-0.004757854156196117,
-0.022019755095243454,
-0.07163910567760468,
-0.02969789318740368,
-0.06521736085414886,
0.035539913922548294,
0.057484738528728485,
-2.905592079150315e-33,
-0.030958600342273712,
0.03152329847216606,
0.06760674715042114,
0.048132386058568954,
-0.01145091187208891,
0.001780114253051579,
-0.018243465572595596,
0.061408642679452896,
-0.0023994161747395992,
-0.03755901753902435,
-0.048469625413417816,
0.0436311699450016,
0.05902663990855217,
-0.018757736310362816,
-0.006931419018656015,
-0.0531696118414402,
0.024345453828573227,
0.08556050062179565,
-0.0031209916342049837,
0.02018122561275959,
0.08822092413902283,
0.1010005846619606,
-0.04056696221232414,
0.06726524233818054,
-0.09966093301773071,
-0.015140303410589695,
-0.056508708745241165,
0.11532872915267944,
-0.016302218660712242,
0.01055445708334446,
0.03761856630444527,
0.023615416139364243,
0.013051174581050873,
-0.022510157898068428,
-0.08336739987134933,
0.12298638373613358,
0.06051948666572571,
0.06474056839942932,
-0.10522669553756714,
0.12346106022596359,
0.08854243159294128,
-0.05004853010177612,
-0.06974967569112778,
0.03649044409394264,
0.016351787373423576,
-0.02269752323627472,
-0.011516813188791275,
0.016690390184521675,
0.007612256333231926,
0.0014696860453113914,
0.004595768637955189,
0.0911855548620224,
-0.09083551168441772,
0.04256559535861015,
-0.08144672214984894,
-0.056923139840364456,
-0.06858060508966446,
0.013266489841043949,
0.08875533193349838,
-0.0011328564723953605,
-0.0032324439380317926,
0.030336596071720123,
-0.04357917234301567,
0.02277642861008644,
-0.062405698001384735,
0.02999127469956875,
0.01709583029150963,
0.07061749696731567,
0.07899995148181915,
0.038013752549886703,
-0.0267789326608181,
-0.009595135226845741,
-0.009024047292768955,
-0.03907347097992897,
-0.009868454188108444,
-0.07388674467802048,
0.026926664635539055,
0.007494466379284859,
-0.014033887535333633,
0.008585250005126,
0.06072939187288284,
-0.03034774214029312,
0.058351561427116394,
-0.04153195396065712,
0.06918276101350784,
-0.0013061007484793663,
0.02857624553143978,
0.04051455110311508,
-0.020330827683210373,
0.09231376647949219,
-0.004334761295467615,
0.0033729143906384706,
0.009278529323637486,
-0.07133207470178604,
-0.049463242292404175,
-4.873649572800787e-8,
-0.10663330554962158,
0.05857820808887482,
0.11604809761047363,
0.06057422235608101,
-0.018948251381516457,
-0.009042519144713879,
-0.027477877214550972,
0.02131018415093422,
0.01263782661408186,
0.020160576328635216,
0.01836535334587097,
0.004839236848056316,
0.0032830829732120037,
-0.11035355180501938,
-0.010308281518518925,
0.09465456753969193,
0.006319721229374409,
-0.032531727105379105,
-0.001686211908236146,
0.033405013382434845,
-0.006201085634529591,
-0.003539786906912923,
0.020523518323898315,
-0.047019921243190765,
0.0298102255910635,
-0.05083945766091347,
-0.09240009635686874,
-0.005637067835777998,
0.042083799839019775,
0.04280594363808632,
0.014701771549880505,
0.043958354741334915,
0.045472126454114914,
-0.07925426214933395,
-0.014833717606961727,
0.014143703505396843,
0.030598819255828857,
-0.04298913851380348,
0.038443949073553085,
-0.053476717323064804,
-0.00795299094170332,
0.03769370913505554,
0.008187620900571346,
-0.0233493410050869,
-0.0016615742351859808,
0.023417100310325623,
-0.00687919557094574,
-0.09597175568342209,
0.07279805839061737,
-0.022374099120497704,
0.10844935476779938,
-0.058915428817272186,
0.01971326395869255,
0.02319893054664135,
0.09469538927078247,
-0.02767082490026951,
0.014787322841584682,
-0.017953066155314445,
0.021516036242246628,
0.08682770282030106,
0.06334634125232697,
-0.014481686986982822,
0.004630003124475479,
0.010440384037792683
] |
https://github.com/ziqian2000/smiling
| 16 | 0 | 0 | 0 |
Behavioral Entropy-Guided Dataset Generation for Offline Reinforcement Learning
|
https://openreview.net/forum?id=LuT2CVrlpU
|
[
"Wesley A. Suttle",
"Aamodh Suresh",
"Carlos Nieto-Granda"
] |
Poster
|
Entropy-based objectives are widely used to perform state space exploration in reinforcement learning (RL) and dataset generation for offline RL. Behavioral entropy (BE), a rigorous generalization of classical entropies that incorporates cognitive and perceptual biases of agents, was recently proposed for discrete settings and shown to be a promising metric for robotic exploration problems. In this work, we propose using BE as a principled exploration objective for systematically generating datasets that provide diverse state space coverage in complex, continuous, potentially high-dimensional domains. To achieve this, we extend the notion of BE to continuous settings, derive tractable $k$-nearest neighbor estimators, provide theoretical guarantees for these estimators, and develop practical reward functions that can be used with standard RL methods to learn BE-maximizing policies. Using standard MuJoCo environments, we experimentally compare the performance of offline RL algorithms for a variety of downstream tasks on datasets generated using BE, R\'{e}nyi, and Shannon entropy-maximizing policies, as well as the SMM and RND algorithms. We find that offline RL algorithms trained on datasets collected using BE outperform those trained on datasets collected using Shannon entropy, SMM, and RND on all tasks considered, and on 80\% of the tasks compared to datasets collected using Renyi entropy.
|
reinforcement learning, offline reinforcement learning, exploration, entropy
|
We explore a new exploration objective for RL and show that it generates superior datasets for subsequent offline RL.
| 11,959 |
2502.04141
|
[
-0.03958401083946228,
-0.06955167651176453,
0.045801978558301926,
-0.005626197904348373,
0.04726434126496315,
0.04226985201239586,
0.00813822541385889,
-0.03486328199505806,
-0.004449492786079645,
0.05632360279560089,
-0.057660628110170364,
-0.03282758966088295,
0.0816144198179245,
0.07039140164852142,
-0.030749550089240074,
0.033004362136125565,
0.13941240310668945,
0.024178069084882736,
-0.10299801081418991,
-0.0015776859363541007,
0.08401995152235031,
0.049216583371162415,
0.10245896875858307,
0.005534093361347914,
0.00487299682572484,
0.009955167770385742,
0.0804426521062851,
-0.04770834371447563,
0.001934064319357276,
0.03863845765590668,
-0.006190039217472076,
-0.027025533840060234,
-0.029986688867211342,
-0.10361016541719437,
-0.02250290848314762,
0.11632709950208664,
-0.12497080117464066,
0.024050092324614525,
-0.08266378939151764,
-0.0030047644395381212,
-0.044022396206855774,
0.0388900451362133,
0.04110628738999367,
0.022517750039696693,
0.024495301768183708,
0.03203364834189415,
-0.043966274708509445,
-0.08923554420471191,
0.060540638864040375,
-0.00900364201515913,
-0.03273935988545418,
0.052666645497083664,
0.03710559755563736,
0.009543485008180141,
0.03323656693100929,
0.05931108444929123,
0.020734479650855064,
0.009571495465934277,
0.04415787756443024,
-0.030927877873182297,
0.00099581154063344,
-0.09883664548397064,
-0.013497509993612766,
-0.044528815895318985,
-0.08858264237642288,
-0.07702603191137314,
0.01493872981518507,
0.06892190873622894,
0.014488469809293747,
-0.05093182250857353,
-0.047224510461091995,
0.038026098161935806,
0.012286784127354622,
-0.03834071010351181,
0.08602802455425262,
0.015508859418332577,
0.03980058431625366,
0.002285952679812908,
0.03357797488570213,
-0.042745642364025116,
-0.022637583315372467,
-0.028356386348605156,
0.025809867307543755,
0.050433751195669174,
0.05263539403676987,
-0.026393309235572815,
0.065508633852005,
0.04782464727759361,
0.15353691577911377,
0.024592021480202675,
-0.012454075738787651,
-0.01901138387620449,
-0.035399653017520905,
-0.03570371866226196,
0.027458662167191505,
-0.010454175993800163,
-0.08117781579494476,
-0.053047411143779755,
0.005375274457037449,
0.024944160133600235,
0.004286316689103842,
0.07217974960803986,
-0.027125302702188492,
0.016289636492729187,
0.00375593570061028,
-0.08081348240375519,
0.05050038546323776,
0.022145215421915054,
0.09761679917573929,
-0.0863729938864708,
0.06214658170938492,
-0.016642654314637184,
0.037433698773384094,
0.030890949070453644,
0.01389682199805975,
0.0007070128340274096,
0.02514038234949112,
0.009768643416464329,
-0.02122633159160614,
0.035177044570446014,
-0.0038341465406119823,
-0.06963362544775009,
0.01952151022851467,
-0.01816086657345295,
0.07417964190244675,
-0.035578660666942596,
0.012440327554941177,
6.818404931599273e-34,
0.05338705703616142,
-0.05955415591597557,
0.03164249658584595,
0.01437845453619957,
0.004516195505857468,
-0.015947086736559868,
0.05775197222828865,
0.03279938921332359,
-0.04325038939714432,
0.013786666095256805,
-0.04503164440393448,
0.07236359268426895,
-0.04777001217007637,
0.031697213649749756,
0.04225427657365799,
0.01492627989500761,
-0.01656506210565567,
0.01949346996843815,
-0.03888741508126259,
-0.020167848095297813,
0.056997183710336685,
-0.012069065123796463,
0.027393745258450508,
-0.07042994350194931,
0.08473309874534607,
-0.018902748823165894,
-0.030536266043782234,
-0.04225670173764229,
-0.023062050342559814,
-0.002074477495625615,
-0.057142697274684906,
0.006360028404742479,
-0.029103785753250122,
0.06953512132167816,
-0.03450321406126022,
0.05448126792907715,
-0.10959706455469131,
0.028167789801955223,
-0.026771316304802895,
-0.0342717319726944,
0.04054978862404823,
0.012134706601500511,
-0.009979229420423508,
-0.0170368030667305,
-0.11071250587701797,
-0.05953017622232437,
0.04937340319156647,
-0.01604291796684265,
-0.005025756079703569,
-0.03418532758951187,
0.013987698592245579,
0.030517814680933952,
0.04066057503223419,
-0.07335589081048965,
-0.025265417993068695,
-0.027130072936415672,
0.02633441612124443,
0.0633399710059166,
-0.009296368807554245,
0.007103986106812954,
-0.032528262585401535,
-0.016285257413983345,
0.05878111347556114,
0.02791452594101429,
0.03509538620710373,
0.05759340152144432,
-0.02664019912481308,
0.023045137524604797,
0.11680973321199417,
-0.001372656668536365,
-0.010372118093073368,
0.073048897087574,
0.03203865513205528,
-0.07145727425813675,
0.08079976588487625,
-0.01425680611282587,
0.0265548974275589,
-0.04753878340125084,
0.02751629240810871,
0.02470025047659874,
0.03890996426343918,
0.00537825794890523,
-0.06771700084209442,
-0.051629189401865005,
0.013043107464909554,
-0.04706656560301781,
-0.0827745795249939,
-0.07205861061811447,
-0.07981888204813004,
0.008739768527448177,
-0.07555746287107468,
-0.07821593433618546,
-0.07013390213251114,
0.031148217618465424,
-0.033722177147865295,
-2.3682126326199986e-33,
-0.02587983012199402,
-0.061683692038059235,
0.08343590050935745,
0.02106461301445961,
-0.050842564553022385,
-0.0017942788545042276,
0.016300151124596596,
-0.02653423883020878,
-0.005903335753828287,
-0.0331762433052063,
-0.10049503296613693,
-0.004039866384118795,
0.0653521791100502,
0.015603397972881794,
-0.014419256709516048,
0.05079236999154091,
-0.014240358956158161,
0.0319790318608284,
-0.002556364517658949,
-0.020669352263212204,
0.028352275490760803,
0.08044254779815674,
-0.11541534960269928,
0.05906330421566963,
0.0036692535504698753,
0.04685382917523384,
0.0015642610378563404,
0.04871008172631264,
-0.03205930441617966,
0.027883749455213547,
-0.04890973120927811,
-0.1013040617108345,
-0.08570370078086853,
0.01083020493388176,
-0.07912474125623703,
0.05718237906694412,
0.07571064680814743,
0.02708340622484684,
-0.11868085712194443,
0.11631296575069427,
-0.013832294382154942,
-0.004181531257927418,
-0.023273855447769165,
0.015682484954595566,
0.04496433213353157,
-0.045266442000865936,
-0.041123904287815094,
0.053800538182258606,
0.0110540259629488,
0.02628682553768158,
-0.03104545921087265,
0.07169736921787262,
-0.0640663206577301,
0.014737374149262905,
-0.08581849187612534,
-0.06261921674013138,
-0.11286421120166779,
0.013432158157229424,
0.025469021871685982,
-0.0061414809897542,
-0.022523146122694016,
-0.04965279623866081,
0.0010595315834507346,
0.1010187566280365,
-0.10072465240955353,
0.05851452797651291,
0.04657558724284172,
0.07051753997802734,
-0.036729153245687485,
-0.025769509375095367,
-0.06290023773908615,
-0.02869115211069584,
0.01510168332606554,
-0.013572675175964832,
0.024762414395809174,
0.015096722170710564,
0.08252297341823578,
-0.05577781796455383,
-0.014976691454648972,
-0.0006101478938944638,
0.02043052576482296,
0.06973637640476227,
0.03583310917019844,
0.05550897493958473,
0.07988055795431137,
0.10000257939100266,
0.040323495864868164,
0.03515702858567238,
0.03393630310893059,
-0.05277346074581146,
-0.021033603698015213,
-0.008674098178744316,
-0.028970113024115562,
0.05357546731829643,
-0.03051513247191906,
-4.210912152302626e-8,
-0.060695864260196686,
-0.022664770483970642,
0.1545935869216919,
-0.016186697408556938,
0.08589166402816772,
0.026438217610120773,
0.0018560780445113778,
0.04196511581540108,
-0.08503081649541855,
-0.030625658109784126,
0.10130264610052109,
0.07654598355293274,
-0.02020159736275673,
-0.043100278824567795,
-0.012421847321093082,
-0.014100886881351471,
0.01482627633959055,
-0.04820307716727257,
0.034771934151649475,
-0.010123488493263721,
0.02565426006913185,
-0.04087298363447189,
0.018091708421707153,
-0.03752076253294945,
0.045771028846502304,
-0.06546854972839355,
-0.046898167580366135,
0.0396767258644104,
0.04034765064716339,
-0.013419239781796932,
0.02605387195944786,
-0.04618482664227486,
0.08400178700685501,
-0.004349735099822283,
0.021589260548353195,
-0.017918959259986877,
0.004850638564676046,
-0.02251729555428028,
-0.016736816614866257,
-0.042907219380140305,
0.004111448768526316,
0.05724972113966942,
0.010859592817723751,
0.005383100360631943,
-0.04616757854819298,
0.03383535146713257,
-0.10940487682819366,
-0.09210174530744553,
0.06104573979973793,
-0.04421352967619896,
-0.0009797250386327505,
-0.03425953909754753,
-0.053528159856796265,
0.06202949211001396,
0.11992849409580231,
-0.022121097892522812,
-0.009474470280110836,
-0.09142738580703735,
0.0638270154595375,
0.05919409543275833,
-0.04210749268531799,
0.004790888633579016,
-0.11888209730386734,
-0.02906196564435959
] | 0 | 0 | 0 | 0 |
|
Context-aware Dynamic Pruning for Speech Foundation Models
|
https://openreview.net/forum?id=u2QdCiOgwA
|
[
"Masao Someki",
"Yifan Peng",
"Siddhant Arora",
"Markus Müller",
"Athanasios Mouchtaris",
"Grant Strimel",
"Jing Liu",
"Shinji Watanabe"
] |
Poster
|
Foundation models, such as large language models, have achieved remarkable success in natural language processing and are evolving into models capable of handling multiple modalities.
Listening ability, in particular, is crucial for many applications, leading to research on building speech foundation models. However, the high computational cost of these large models presents a significant challenge for real-world applications. Although substantial efforts have been made to reduce computational costs, such as through pruning techniques, the majority of these approaches are applied primarily during the training phase for specific downstream tasks. In this study, we hypothesize that optimal pruned networks may vary based on contextual factors such as speaker characteristics, languages, and tasks. To address this, we propose a dynamic pruning technique that adapts to these contexts during inference without altering the underlying model. We demonstrated that we could successfully reduce inference time by approximately 30\% while maintaining accuracy in multilingual/multi-task scenarios. We also found that the obtained pruned structure offers meaningful interpretations based on the context, e.g., task-related information emerging as the dominant factor for efficient pruning.
|
Pruning, Speech Foundation Model, Automatic Speech Recognition, Speech Translation
| null | 11,953 | null |
[
0.0003132678393740207,
-0.06784184277057648,
0.004159791860729456,
0.017145831137895584,
0.09224877506494522,
-0.0017782255308702588,
0.02094857580959797,
-0.023485524579882622,
0.015671906992793083,
-0.10483356565237045,
-0.05511345714330673,
-0.06349608302116394,
0.01724940352141857,
0.06358682364225388,
0.015271512791514397,
-0.015977688133716583,
0.022187400609254837,
0.11771921068429947,
-0.1247214674949646,
-0.10268136858940125,
0.002719338284805417,
0.04128984361886978,
0.025995654985308647,
0.02783428505063057,
0.040393583476543427,
-0.016054945066571236,
-0.06769374012947083,
-0.05699459835886955,
0.056486986577510834,
0.010812288150191307,
0.06531795114278793,
0.03396407887339592,
0.0693649873137474,
0.016045553609728813,
-0.05874178558588028,
0.04403810575604439,
-0.043330270797014236,
0.05380432680249214,
-0.018014006316661835,
-0.008435015566647053,
-0.045937731862068176,
0.040002044290304184,
-0.030411824584007263,
0.013709090650081635,
0.03216004744172096,
-0.00013446925731841475,
-0.04710978642106056,
-0.012451008893549442,
0.030617229640483856,
0.06525325775146484,
-0.06612846255302429,
0.013909189030528069,
0.00027660190244205296,
0.09211312234401703,
-0.035855669528245926,
0.027066439390182495,
-0.033033598214387894,
0.07130005210638046,
-0.02221762016415596,
-0.00898241437971592,
-0.027403144165873528,
-0.16673140227794647,
-0.06277035176753998,
-0.00916288048028946,
-0.02707732655107975,
0.0010214251233264804,
-0.0005179521394893527,
0.03749528154730797,
-0.01790194772183895,
0.04883405938744545,
-0.01759970188140869,
0.09393473714590073,
-0.04717393219470978,
0.07646748423576355,
0.058320675045251846,
0.06972381472587585,
0.03445963189005852,
-0.048794448375701904,
0.025519708171486855,
0.027970964089035988,
-0.0016910803969949484,
-0.013608168810606003,
0.014829641208052635,
-0.04147346690297127,
0.12200719118118286,
-0.023472096771001816,
0.017878640443086624,
-0.019564973190426826,
0.002267330652102828,
-0.012343032285571098,
-0.10743886232376099,
-0.058622658252716064,
0.06111622601747513,
0.04658667370676994,
0.049331605434417725,
0.06834227591753006,
-0.05247495695948601,
-0.05200947821140289,
-0.02700427733361721,
0.08473934233188629,
-0.029066860675811768,
0.03195580467581749,
0.0018602749332785606,
-0.10248861461877823,
-0.04482905566692352,
0.005481655243784189,
-0.0337950624525547,
0.05454041808843613,
-0.0031405126210302114,
-0.07003916054964066,
0.031317148357629776,
0.08952808380126953,
-0.00664279842749238,
0.01807975210249424,
0.027717990800738335,
-0.012601232156157494,
-0.007270346861332655,
-0.07341041415929794,
0.04384820535778999,
0.0436541847884655,
0.0065714651718735695,
0.08811870962381363,
-0.030502527952194214,
-0.0231606587767601,
-0.000013748752280662302,
-0.05368201434612274,
-0.00678190728649497,
1.0360768637239256e-33,
0.07046963274478912,
0.09749258309602737,
0.06273426860570908,
0.019864197820425034,
0.09238216280937195,
-0.02297060936689377,
-0.0009863108862191439,
0.037756651639938354,
0.02015252783894539,
0.0030706769321113825,
0.014676939696073532,
0.03852327540516853,
-0.06028653308749199,
0.00404455978423357,
0.03951965644955635,
-0.034460391849279404,
-0.00423450255766511,
0.10838665068149567,
-0.03653492033481598,
-0.03432358428835869,
0.034509338438510895,
0.03473172336816788,
-0.026483934372663498,
-0.047778695821762085,
0.0946611538529396,
-0.021063577383756638,
0.09036363661289215,
-0.03552006930112839,
-0.06615031510591507,
-0.003780120052397251,
-0.09892725944519043,
-0.04053928703069687,
-0.06024368479847908,
0.05656161159276962,
0.04096085950732231,
-0.0839184895157814,
0.02084733545780182,
-0.01660141721367836,
-0.01048040110617876,
-0.05134202912449837,
-0.06788477301597595,
0.0011117657413706183,
0.05011152848601341,
-0.0081774378195405,
-0.08606120944023132,
-0.09793911129236221,
-0.007194361183792353,
0.016971208155155182,
-0.02536611072719097,
0.010760181583464146,
0.04571130499243736,
0.022434571757912636,
0.04112280532717705,
-0.008355243131518364,
-0.05932428315281868,
-0.0685926228761673,
0.01057138666510582,
0.013120199553668499,
0.062117382884025574,
0.04723702371120453,
0.0024352672044187784,
0.02052822709083557,
-0.028096290305256844,
0.10131220519542694,
0.08026246726512909,
0.012704706750810146,
-0.11240837723016739,
0.03156698867678642,
0.04849601909518242,
-0.03268290311098099,
-0.04815211519598961,
0.042410142719745636,
0.024608930572867393,
-0.05089978501200676,
0.019178181886672974,
0.03015081025660038,
0.0647941380739212,
-0.039111699908971786,
-0.003858729265630245,
0.06277406215667725,
-0.051765576004981995,
0.020288201048970222,
0.023734500631690025,
-0.013517019338905811,
-0.05154717341065407,
-0.022964246571063995,
0.05372878909111023,
-0.09588434547185898,
-0.016027504578232765,
-0.03144824877381325,
-0.05747310444712639,
-0.003689178265631199,
-0.024625621736049652,
-0.02512148767709732,
0.0523165687918663,
-1.4078253909680393e-33,
-0.029806403443217278,
0.016495320945978165,
0.020545298233628273,
0.023677373304963112,
-0.00929687637835741,
-0.011813482269644737,
0.0437849722802639,
0.08301327377557755,
-0.08661437779664993,
-0.12928993999958038,
-0.04785196855664253,
-0.04205156862735748,
0.09245914220809937,
-0.021926026791334152,
0.01493356004357338,
0.01615665666759014,
-0.022310622036457062,
0.03263514116406441,
0.04143966734409332,
0.09397468715906143,
0.013517402112483978,
0.07096583396196365,
-0.08092217892408371,
-0.018595034256577492,
0.006861393339931965,
-0.014713200740516186,
-0.038727715611457825,
0.0005098871188238263,
-0.010645371861755848,
-0.04093604534864426,
-0.11316592246294022,
0.0002109218912664801,
-0.09123458713293076,
-0.06052341312170029,
-0.052824974060058594,
0.0483592227101326,
-0.015933971852064133,
-0.029732542112469673,
0.04048041254281998,
0.10195747017860413,
0.03437815606594086,
-0.04316436126828194,
-0.058737657964229584,
-0.05699761211872101,
0.013403233140707016,
0.005118906497955322,
-0.07562821358442307,
-0.007858737371861935,
-0.02836994268000126,
-0.015780068933963776,
0.012247317470610142,
0.05906132981181145,
-0.012244059704244137,
0.01778561621904373,
-0.0019686692394316196,
-0.10174999386072159,
0.0020392185542732477,
-0.049056943506002426,
0.07151231169700623,
-0.04275985062122345,
-0.015631182119250298,
-0.009674095548689365,
0.02489200048148632,
-0.02338428981602192,
0.008869126439094543,
-0.06569185853004456,
0.006777443923056126,
0.01776571199297905,
0.011694847606122494,
-0.01509551890194416,
-0.014290925115346909,
0.03019608184695244,
-0.01522686704993248,
0.013026993721723557,
-0.04714788869023323,
0.023858098313212395,
-0.02079852484166622,
-0.12571826577186584,
-0.05425553023815155,
-0.0034496283624321222,
0.0455731600522995,
-0.019198672845959663,
0.012826981022953987,
-0.06368467956781387,
0.09457654505968094,
0.13881097733974457,
0.024536238983273506,
0.05374262109398842,
0.10356716811656952,
0.06993013620376587,
-0.03057723492383957,
0.016866516321897507,
0.0438355915248394,
0.10393185168504715,
-0.06038684770464897,
-4.076347792647539e-8,
-0.058486469089984894,
0.10740041732788086,
-0.052633680403232574,
0.033628955483436584,
0.00963558815419674,
-0.08764111995697021,
-0.019488146528601646,
0.05742673948407173,
-0.01089791115373373,
-0.03725457191467285,
0.05013169348239899,
-0.03903387486934662,
0.006147455889731646,
0.0009273046744056046,
-0.025013819336891174,
0.05203358456492424,
-0.017097491770982742,
0.0545540414750576,
-0.0033156301360577345,
-0.021230680868029594,
0.0900801420211792,
0.07964100688695908,
-0.04296163469552994,
0.09120761603116989,
0.04710947349667549,
-0.05568339303135872,
0.057815466076135635,
0.11542654782533646,
0.052991196513175964,
-0.018307799473404884,
-0.0458516925573349,
0.049816813319921494,
-0.10660005360841751,
-0.04291429743170738,
-0.015110166743397713,
0.072608582675457,
0.00270750536583364,
-0.009825450368225574,
-0.048427559435367584,
-0.03662729635834694,
0.01730114221572876,
0.006730074528604746,
-0.08191930502653122,
0.08440058678388596,
0.05807218328118324,
-0.015002565458416939,
-0.04643188416957855,
-0.05304042622447014,
0.06563972681760788,
-0.05242861062288284,
0.022932544350624084,
0.02845763973891735,
-0.05547460541129112,
-0.03544953465461731,
0.05204879119992256,
0.03345368802547455,
0.009281224571168423,
0.03354907035827637,
-0.022703278809785843,
-0.007010521832853556,
-0.00744252884760499,
0.0315413735806942,
-0.07703892886638641,
0.01528103370219469
] | 0 | 0 | 0 | 0 |
|
Infinite-Resolution Integral Noise Warping for Diffusion Models
|
https://openreview.net/forum?id=Y6LPWBo2HP
|
[
"Yitong Deng",
"Winnie Lin",
"Lingxiao Li",
"Dmitriy Smirnov",
"Ryan D Burgert",
"Ning Yu",
"Vincent Dedun",
"Mohammad H. Taghavi"
] |
Poster
|
Adapting pretrained image-based diffusion models to generate temporally consistent videos has become an impactful generative modeling research direction. Training-free noise-space manipulation has proven to be an effective technique, where the challenge is to preserve the Gaussian white noise distribution while adding in temporal consistency. Recently, Chang et al. (2024) formulated this problem using an integral noise representation with distribution-preserving guarantees, and proposed an upsampling-based algorithm to compute it. However, while their mathematical formulation is advantageous, the algorithm incurs a high computational cost. Through analyzing the limiting-case behavior of their algorithm as the upsampling resolution goes to infinity, we develop an alternative algorithm that, by gathering increments of multiple Brownian bridges, achieves their infinite-resolution accuracy while simultaneously reducing the computational cost by orders of magnitude. We prove and experimentally validate our theoretical claims, and demonstrate our method's effectiveness in real-world applications. We further show that our method can readily extend to the 3-dimensional space.
|
diffusion models; video generation; temporal consistency; multi-view consistency; noise warping; white Gaussian noise
| null | 11,949 |
2411.01212
|
[
-0.1051664724946022,
-0.10876865684986115,
0.085186667740345,
-0.010783033445477486,
0.023622555658221245,
-0.05946648493409157,
-0.0561140701174736,
-0.11387618631124496,
0.02944614179432392,
-0.07941212505102158,
-0.0987306460738182,
-0.016706902533769608,
0.047637857496738434,
0.047117579728364944,
-0.10466791689395905,
-0.00024490567739121616,
0.03495693951845169,
0.11865799874067307,
-0.12226106971502304,
-0.00966706033796072,
-0.007351688574999571,
-0.001696772058494389,
-0.0063159228302538395,
-0.008351673372089863,
0.09422127157449722,
-0.027866238728165627,
0.07091187685728073,
-0.053833238780498505,
0.030386798083782196,
-0.03899727389216423,
0.07937809079885483,
-0.027602476999163628,
-0.019147155806422234,
-0.014102858491241932,
-0.017581474035978317,
-0.018825026229023933,
-0.04970461502671242,
0.03868497535586357,
0.06890300661325455,
-0.015554718673229218,
0.04676142334938049,
0.0735592320561409,
0.025947330519557,
-0.011938128620386124,
0.04326045140624046,
0.0013707529287785292,
0.03856206312775612,
-0.06385479122400284,
0.07099863886833191,
0.019195618107914925,
0.012386632151901722,
0.021389415487647057,
-0.04767298698425293,
0.0852290615439415,
-0.03197348862886429,
0.019491609185934067,
0.06696687638759613,
0.02346278540790081,
-0.008132239803671837,
0.027034206315875053,
0.022427603602409363,
-0.08381132036447525,
0.0005069790640845895,
-0.01775362156331539,
-0.055539730936288834,
-0.015120431780815125,
0.016548676416277885,
0.02475840412080288,
-0.06416766345500946,
0.03189108520746231,
-0.07031089067459106,
0.09992461651563644,
-0.010083234868943691,
0.08234445750713348,
0.02142854779958725,
0.015122948214411736,
0.058299556374549866,
0.038057487457990646,
0.02881254442036152,
-0.029358776286244392,
0.10753589868545532,
-0.023320157080888748,
0.06696319580078125,
-0.011542100459337234,
0.07228586822748184,
0.008124606683850288,
-0.004489589482545853,
0.061952315270900726,
0.013765811920166016,
0.07109824568033218,
-0.06593457609415054,
0.052037592977285385,
-0.12511484324932098,
-0.02170776203274727,
-0.03598196431994438,
-0.010138657875359058,
0.02700844407081604,
-0.06974969804286957,
0.07220610976219177,
0.10528095066547394,
0.0006681618979200721,
-0.029031433165073395,
0.0642189234495163,
-0.010616141371428967,
0.1176709458231926,
-0.06391967087984085,
0.10251854360103607,
0.05037868395447731,
0.024418577551841736,
-0.02270738035440445,
0.07519174367189407,
-0.048923734575510025,
0.05297822132706642,
0.0466228611767292,
0.024768782779574394,
-0.022294435650110245,
-0.0038743563927710056,
0.008054699748754501,
-0.027138279750943184,
0.0017818888882175088,
-0.015284513123333454,
0.01640847511589527,
-0.048985403031110764,
0.03675193339586258,
-0.005369578022509813,
-0.06004875898361206,
-0.025000544264912605,
8.735130089412223e-35,
-0.0685952752828598,
-0.02589578554034233,
0.06102366000413895,
0.06114460155367851,
0.1123591884970665,
0.03017566353082657,
-0.045497696846723557,
-0.09511411935091019,
0.018508082255721092,
-0.013715698383748531,
-0.002026757225394249,
-0.04438662528991699,
-0.056289445608854294,
0.033896561712026596,
0.004518014378845692,
0.00019794257241301239,
-0.031904611736536026,
0.09247598052024841,
0.0899948850274086,
0.02501152828335762,
0.06534195691347122,
-0.03222766891121864,
0.006518782116472721,
0.017793962731957436,
-0.05332857742905617,
-0.03896269574761391,
0.027972135692834854,
-0.048697151243686676,
0.10941625386476517,
-0.008181133307516575,
-0.08931595087051392,
-0.007546460721641779,
-0.007678166497498751,
0.0023692711256444454,
0.013729045167565346,
-0.010674800723791122,
0.024424707517027855,
-0.016927096992731094,
-0.09315589815378189,
-0.02429523505270481,
0.011273840442299843,
0.08396916091442108,
-0.003030286869034171,
-0.03733697533607483,
-0.11910799145698547,
-0.05177051201462746,
0.08939066529273987,
0.020112449303269386,
-0.01757952757179737,
-0.006171742919832468,
0.05044655129313469,
-0.03792453557252884,
-0.00876516755670309,
-0.07678348571062088,
-0.025338031351566315,
0.01620345376431942,
0.09574587643146515,
-0.06422584503889084,
0.015961328521370888,
-0.0412684865295887,
0.020275302231311798,
-0.011382423341274261,
-0.024263635277748108,
0.0628710463643074,
0.061233822256326675,
-0.02834324538707733,
-0.02035055123269558,
0.05858568102121353,
-0.03185240179300308,
-0.00025827967328950763,
-0.038872990757226944,
0.006591561250388622,
-0.04167703539133072,
-0.06844229251146317,
-0.011989264748990536,
-0.017587004229426384,
-0.019899537786841393,
0.0437282994389534,
0.014853935688734055,
0.042126432061195374,
-0.1039862111210823,
-0.05719756707549095,
0.028006091713905334,
-0.02580282650887966,
0.00799214094877243,
-0.008076136931777,
-0.03474729508161545,
0.022569317370653152,
-0.042442601174116135,
-0.12105809897184372,
-0.05372006818652153,
0.010451997630298138,
-0.007859739474952221,
0.02338443137705326,
-0.006596241146326065,
7.2077869761316745e-34,
0.007193208672106266,
0.08564046770334244,
0.09353503584861755,
0.06706220656633377,
0.013126270845532417,
0.01451951265335083,
-0.015324137173593044,
0.059681233018636703,
0.10518389195203781,
-0.0765465646982193,
-0.05827559903264046,
-0.020152727141976357,
-0.04810728505253792,
-0.02547847293317318,
-0.01346119586378336,
-0.09025204181671143,
0.03412902355194092,
-0.006226884666830301,
-0.10589224845170975,
0.011374171823263168,
0.04183002561330795,
0.03741501644253731,
-0.028306039050221443,
-0.020505405962467194,
-0.05735129117965698,
0.0783296450972557,
-0.00956552941352129,
0.08732789009809494,
0.003920595161616802,
0.03281703218817711,
-0.03522243723273277,
-0.07035904377698898,
0.012278741225600243,
-0.04955156892538071,
-0.10197991132736206,
0.02886779047548771,
0.06659027934074402,
0.015671484172344208,
-0.0012785999570041895,
0.054191797971725464,
0.025914421305060387,
0.001909410348162055,
-0.03430852293968201,
0.00786882545799017,
0.007865950465202332,
-0.036358509212732315,
-0.03372412919998169,
0.03606389835476875,
-0.023317568004131317,
0.02394108846783638,
-0.017922906205058098,
0.08610326796770096,
-0.035698216408491135,
0.04325869306921959,
0.017992332577705383,
-0.060026224702596664,
-0.07302150130271912,
-0.03698645904660225,
0.11023572832345963,
0.00000914784050110029,
-0.05618373677134514,
-0.02276463434100151,
-0.06854908913373947,
-0.05736863240599632,
-0.05548907071352005,
0.05554228276014328,
-0.07452728599309921,
-0.012097544968128204,
-0.006105675362050533,
0.06461729854345322,
0.03503178060054779,
-0.02381916344165802,
-0.07111784815788269,
0.024267980828881264,
0.029209032654762268,
-0.059504054486751556,
0.09173078089952469,
-0.05109596252441406,
0.0713099017739296,
-0.04935001954436302,
-0.026469027623534203,
-0.01665533147752285,
0.05509676784276962,
0.024633105844259262,
0.06629958003759384,
0.032312579452991486,
0.047375842928886414,
-0.024609696120023727,
0.07642960548400879,
-0.025026049464941025,
0.025341486558318138,
0.06173718348145485,
0.05180598050355911,
0.024932458996772766,
-0.10378237068653107,
-4.683671050997873e-8,
-0.07636906206607819,
0.05518840625882149,
-0.04708952084183693,
-0.04354563355445862,
0.0023285995703190565,
-0.06700490415096283,
0.05923891440033913,
0.04188276827335358,
0.028350548818707466,
-0.10248256474733353,
0.08727892488241196,
-0.037843212485313416,
-0.04901896044611931,
0.028265926986932755,
-0.0602140836417675,
-0.032131217420101166,
0.0020951037295162678,
0.009508941322565079,
-0.028464851900935173,
0.004250214900821447,
-0.13353268802165985,
0.048177413642406464,
0.026079436764121056,
-0.02843894623219967,
0.07802386581897736,
-0.0028943237848579884,
-0.03106546215713024,
0.010248223319649696,
0.04741504788398743,
-0.0008111747447401285,
-0.0037323152646422386,
0.030707715079188347,
0.03461138904094696,
0.00020943359413649887,
0.0007238795515149832,
0.010482672601938248,
-0.008366987109184265,
0.05072156339883804,
-0.030767368152737617,
-0.07199351489543915,
-0.04275822639465332,
0.06966531276702881,
0.03756404668092728,
-0.03196666017174721,
-0.0030008007306605577,
0.060726962983608246,
0.05323932692408562,
0.019974272698163986,
-0.006612689234316349,
0.05465730279684067,
0.07296909391880035,
0.04910171031951904,
0.010299100540578365,
-0.007658918853849173,
0.079593725502491,
0.0343751423060894,
0.06550981849431992,
-0.04562324285507202,
0.024148326367139816,
0.027630096301436424,
0.0117791211232543,
-0.0028580024372786283,
-0.03812865540385246,
0.024983208626508713
] | 0 | 0 | 0 | 0 |
|
Injective flows for star-like manifolds
|
https://openreview.net/forum?id=Jyh0DR4fFE
|
[
"Marcello Massimo Negri",
"Jonathan Aellen",
"Volker Roth"
] |
Poster
|
Normalizing Flows (NFs) are powerful and efficient models for density estimation. When modeling densities on manifolds, NFs can be generalized to injective flows but the Jacobian determinant becomes computationally prohibitive. Current approaches either consider bounds on the log-likelihood or rely on some approximations of the Jacobian determinant. In contrast, we propose injective flows for star-like manifolds and show that for such manifolds we can compute the Jacobian determinant exactly and efficiently. This aspect is particularly relevant for variational inference settings, where no samples are available and only some unnormalized target is known. Among many, we showcase the relevance of modeling densities on star-like manifolds in two settings. Firstly, we introduce a novel Objective Bayesian approach for penalized likelihood models by interpreting level-sets of the penalty as star-like manifolds. Secondly, we consider probabilistic mixing models and introduce a general method for variational inference by defining the posterior of mixture weights on the probability simplex.
|
Normalizing Flows, Injective Flows, Bayesian Inference, Variational Inference, Objective Bayes
|
We propose injective flows for star-like manifolds and show that they allow for an exact and efficient Jacobian determinant.
| 11,933 |
2406.09116
|
[
-0.0034187620040029287,
-0.07071609795093536,
0.0982554480433464,
-0.008206506259739399,
0.04769124463200569,
0.02050163410604,
0.044497206807136536,
-0.06702890247106552,
0.019548408687114716,
-0.007181942462921143,
-0.026517177000641823,
-0.03084135800600052,
0.02655915915966034,
0.05308742821216583,
0.003446432761847973,
-0.04340614378452301,
0.034997373819351196,
0.09353905916213989,
-0.043115947395563126,
0.04046395793557167,
-0.006443734746426344,
0.14019310474395752,
-0.0393587164580822,
0.019470909610390663,
0.07062532007694244,
-0.04829424247145653,
0.06857190281152725,
0.008836270309984684,
-0.0032740982715040445,
-0.038401372730731964,
0.04009620472788811,
0.04751969501376152,
-0.09526940435171127,
-0.0058286674320697784,
-0.0573955774307251,
0.04068711772561073,
-0.01854182593524456,
0.01419026218354702,
-0.024396652355790138,
0.03027229756116867,
-0.008792184293270111,
0.026673251762986183,
0.03312400355935097,
0.027229949831962585,
-0.005826159846037626,
-0.012472358532249928,
-0.08078005164861679,
-0.053869184106588364,
0.03572637587785721,
0.03654354065656662,
-0.02351432852447033,
-0.060684461146593094,
-0.04983312264084816,
-0.0127452677115798,
-0.02977070026099682,
-0.06380144506692886,
0.04130353778600693,
-0.1014976054430008,
0.08886200189590454,
-0.001649941666983068,
-0.003985071089118719,
-0.07432128489017487,
-0.05216461420059204,
0.02521640621125698,
0.01116950437426567,
-0.060373157262802124,
-0.04285041615366936,
0.09088066220283508,
-0.014542707242071629,
0.09327322989702225,
-0.05270358920097351,
0.09911198168992996,
-0.06917770206928253,
-0.05850943922996521,
-0.0411103293299675,
-0.02805161103606224,
0.018693698570132256,
-0.04908641800284386,
-0.030361631885170937,
-0.02449817769229412,
0.03497977927327156,
-0.008236894384026527,
0.011313962750136852,
0.007079634815454483,
0.03268037363886833,
0.06744634360074997,
-0.07566908001899719,
0.04952069744467735,
0.03538621962070465,
-0.005669667385518551,
-0.06667716801166534,
0.02266163006424904,
-0.06419811397790909,
-0.039360299706459045,
0.02644970640540123,
0.014947577379643917,
-0.028266413137316704,
-0.0880998820066452,
0.1313038021326065,
0.11295779794454575,
-0.012067736126482487,
-0.03159993141889572,
0.06678853929042816,
0.022718200460076332,
0.04895582050085068,
-0.051400065422058105,
0.01965806633234024,
0.018512355163693428,
-0.018556904047727585,
-0.00922226533293724,
0.04676136374473572,
-0.045568399131298065,
0.05861950293183327,
-0.007265189196914434,
-0.07452701032161713,
-0.018370918929576874,
0.07584041357040405,
0.004889458417892456,
-0.012255401350557804,
-0.0032527840230613947,
-0.046304307878017426,
0.03582806885242462,
0.020153872668743134,
-0.010883412323892117,
0.013193623162806034,
0.06167050078511238,
-0.024543123319745064,
1.293524912068888e-33,
0.045594584196805954,
-0.046977728605270386,
0.06069415062665939,
0.12722896039485931,
-0.03211238980293274,
0.057531408965587616,
-0.03663894906640053,
-0.02960434928536415,
0.004829653538763523,
-0.011891774833202362,
-0.04269183427095413,
0.013953888788819313,
-0.09294231981039047,
0.06022598594427109,
-0.018850840628147125,
-0.03277900442481041,
0.05476462095975876,
0.051097843796014786,
-0.06835341453552246,
0.006942731328308582,
0.01969580166041851,
-0.004118597134947777,
-0.0665985643863678,
-0.01606687903404236,
-0.06214079633355141,
0.0020461278036236763,
0.014770914800465107,
0.013000131584703922,
-0.04770364984869957,
-0.011902578175067902,
-0.025443287566304207,
-0.003060446586459875,
-0.031409040093421936,
0.05916222184896469,
0.07389289885759354,
-0.0016140233492478728,
-0.04233361408114433,
0.03980676457285881,
-0.052405886352062225,
-0.1048109382390976,
-0.015307466499507427,
0.04651554301381111,
-0.028330599889159203,
-0.020043637603521347,
-0.14335623383522034,
-0.015806084498763084,
0.023919258266687393,
0.00011701359471771866,
0.05714957043528557,
-0.040977075695991516,
0.0720609650015831,
0.056981686502695084,
0.037486907094717026,
-0.04741193354129791,
-0.059058330953121185,
0.0020054555498063564,
0.015609608963131905,
0.0021767152938991785,
0.03546522557735443,
0.0006084741908125579,
-0.016221897676587105,
0.04405898228287697,
0.011704593896865845,
-0.08147218823432922,
0.038038138300180435,
0.12140205502510071,
0.022526076063513756,
0.03418152779340744,
-0.007890823297202587,
-0.008459441363811493,
-0.022939082235097885,
0.053927984088659286,
0.026645848527550697,
-0.023565689101815224,
0.0016026787925511599,
0.09401974827051163,
-0.030980104580521584,
-0.027538800612092018,
0.0354856513440609,
0.01235194317996502,
-0.05103190988302231,
-0.01654241234064102,
-0.057619158178567886,
0.014107578434050083,
0.037105441093444824,
-0.037728436291217804,
-0.003825474763289094,
-0.019590334966778755,
-0.018140563741326332,
-0.06547845900058746,
-0.01093202643096447,
-0.048955563455820084,
-0.054027993232011795,
-0.054859600961208344,
-0.011755755171179771,
-3.248964264601075e-33,
-0.06563658267259598,
0.009128697216510773,
0.027277342975139618,
0.0841865986585617,
0.03851626440882683,
-0.05152195692062378,
0.042377032339572906,
0.04381091892719269,
0.024084212258458138,
-0.06205349043011665,
-0.06258098036050797,
0.08413542062044144,
0.04843715578317642,
-0.06265588849782944,
0.052327536046504974,
0.02073616348206997,
-0.06984277814626694,
0.05422802269458771,
-0.0672435462474823,
-0.016747483983635902,
0.014645918272435665,
0.030910061672329903,
0.004093748517334461,
-0.014309248887002468,
-0.06648676097393036,
0.030744997784495354,
-0.07715120166540146,
0.049059946089982986,
-0.07593526691198349,
-0.003874917747452855,
-0.02214631997048855,
-0.07199765741825104,
-0.06026165559887886,
-0.00583526398986578,
-0.09910894185304642,
0.03840326517820358,
0.03696591779589653,
0.11902512609958649,
-0.0951032042503357,
0.04020446538925171,
-0.06515093892812729,
0.0035163206048309803,
-0.04298793151974678,
0.09511613845825195,
0.014394288882613182,
-0.03797611594200134,
0.06189471110701561,
-0.07491670548915863,
0.0321647934615612,
-0.012762845493853092,
-0.10512430965900421,
0.09311302751302719,
0.02966098114848137,
0.1090996041893959,
-0.035635627806186676,
-0.03690491244196892,
-0.03710738569498062,
-0.0793699100613594,
0.09840911626815796,
-0.017091890797019005,
-0.03769697621464729,
-0.06227799877524376,
-0.07503121346235275,
0.004192758817225695,
-0.04869711399078369,
-0.12242302298545837,
-0.047840416431427,
-0.07125190645456314,
-0.032331954687833786,
0.08988962322473526,
0.013021895661950111,
-0.020544929429888725,
0.03361388295888901,
0.015141093172132969,
0.036605484783649445,
0.006131282076239586,
0.06507446616888046,
-0.03522036597132683,
-0.0033114280086010695,
0.023113977164030075,
0.07539783418178558,
-0.02417367696762085,
0.0626981183886528,
0.032612964510917664,
0.1338585466146469,
0.019619908183813095,
0.023927032947540283,
-0.1436469554901123,
0.012242899276316166,
-0.02677139826118946,
-0.04252732917666435,
-0.006263336166739464,
0.001032047439366579,
0.04959241300821304,
-0.05952000990509987,
-3.705857309910243e-8,
-0.014727933332324028,
-0.013792933896183968,
0.006488831248134375,
-0.055549658834934235,
0.02779478207230568,
-0.0006306961877271533,
-0.061131034046411514,
0.08978953212499619,
0.03266039118170738,
0.007283169776201248,
0.07917418330907822,
0.00520807271823287,
-0.013429849408566952,
-0.06643430888652802,
-0.05362481251358986,
0.05446496605873108,
0.012272185645997524,
0.05254627764225006,
-0.014998193830251694,
0.05111358314752579,
0.027670780196785927,
0.08102347701787949,
-0.0048218825832009315,
-0.006080446299165487,
0.08058420568704605,
-0.08825336396694183,
0.040438611060380936,
0.04240107163786888,
0.05889998376369476,
-0.053965386003255844,
-0.05706758797168732,
0.046215832233428955,
-0.03234430402517319,
0.0019956761971116066,
-0.05734699219465256,
0.04537009447813034,
-0.017607899382710457,
0.033335015177726746,
-0.025246554985642433,
-0.0033241084311157465,
0.023963723331689835,
0.013670933432877064,
-0.04994748532772064,
-0.002379615092650056,
0.033393215388059616,
0.19364382326602936,
0.029403842985630035,
-0.06140199676156044,
-0.015581995248794556,
0.08189276605844498,
-0.011023856699466705,
0.006404837127774954,
0.006082614418119192,
0.06811411678791046,
-0.004330684896558523,
0.019590482115745544,
-0.03556271642446518,
0.06420251727104187,
0.01897282712161541,
0.014408799819648266,
-0.02388916164636612,
0.04192197695374489,
-0.007865960709750652,
-0.07663223892450333
] | 0 | 0 | 0 | 0 |
|
Mixture of In-Context Prompters for Tabular PFNs
|
https://openreview.net/forum?id=2fojNANZSv
|
[
"Derek Qiang Xu",
"F Olcay Cirit",
"Reza Asadi",
"Yizhou Sun",
"Wei Wang"
] |
Poster
|
Recent benchmarks find In-Context Learning (ICL) outperforms both deep learning and tree-based algorithms on small tabular datasets. However, on larger datasets, ICL for tabular learning suffers in both efficiency and effectiveness. In terms of efficiency, transformers incur linear space and quadratic time complexity w.r.t. context size. In terms of effectiveness, contexts at inference encounter distribution shift compared to contexts from pretraining. We propose MixturePFN, which extends Sparse Mixture of Experts to the state-of-the-art ICL for tabular learning model. Specifically, MixturePFN finetunes a specialized ICL expert on each cluster of tabular data and routes new test samples to appropriate experts at inference. MixturePFN supports constant-size contexts by splitting large training datasets into more manageable clusters. MixturePFN addresses distribution shift by finetuning an expert on each training dataset cluster via bootstrapping. Extensive experimental results shows MixturePFN outperforms 19 baselines both in mean rank and as the Condorcet winner across 36 diverse tabular datasets under both accuracy and F1 score with statistical significance.
|
Prior-Fitted Networks, Tabular Learning, Sparse Mixture of Experts.
|
We propose a mixture of prompters technique for tabular in-context learning.
| 11,928 |
2405.16156
|
[
-0.02055584080517292,
-0.08959954977035522,
0.034342993050813675,
0.07018731534481049,
0.08912235498428345,
0.002632993971928954,
0.024449089542031288,
-0.021088795736432076,
-0.027663221582770348,
-0.04481123760342598,
-0.05957775563001633,
-0.0797920972108841,
0.0022699080873280764,
0.03763359785079956,
-0.05746668204665184,
-0.0376967191696167,
0.08457178622484207,
0.0631885752081871,
-0.12536337971687317,
-0.013189846649765968,
-0.028356585651636124,
0.003863521618768573,
0.061306532472372055,
-0.04552753269672394,
0.0028391198720782995,
0.07188811898231506,
-0.02421683259308338,
-0.06608359515666962,
0.03915908560156822,
-0.013381382450461388,
0.03112727589905262,
0.07967476546764374,
0.0103734927251935,
0.04815094172954559,
-0.05063900724053383,
0.06214384734630585,
-0.025601426139473915,
0.08343485742807388,
0.027844201773405075,
0.032748159021139145,
-0.0019566102419048548,
-0.02910338155925274,
0.00298442249186337,
0.009542486630380154,
0.05869007855653763,
0.013081553392112255,
0.02196136862039566,
-0.032689180225133896,
0.04502813145518303,
0.027136122807860374,
-0.06077555567026138,
0.03203994035720825,
-0.015184248797595501,
0.04567093029618263,
-0.027862735092639923,
0.06834757328033447,
0.057726528495550156,
0.00260725780390203,
0.016217147931456566,
0.05369985103607178,
-0.08328273892402649,
-0.09756600856781006,
-0.049153078347444534,
0.006960432045161724,
-0.04669509455561638,
0.04540299251675606,
-0.012282834388315678,
0.04913660138845444,
0.05199175328016281,
-0.05724518001079559,
0.005625533405691385,
0.11198842525482178,
-0.04208884760737419,
0.07807502150535583,
-0.01483960822224617,
-0.003001187928020954,
0.06597120314836502,
-0.04497220739722252,
0.07226339727640152,
-0.06200551614165306,
0.02059061825275421,
0.08102943748235703,
-0.002633335068821907,
-0.04465727135539055,
0.14950384199619293,
-0.04068421572446823,
0.02855822630226612,
0.05217501148581505,
-0.051676250994205475,
-0.0012693246826529503,
-0.0033691851422190666,
0.009804906323552132,
0.042304627597332,
-0.003913463093340397,
-0.002196984365582466,
0.019733041524887085,
0.05012596398591995,
-0.09492213279008865,
-0.03433378040790558,
0.11569932848215103,
-0.014837343245744705,
-0.0004201573901809752,
0.02007495053112507,
-0.002506787423044443,
-0.08100161701440811,
-0.049359362572431564,
0.014985056594014168,
0.0384652204811573,
0.0755746141076088,
-0.02161579765379429,
0.043135739862918854,
0.04590999335050583,
-0.02515449933707714,
-0.05718155950307846,
-0.012702030129730701,
-0.006419907324016094,
0.0009907747153192759,
-0.08414701372385025,
0.04236738756299019,
0.08479256182909012,
-0.06681952625513077,
-0.03291963040828705,
-0.03577951714396477,
0.04827360436320305,
0.016404757276177406,
-0.038100019097328186,
-0.09070554375648499,
5.835419545736724e-33,
0.02797619253396988,
0.026159942150115967,
0.08195367455482483,
0.014270143583416939,
0.05877838656306267,
0.014925363473594189,
0.040929701179265976,
0.04001481831073761,
0.008810605853796005,
0.013378120958805084,
-0.030678004026412964,
0.03393920138478279,
-0.07433692365884781,
0.0700443908572197,
0.05366576835513115,
0.038074132055044174,
-0.09792247414588928,
0.038029056042432785,
-0.08766666799783707,
-0.028364716097712517,
0.02070367895066738,
-0.04314322769641876,
-0.0381435863673687,
-0.0006591047276742756,
0.0010522229131311178,
-0.005762441549450159,
0.09279875457286835,
0.03014761582016945,
0.06127398833632469,
0.015454565174877644,
-0.06091175973415375,
-0.009140117093920708,
-0.05319001525640488,
0.0690978541970253,
0.02149038389325142,
-0.020200606435537338,
0.0010145752457901835,
-0.025601675733923912,
0.006129268556833267,
0.01696443371474743,
-0.01865413598716259,
0.03442829102277756,
0.052932485938072205,
-0.04900236055254936,
-0.09402397274971008,
-0.09700892120599747,
0.024663424119353294,
0.014341711066663265,
-0.03283064067363739,
0.01126181986182928,
0.05228457972407341,
-0.0327734500169754,
0.048544466495513916,
-0.011791559867560863,
-0.024892713874578476,
0.04056185856461525,
0.012443643994629383,
0.029139773920178413,
0.10912618786096573,
0.08268185704946518,
0.004665694199502468,
-0.038381051272153854,
-0.026819851249456406,
0.0994245782494545,
-0.0005943427095189691,
0.05099625512957573,
-0.045382048934698105,
0.041601039469242096,
0.032108403742313385,
-0.07153401523828506,
-0.0066332099959254265,
0.056017857044935226,
-0.025984100997447968,
-0.03824998065829277,
0.07583816349506378,
0.0068638138473033905,
0.01920008286833763,
-0.008982277475297451,
-0.06725170463323593,
0.03322317823767662,
0.03169618546962738,
0.09704605489969254,
-0.011998980306088924,
-0.05961282178759575,
-0.0828666165471077,
0.013963037170469761,
0.04005531966686249,
-0.07363279163837433,
-0.02375563606619835,
-0.04928762838244438,
-0.03812015801668167,
-0.05111756920814514,
-0.00840638019144535,
-0.016014721244573593,
-0.012971075251698494,
-3.8550252746560453e-33,
-0.014153549447655678,
-0.0033708149567246437,
-0.052661724388599396,
0.04718039184808731,
0.008452456444501877,
0.05260995402932167,
0.03436541184782982,
0.013303396292030811,
-0.01975500024855137,
-0.05007877200841904,
0.030049504712224007,
-0.015270353294909,
0.10360214114189148,
-0.06171366199851036,
0.017107177525758743,
-0.0114181749522686,
-0.07268764823675156,
0.03465399891138077,
0.03209545090794563,
0.03004269488155842,
0.07223369181156158,
-0.008933566510677338,
-0.07152858376502991,
0.03350638598203659,
-0.10418301820755005,
-0.04707573354244232,
-0.06978164613246918,
0.08261504769325256,
-0.025931397452950478,
0.027888545766472816,
-0.0014372339937835932,
0.0070343888364732265,
-0.0710253044962883,
-0.04587429016828537,
-0.08586902916431427,
0.04130677878856659,
0.02286149002611637,
-0.03721456974744797,
-0.09520911425352097,
0.05409930273890495,
0.04301527142524719,
0.03364358842372894,
-0.028563929721713066,
0.04030171409249306,
-0.06703733652830124,
0.00041098997462540865,
0.018987540155649185,
-0.10458813607692719,
0.026457654312253,
0.043238211423158646,
-0.06192581728100777,
-0.005538556724786758,
-0.08029008656740189,
0.004981028847396374,
-0.03740289434790611,
-0.036139924079179764,
-0.007272757589817047,
0.00366971199400723,
0.02309275232255459,
-0.08500123769044876,
-0.07109946757555008,
-0.021898629143834114,
-0.008249608799815178,
-0.007666592951864004,
0.05262106657028198,
-0.0660105049610138,
-0.03233308717608452,
-0.01361663918942213,
0.003196973353624344,
0.026836659759283066,
-0.03680622577667236,
-0.04221407696604729,
0.0422552153468132,
-0.01832803711295128,
0.013456806540489197,
0.04889164865016937,
0.027503661811351776,
-0.03531685844063759,
-0.04000084102153778,
-0.04154713824391365,
0.053479552268981934,
-0.09613268077373505,
-0.01655695214867592,
-0.02023581974208355,
0.05586445704102516,
0.15253818035125732,
0.07347762584686279,
0.05087331682443619,
-0.00184217921923846,
0.05622221529483795,
-0.06203725188970566,
-0.003378705121576786,
-0.03065640479326248,
0.056543078273534775,
-0.04930230230093002,
-3.92354877476464e-8,
-0.09184597432613373,
0.1016659289598465,
-0.03786899521946907,
-0.0050628818571567535,
0.03523360937833786,
-0.048152510076761246,
0.03882015123963356,
0.14937591552734375,
-0.0374179482460022,
-0.01012010220438242,
0.07227905094623566,
0.026241520419716835,
-0.04735492542386055,
-0.07201670855283737,
0.02962610311806202,
0.1278139352798462,
0.019167814403772354,
-0.009372016414999962,
-0.011811907403171062,
0.028268778696656227,
0.10467733442783356,
0.04032803699374199,
0.045990440994501114,
-0.0034743647556751966,
0.08634071797132492,
-0.10077933967113495,
-0.0055106994695961475,
0.12357980757951736,
0.01940673030912876,
-0.09412258118391037,
-0.0816904678940773,
-0.0113186314702034,
-0.019765615463256836,
-0.0742882490158081,
0.0004936348996125162,
0.13080401718616486,
0.020152894780039787,
-0.013811290264129639,
-0.020122645422816277,
-0.04206744208931923,
-0.006678758189082146,
-0.030126996338367462,
-0.028821712359786034,
-0.02090875431895256,
-0.02158975973725319,
0.06353107839822769,
-0.01070533599704504,
-0.027733908966183662,
0.03847601264715195,
-0.04248926788568497,
0.0076989377848804,
-0.00043262200779281557,
-0.05032093822956085,
0.027408774942159653,
0.04026464745402336,
0.054131317883729935,
-0.004122250713407993,
-0.07348545640707016,
0.045062024146318436,
0.06817005574703217,
-0.023812120780348778,
-0.035436104983091354,
-0.0667819008231163,
0.03517743945121765
] | 0 | 0 | 0 | 0 |
|
Policy Design in Long-run Welfare Dynamics
|
https://openreview.net/forum?id=d8hYXbxX71
|
[
"Jiduan Wu",
"Rediet Abebe",
"Moritz Hardt",
"Ana-Andreea Stoica"
] |
Poster
|
Improving social welfare is a complex challenge requiring policymakers to optimize objectives across multiple time horizons. Evaluating the impact of such policies presents a fundamental challenge, as those that appear suboptimal in the short run may yield significant long-term benefits. We tackle this challenge by analyzing the long-term dynamics of two prominent policy frameworks: Rawlsian policies, which prioritize those with the greatest need, and utilitarian policies, which maximize immediate welfare gains. Conventional wisdom suggests these policies are at odds, as Rawlsian policies are assumed to come at the cost of reducing the average social welfare, which their utilitarian counterparts directly optimize. We challenge this assumption by analyzing these policies in a sequential decision-making framework where individuals' welfare levels stochastically decay over time, and policymakers can intervene to prevent this decay. Under reasonable assumptions, we prove that interventions following Rawlsian policies can outperform utilitarian policies in the long run, even when the latter dominate in the short run. We characterize the exact conditions under which Rawlsian policies can outperform utilitarian policies. We further illustrate our theoretical findings using simulations, which highlight the risks of evaluating policies based solely on their short-term effects. Our results underscore the necessity of considering long-term horizons in designing and evaluating welfare policies; the true efficacy of even well-established policies may only emerge over time.
|
Long-run welfare, policy design, Rawlsian policy and utilitarianism
| null | 11,925 |
2503.00632
|
[
-0.1037265956401825,
-0.05405011773109436,
0.05646869167685509,
-0.009672052226960659,
0.04335428774356842,
0.018271958455443382,
-0.04533432796597481,
-0.008656738325953484,
-0.027388058602809906,
0.048452530056238174,
-0.015461878851056099,
0.027622397989034653,
0.005800360348075628,
0.032323602586984634,
0.01981496438384056,
0.009812841191887856,
0.07229039818048477,
0.052578046917915344,
-0.09910459816455841,
0.018273284658789635,
0.021083831787109375,
0.006773792672902346,
-0.03296828642487526,
-0.006472387816756964,
0.0175655335187912,
-0.0892571210861206,
-0.002999871037900448,
-0.040712472051382065,
0.016972361132502556,
0.042110878974199295,
0.09319708496332169,
-0.11558838933706284,
0.03969958797097206,
-0.02609892562031746,
0.008045872673392296,
0.06255999207496643,
-0.07000795751810074,
-0.10388761013746262,
-0.04408659785985947,
0.008692597039043903,
-0.0047402530908584595,
-0.048693083226680756,
-0.09790047258138657,
0.07231941819190979,
-0.015798205509781837,
-0.05432978644967079,
0.1091657504439354,
-0.03427901491522789,
-0.001852374174632132,
-0.036116816103458405,
-0.014732819981873035,
0.05440907925367355,
0.04477082937955856,
0.013773375190794468,
0.03946579992771149,
0.02354101836681366,
0.012057865969836712,
0.009168573655188084,
-0.04152952507138252,
-0.011287805624306202,
-0.09114257991313934,
-0.13361307978630066,
-0.04210777208209038,
-0.050822917371988297,
-0.024620285257697105,
0.00011345718667143956,
-0.012656590901315212,
0.04078001156449318,
0.01121801882982254,
0.15495069324970245,
-0.019554998725652695,
0.018368445336818695,
-0.018573561683297157,
-0.015467977151274681,
0.02698737196624279,
-0.05395670235157013,
0.06347084045410156,
0.03111845627427101,
0.011223279871046543,
-0.11499122530221939,
-0.0037832462694495916,
-0.018593380227684975,
-0.006167080253362656,
-0.02760765142738819,
0.07650209963321686,
-0.017831383273005486,
-0.027423402294516563,
-0.028983578085899353,
0.08557506650686264,
-0.02436661906540394,
0.01742946356534958,
0.024136381223797798,
0.012679515406489372,
-0.03438572585582733,
0.07522273063659668,
0.03301361948251724,
0.04485515132546425,
-0.07263343781232834,
-0.06718581914901733,
0.08649507910013199,
-0.08077523857355118,
-0.020911525934934616,
0.001464742235839367,
0.026835909113287926,
0.03275156766176224,
-0.16167742013931274,
-0.08779175579547882,
0.08486078679561615,
0.007810400333255529,
-0.009487302973866463,
0.10442660748958588,
-0.03201358765363693,
0.04478726536035538,
0.03409383073449135,
-0.0470876507461071,
-0.05064292252063751,
0.024660887196660042,
-0.03291163593530655,
0.059791598469018936,
0.033028069883584976,
0.045515026897192,
0.02397223189473152,
0.023804180324077606,
0.025423789396882057,
-0.012300953269004822,
0.07298129051923752,
0.010647578164935112,
1.66463646831165e-33,
-0.10505395382642746,
-0.017712993547320366,
-0.024639397859573364,
-0.006383844185620546,
-0.07091087847948074,
-0.06394650787115097,
-0.009129862301051617,
0.03799155727028847,
-0.07830500602722168,
0.012647805735468864,
0.06742990016937256,
0.04772728681564331,
-0.03680015355348587,
0.11821671575307846,
0.04301074892282486,
-0.08732490986585617,
-0.06422636657953262,
0.07118361443281174,
0.01453715655952692,
-0.03114514984190464,
0.04751204699277878,
-0.014159579761326313,
-0.001405362389050424,
-0.07103562355041504,
0.006207927595824003,
-0.06739435344934464,
0.0057338932529091835,
-0.018795328214764595,
0.0049816919490695,
0.014466260559856892,
-0.02998516336083412,
0.07946176081895828,
-0.01850857213139534,
-0.02098432183265686,
-0.06383408606052399,
-0.00564385624602437,
-0.02371695451438427,
-0.004428292158991098,
0.007711900398135185,
-0.0460614413022995,
-0.04049299284815788,
0.02044859528541565,
0.04040075093507767,
0.06924082338809967,
-0.01768578588962555,
-0.12259963154792786,
0.031031964346766472,
0.03116212785243988,
-0.06123487278819084,
0.03243013843894005,
0.034318141639232635,
0.04526178911328316,
-0.013125563971698284,
-0.07738439738750458,
-0.13841646909713745,
-0.04204340651631355,
0.02371903881430626,
-0.03859364613890648,
-0.0828639417886734,
-0.04646555706858635,
0.061775777488946915,
-0.052097320556640625,
0.005675380118191242,
-0.03939511626958847,
-0.033688098192214966,
0.11350467056035995,
-0.030555617064237595,
0.06347962468862534,
0.03653969243168831,
-0.009022217243909836,
0.030965708196163177,
0.08583978563547134,
0.019280049949884415,
-0.06294257938861847,
0.014429558999836445,
0.073617123067379,
0.06038646772503853,
-0.00534052262082696,
-0.026025010272860527,
-0.026970425620675087,
0.01862509921193123,
0.0002597621059976518,
-0.033827755600214005,
-0.014379460364580154,
0.055363476276397705,
0.026475094258785248,
0.006959747057408094,
-0.05043443664908409,
0.027306368574500084,
-0.05303783714771271,
-0.06017086282372475,
-0.00530619639903307,
0.009740740060806274,
0.04572940617799759,
0.02040354162454605,
-4.06016850799972e-33,
-0.01880139857530594,
-0.01517172995954752,
0.004384081345051527,
0.031693778932094574,
0.054457951337099075,
-0.028833044692873955,
-0.0844886302947998,
-0.05851267650723457,
-0.036904219537973404,
-0.012575695291161537,
-0.04163945093750954,
0.010768341831862926,
0.04834901914000511,
0.00016428460367023945,
-0.04298270121216774,
-0.0382063128054142,
0.043655913323163986,
-0.005429629702121019,
0.03620319440960884,
-0.043582066893577576,
0.07522310316562653,
0.09987916052341461,
-0.10455247014760971,
0.10856330394744873,
0.01899564079940319,
0.056662291288375854,
-0.03742779791355133,
0.026732010766863823,
0.014860098250210285,
-0.03648102656006813,
-0.10934435576200485,
-0.052579768002033234,
-0.032938092947006226,
-0.0367254838347435,
-0.02407597191631794,
0.0811699777841568,
-0.012005932629108429,
0.013297968544065952,
-0.05596258118748665,
0.09015974402427673,
0.114990234375,
-0.08420975506305695,
-0.02495681680738926,
0.04773172736167908,
0.045894816517829895,
-0.04392172768712044,
-0.0024027808103710413,
-0.00599232641980052,
-0.03867543861269951,
0.015591204166412354,
0.060495853424072266,
0.03809276223182678,
0.010696270503103733,
0.07553785294294357,
-0.10318685322999954,
-0.09683335572481155,
0.001096538151614368,
-0.0476166307926178,
0.0421285554766655,
0.026926299557089806,
-0.04472099617123604,
0.06626482307910919,
-0.030314572155475616,
0.053740255534648895,
-0.01600799523293972,
-0.043803922832012177,
-0.003626365913078189,
-0.10109324008226395,
0.07308834791183472,
-0.05863751098513603,
0.040350791066884995,
0.002394637558609247,
-0.08805960416793823,
-0.04326917603611946,
0.005572803318500519,
0.08841021358966827,
-0.012538476847112179,
0.03713042661547661,
-0.04863458499312401,
0.04000892862677574,
0.027481164783239365,
-0.00020975092775188386,
0.01289280317723751,
-0.03796283155679703,
-0.025587832555174828,
-0.0026675094850361347,
0.026301603764295578,
0.02019359916448593,
-0.021665945649147034,
0.03537869453430176,
-0.0062369247898459435,
-0.03047238662838936,
0.03586982563138008,
0.03561605513095856,
-0.024560198187828064,
-4.717173140988962e-8,
0.08119922876358032,
0.06007158011198044,
-0.005603224039077759,
0.0119935879483819,
0.04713164269924164,
-0.004706383682787418,
-0.0655415952205658,
-0.0035197800025343895,
-0.03244123235344887,
-0.042406193912029266,
0.053936850279569626,
0.10396703332662582,
0.07784264534711838,
-0.08756612986326218,
-0.07536941766738892,
0.07291655987501144,
-0.04374302178621292,
-0.05022620037198067,
0.01134552527219057,
0.037053342908620834,
0.01732511818408966,
0.08378936350345612,
-0.014691151678562164,
-0.05529220402240753,
-0.030479492619633675,
0.06572704762220383,
0.014850255101919174,
0.08417344838380814,
0.07829316705465317,
0.023309575393795967,
-0.006623928900808096,
0.024597885087132454,
-0.018447251990437508,
0.028040515258908272,
-0.040012650191783905,
-0.04163709655404091,
-0.07089380919933319,
-0.005089121870696545,
0.03984453156590462,
-0.027219410985708237,
0.018153274431824684,
0.052353665232658386,
-0.08389150351285934,
0.030079858377575874,
-0.020747721195220947,
-0.004047630820423365,
-0.11170949786901474,
-0.059217214584350586,
0.04816724359989166,
-0.01883339323103428,
0.03411790728569031,
-0.022446854040026665,
0.022777186706662178,
0.05496944114565849,
0.026470419019460678,
-0.052865639328956604,
0.02065091021358967,
-0.0029865908436477184,
-0.03094150312244892,
0.08084896951913834,
0.027833905071020126,
0.01905929110944271,
-0.028102897107601166,
0.04545711725950241
] |
https://github.com/wujiduan/Rawls_vs_Utils
| 0 | 0 | 0 | 0 |
Dynamic Modeling of Patients, Modalities and Tasks via Multi-modal Multi-task Mixture of Experts
|
https://openreview.net/forum?id=NJxCpMt0sf
|
[
"Chenwei Wu",
"Zitao Shuai",
"Zhengxu Tang",
"Luning Wang",
"Liyue Shen"
] |
Poster
|
Multi-modal multi-task learning holds significant promise in tackling complex diagnostic tasks and many significant medical imaging problems. It fulfills the needs in real-world diagnosis protocol to leverage information from different data sources and simultaneously perform mutually informative tasks. However, medical imaging domains introduce two key challenges: dynamic modality fusion and modality-task dependence. The quality and amount of task-related information from different modalities could vary significantly across patient samples, due to biological and demographic factors. Traditional fusion methods apply fixed combination strategies that fail to capture this dynamic relationship, potentially underutilizing modalities that carry stronger diagnostic signals for specific patients. Additionally, different clinical tasks may require dynamic feature selection and combination from various modalities, a phenomenon we term “modality-task dependence.” To address these issues, we propose M4oE, a novel Multi-modal Multi-task Mixture of Experts framework for precise Medical diagnosis. M4oE comprises Modality-Specific (MSoE) modules and a Modality-shared Modality-Task MoE (MToE) module. With collaboration from both modules, our model dynamically decomposes and learns distinct and shared information from different modalities and achieves dynamic fusion. MToE provides a joint probability model of modalities and tasks by using experts as a link and encourages experts to learn modality-task dependence via conditional mutual information loss. By doing so, M4oE offers sample and population-level interpretability of modality contributions. We evaluate M4oE on four public multi-modal medical benchmark datasets for solving two important medical diagnostic problems including breast cancer screening and retinal disease diagnosis. Results demonstrate our method's superiority over state-of-the-art methods under different metrics of classification and segmentation tasks like Accuracy, AUROC, AUPRC, and DICE.
|
Multimodal Learning, Medical Imaging
|
A Sample and Task-dynamic Model for Multi-modal Multi-task Medical Image Learning
| 11,923 | null |
[
0.02059340476989746,
-0.07042152434587479,
0.013603274710476398,
0.009716233238577843,
0.07274378091096878,
-0.05952032655477524,
0.05388383939862251,
0.015815600752830505,
-0.019272295758128166,
-0.10470446944236755,
-0.040085695683956146,
-0.10547903925180435,
0.03346410393714905,
0.12029432505369186,
-0.0326029434800148,
0.020494546741247177,
-0.0087966900318861,
0.06560934334993362,
-0.1188177615404129,
-0.0004662483697757125,
0.034827638417482376,
-0.03118770942091942,
0.004787759855389595,
0.010369103401899338,
0.009312417358160019,
0.05847010388970375,
-0.004584120120853186,
-0.062324535101652145,
-0.01633027009665966,
-0.0321720652282238,
0.12015242129564285,
0.04174545034766197,
-0.04401981458067894,
0.006655914708971977,
-0.05052661523222923,
0.006136871874332428,
-0.05982097238302231,
0.07258626073598862,
-0.057276733219623566,
-0.04656650125980377,
-0.0009779698448255658,
0.038945674896240234,
-0.02779505029320717,
0.02022562175989151,
-0.004912913776934147,
0.021289026364684105,
-0.06990250200033188,
-0.02850859798491001,
0.09074798971414566,
0.05552641302347183,
-0.0693659782409668,
-0.039143096655607224,
-0.01356944814324379,
0.07152608782052994,
0.05564286559820175,
0.06934285163879395,
-0.05201686546206474,
-0.03842918202280998,
-0.03318844735622406,
-0.0251393411308527,
-0.045267097651958466,
-0.05905493348836899,
0.012219779193401337,
-0.003713050624355674,
0.12610746920108795,
0.07258595526218414,
-0.024501938372850418,
-0.02722257375717163,
-0.026264837011694908,
0.003766811452805996,
-0.08625831454992294,
0.032893601804971695,
-0.001570197637192905,
-0.026027046144008636,
-0.02278992161154747,
0.017078762874007225,
0.036531541496515274,
0.0328030027449131,
0.05999241769313812,
-0.009942645207047462,
-0.006072507239878178,
0.07462336122989655,
0.0063074869103729725,
-0.09135409444570541,
0.15714967250823975,
-0.010451962240040302,
-0.059428103268146515,
0.024314546957612038,
-0.01700645685195923,
-0.023949475958943367,
-0.04851992800831795,
-0.06799505650997162,
0.04724178835749626,
-0.005479687359184027,
0.05205657705664635,
-0.0034975807648152113,
0.03219301626086235,
-0.024585800245404243,
0.05809265002608299,
0.05805862322449684,
-0.09017427265644073,
-0.054606981575489044,
-0.017546283081173897,
0.03082037717103958,
0.02662079967558384,
-0.07026136666536331,
0.003756200196221471,
0.002233927370980382,
0.042815398424863815,
-0.058286186307668686,
0.024872897192835808,
0.009695055894553661,
0.06257780641317368,
-0.02801261842250824,
0.018484003841876984,
0.004575344733893871,
-0.0009112244006246328,
0.05560576170682907,
0.05096610635519028,
0.04275142773985863,
-0.0019649772439152002,
-0.07878521829843521,
-0.01746623031795025,
-0.047897275537252426,
0.09013577550649643,
0.05190744623541832,
-0.09859772771596909,
2.2755280254645306e-33,
0.07595587521791458,
0.014440631493926048,
0.0781155377626419,
0.07313025742769241,
0.029840897768735886,
-0.026428481563925743,
-0.07230161130428314,
0.050431668758392334,
-0.023498233407735825,
0.021114448085427284,
-0.02483825758099556,
0.054270338267087936,
-0.04438801854848862,
0.06907683610916138,
0.031096076592803,
-0.019632162526249886,
0.007444646675139666,
0.04615785926580429,
-0.013317546807229519,
0.000020609149942174554,
0.029645048081874847,
-0.00778921926394105,
-0.044192079454660416,
0.02283407747745514,
-0.020706122741103172,
0.002697476651519537,
0.0893794596195221,
0.00023460945521946996,
0.004249103832989931,
0.009375742636620998,
-0.1947997361421585,
-0.012122094631195068,
-0.046719156205654144,
0.005417537875473499,
-0.007977567613124847,
-0.025979656726121902,
-0.08337943255901337,
-0.03154372051358223,
0.01400080043822527,
0.02841089852154255,
-0.013107377104461193,
-0.0518498457968235,
-0.019171830266714096,
-0.010686944238841534,
-0.025660371407866478,
-0.05943497642874718,
-0.008869433775544167,
0.032385870814323425,
-0.06565061956644058,
-0.05987244099378586,
-0.01967649720609188,
-0.03164411708712578,
0.053089823573827744,
-0.11718627065420151,
-0.039625588804483414,
0.04824000597000122,
-0.04704572632908821,
0.020201293751597404,
0.09084955602884293,
-0.004746459424495697,
0.004532345570623875,
0.00823406595736742,
-0.057041238993406296,
0.053550273180007935,
0.05493277683854103,
0.05988067388534546,
-0.004187589045614004,
0.043437641113996506,
0.08334552496671677,
0.027177682146430016,
-0.007903428748250008,
0.05783888325095177,
0.030820928514003754,
-0.03594212979078293,
0.05208106338977814,
0.04628658667206764,
0.0587153285741806,
-0.06717634201049805,
-0.03026810847222805,
0.024625178426504135,
-0.04265123978257179,
0.05569285899400711,
-0.042357202619314194,
-0.02703372947871685,
-0.008710920810699463,
0.025926493108272552,
0.023018207401037216,
-0.02260882407426834,
-0.11044950783252716,
-0.05870406702160835,
-0.05140214413404465,
0.05812352895736694,
-0.012269379571080208,
-0.020758170634508133,
-0.0016592098399996758,
-1.7307499440217417e-33,
0.011539307422935963,
0.06405185908079147,
-0.056370723992586136,
-0.020867545157670975,
0.08527863770723343,
-0.023190954700112343,
0.017744384706020355,
-0.028851879760622978,
-0.04376155510544777,
-0.03911294788122177,
0.027315231040120125,
-0.09237895905971527,
0.0067018079571425915,
-0.07696180045604706,
-0.03190106526017189,
0.10729508101940155,
-0.13089284300804138,
0.06514137238264084,
0.013295372016727924,
0.04941151663661003,
0.011401708237826824,
0.04948152229189873,
-0.07474011927843094,
0.054905038326978683,
-0.06758873164653778,
0.03140843287110329,
-0.005277945660054684,
0.005041187163442373,
0.014973263256251812,
-0.033452920615673065,
-0.05624763295054436,
-0.033408187329769135,
0.0013167093275114894,
-0.04011467844247818,
-0.0029738303273916245,
0.08502835035324097,
-0.03617719188332558,
-0.04271385818719864,
-0.0020833092276006937,
0.0718299075961113,
0.017750568687915802,
0.0010844948701560497,
-0.0864245817065239,
0.03845876082777977,
0.04396231472492218,
-0.041304223239421844,
0.0026351269334554672,
-0.01308513805270195,
0.09331623464822769,
0.023424534127116203,
-0.026319097727537155,
-0.029940608888864517,
-0.13865575194358826,
0.019960103556513786,
-0.07383007556200027,
-0.06013384833931923,
-0.0004996781353838742,
-0.06805998831987381,
0.011560185812413692,
0.021007267758250237,
-0.009847382083535194,
-0.06546390056610107,
0.060549236834049225,
0.019603244960308075,
0.034638408571481705,
0.0404970720410347,
0.07168953120708466,
-0.0371219776570797,
-0.045508380979299545,
0.085419662296772,
-0.04228683561086655,
-0.020827196538448334,
0.06291799992322922,
0.0034537871833890676,
0.11450508236885071,
-0.015316437929868698,
-0.035216279327869415,
-0.03769949451088905,
-0.049370232969522476,
0.009690367616713047,
0.013651391491293907,
-0.11472631245851517,
-0.00043934767018072307,
0.051715780049562454,
-0.03179140388965607,
0.060869842767715454,
0.06154982000589371,
0.04724433273077011,
0.07034564018249512,
-0.042083993554115295,
0.042303409427404404,
0.0339834950864315,
-0.005764562636613846,
0.013466483913362026,
-0.02970241941511631,
-3.979893037353577e-8,
0.014702052809298038,
-0.005020929966121912,
-0.021580588072538376,
-0.08256129920482635,
0.005344795063138008,
-0.06717009097337723,
-0.04067365825176239,
0.009562279097735882,
0.02894456498324871,
0.0046489122323691845,
0.059658363461494446,
-0.1048947125673294,
-0.011862828396260738,
-0.06141577661037445,
0.04976293444633484,
0.03982105478644371,
-0.0034827759955078363,
0.08242516964673996,
-0.0413324274122715,
-0.05141366273164749,
0.08654194325208664,
-0.05311945825815201,
0.05681786313652992,
0.01004860084503889,
0.09702141582965851,
-0.052050020545721054,
-0.057639747858047485,
0.055541954934597015,
0.01270871702581644,
0.035223402082920074,
-0.020708130672574043,
0.017302345484495163,
0.025856664404273033,
0.013045604340732098,
-0.058527663350105286,
0.03704399615526199,
0.08816102892160416,
-0.04688040539622307,
-0.03991304710507393,
0.03679397702217102,
0.0035662923473864794,
0.03339679539203644,
-0.06088840216398239,
0.07132028043270111,
0.0415848083794117,
-0.016629669815301895,
0.05070365220308304,
-0.054378967732191086,
0.032541558146476746,
0.015151886269450188,
0.028537094593048096,
0.030213970690965652,
0.06624013185501099,
0.11446760594844818,
0.0787358507514,
0.0489020012319088,
0.06808701902627945,
0.010189570486545563,
0.08635135740041733,
-0.02681766450405121,
0.059714529663324356,
0.010806539095938206,
-0.05254774168133736,
-0.055134113878011703
] | 0 | 0 | 0 | 0 |
|
Procedural Synthesis of Synthesizable Molecules
|
https://openreview.net/forum?id=OGfyzExd69
|
[
"Michael Sun",
"Alston Lo",
"Minghao Guo",
"Jie Chen",
"Connor W. Coley",
"Wojciech Matusik"
] |
Poster
|
Designing synthetically accessible molecules and recommending analogs to unsynthesizable molecules are important problems for accelerating molecular discovery. We reconceptualize both problems using ideas from program synthesis. Drawing inspiration from syntax-guided synthesis approaches, we decouple the syntactic skeleton from the semantics of a synthetic tree to create a bilevel framework for reasoning about the combinatorial space of synthesis pathways. Given a molecule we aim to generate analogs for, we iteratively refine its skeletal characteristics via Markov Chain Monte Carlo simulations over the space of syntactic skeletons. Given a black-box oracle to optimize, we formulate a joint design space over syntactic templates and molecular descriptors and introduce evolutionary algorithms that optimize both syntactic and semantic dimensions synergistically. Our key insight is that once the syntactic skeleton is set, we can amortize over the search complexity of deriving the program's semantics by training policies to fully utilize the fixed horizon Markov Decision Process imposed by the syntactic template. We demonstrate performance advantages of our bilevel framework for synthesizable analog generation and synthesizable molecule design. Notably, our approach offers the user explicit control over the resources required to perform synthesis and biases the design space towards simpler solutions, making it particularly promising for autonomous synthesis platforms. Supporting code is at https://github.com/shiningsunnyday/SynthesisNet.
|
molecular design, synthesis planning, tree generation, graph generation
|
We tackle synthesizable analog generation and synthesizable molecule design by modeling the design of synthesis pathways as a conditional program synthesis problem.
| 11,920 |
2409.05873
|
[
-0.05186205729842186,
-0.060985732823610306,
0.014870786108076572,
-0.02912253700196743,
-0.018559863790869713,
-0.0814858227968216,
-0.041590042412281036,
0.007045794744044542,
-0.014676465652883053,
-0.014400138519704342,
-0.06735879182815552,
-0.08466806262731552,
0.028606446459889412,
0.0016055997693911195,
0.0016311872750520706,
0.03926762193441391,
-0.012743531726300716,
0.11420129984617233,
-0.011530670337378979,
-0.030536899343132973,
0.029129354283213615,
0.07287590205669403,
0.031336698681116104,
-0.013711882755160332,
0.0011824413668364286,
0.06723956018686295,
0.031025966629385948,
-0.031656038016080856,
-0.020612437278032303,
0.028382837772369385,
0.07292699813842773,
0.01802673190832138,
-0.010892347432672977,
-0.08795954287052155,
0.0710512027144432,
0.0587090440094471,
-0.07930821180343628,
-0.03261473402380943,
-0.0048705716617405415,
-0.017940787598490715,
-0.013004885986447334,
0.03630071505904198,
-0.021063147112727165,
0.025444786995649338,
0.047156788408756256,
-0.048566415905952454,
-0.02415694296360016,
0.018178969621658325,
-0.02527882158756256,
-0.022650040686130524,
-0.06679472327232361,
-0.044866953045129776,
-0.07036047428846359,
-0.03504731133580208,
0.0334102138876915,
0.09476137906312943,
-0.016049206256866455,
-0.03924161195755005,
-0.018623298034071922,
-0.05397307500243187,
0.004029015079140663,
-0.07716111838817596,
0.0023784420918673277,
-0.057828981429338455,
0.007120370864868164,
-0.0056890989653766155,
0.10947176814079285,
0.0605708584189415,
0.08098018914461136,
0.08460092544555664,
-0.10205356031656265,
0.0934276133775711,
-0.03137115016579628,
0.08726145327091217,
0.0023281248286366463,
0.09562113881111145,
0.023687593638896942,
0.00482584536075592,
0.05094965547323227,
-0.10475947707891464,
-0.018928416073322296,
0.007479085121303797,
0.028702156618237495,
0.010868677869439125,
0.03476016968488693,
0.02594987489283085,
0.04116367921233177,
0.10250730812549591,
0.09958535432815552,
0.010943655855953693,
-0.06761320680379868,
-0.07135520875453949,
0.05217854306101799,
-0.09259495884180069,
-0.060153063386678696,
0.06010245531797409,
0.031068921089172363,
-0.01954810693860054,
0.06863249838352203,
0.041697654873132706,
0.011566033586859703,
0.03223387524485588,
0.08792858570814133,
-0.08740142732858658,
-0.0007049718988128006,
-0.002126471372321248,
0.007935353554785252,
0.05006876960396767,
0.0028542557265609503,
-0.039304908365011215,
-0.01089070551097393,
0.06530574709177017,
0.056535277515649796,
-0.0310304444283247,
-0.047429971396923065,
0.07618742436170578,
-0.027767905965447426,
-0.057596586644649506,
0.0763339251279831,
0.1031745970249176,
-0.017539924010634422,
0.004067536443471909,
-0.04056194797158241,
0.047620102763175964,
-0.03410571441054344,
-0.039333947002887726,
-0.08934208750724792,
1.5511902431977602e-33,
0.0505121573805809,
0.063304103910923,
0.12654268741607666,
0.06895317137241364,
0.09673909842967987,
-0.033514875918626785,
0.039463479071855545,
-0.005125876050442457,
-0.09636485576629639,
0.015372367575764656,
0.0027378713712096214,
-0.0522996224462986,
-0.0852171927690506,
0.05619657784700394,
0.08867957442998886,
-0.005807456094771624,
-0.032271839678287506,
0.08377300202846527,
-0.07887794822454453,
-0.04221589118242264,
-0.013401715084910393,
0.040625158697366714,
-0.03515526279807091,
-0.031375180929899216,
-0.003611012827605009,
0.017167838290333748,
-0.011261885985732079,
-0.041770320385694504,
-0.03332839161157608,
-0.020685294643044472,
-0.09973087906837463,
-0.041858185082674026,
-0.01062371302396059,
0.05816385895013809,
0.03530802205204964,
-0.09659921377897263,
-0.015573094598948956,
-0.1030377447605133,
-0.011230059899389744,
-0.022134695202112198,
0.03568064793944359,
-0.011652754619717598,
0.03395410627126694,
0.040697839111089706,
-0.04808123782277107,
0.051865339279174805,
-0.015165433287620544,
-0.0027712436858564615,
-0.009982976131141186,
-0.057498857378959656,
0.08129284530878067,
0.0462430939078331,
0.09056555479764938,
-0.006987620145082474,
-0.0313401073217392,
0.0007338174618780613,
-0.03442830964922905,
-0.0025389529764652252,
0.06116224825382233,
0.07182793319225311,
-0.027460195124149323,
0.06284084916114807,
-0.01693716086447239,
0.01633474789559841,
-0.02876359410583973,
0.05976776406168938,
-0.0370587483048439,
0.007179757114499807,
0.08278528600931168,
0.04354926571249962,
-0.019746657460927963,
-0.018982352688908577,
-0.015251007862389088,
-0.007744234520941973,
0.04322872683405876,
0.030742008239030838,
-0.003378475783392787,
-0.07067167013883591,
-0.024547945708036423,
0.008026989176869392,
-0.037627924233675,
0.04873333498835564,
-0.07078889012336731,
0.0008214110275730491,
0.013670899905264378,
-0.1783004105091095,
-0.035267338156700134,
-0.06847866624593735,
-0.07683233916759491,
-0.007084534503519535,
-0.08595693111419678,
-0.10947148501873016,
-0.034110911190509796,
-0.054786451160907745,
0.02544931136071682,
-4.19314777580443e-33,
-0.01860884763300419,
-0.042415764182806015,
0.013567887246608734,
0.023768190294504166,
-0.0749422088265419,
-0.08293706178665161,
-0.07704965770244598,
-0.06172214820981026,
0.05255907028913498,
-0.09237765520811081,
-0.04563838243484497,
-0.012319663539528847,
0.04477599635720253,
-0.06668800115585327,
0.014787664636969566,
-0.02043139562010765,
0.00519073149189353,
-0.0015825949376448989,
0.03401392325758934,
0.07961469143629074,
0.0430084727704525,
0.08914492279291153,
-0.0675235316157341,
0.006867211777716875,
-0.010968316346406937,
0.0034465633798390627,
0.014129922725260258,
0.03818758577108383,
0.03014315664768219,
0.035569705069065094,
-0.06367477774620056,
0.015487785451114178,
-0.07297256588935852,
0.0730038583278656,
-0.06002030894160271,
0.025282423943281174,
0.025554301217198372,
-0.011995578184723854,
-0.009068025276064873,
-0.004792812746018171,
0.00664279330521822,
-0.052332889288663864,
-0.04695955663919449,
0.04436134546995163,
0.04744512587785721,
0.007822353392839432,
-0.055221352726221085,
0.024968793615698814,
0.04467599093914032,
-0.05107937008142471,
0.04718142747879028,
0.02471490204334259,
0.0038950748275965452,
-0.04365057125687599,
-0.0440731905400753,
-0.03954152390360832,
-0.06441538035869598,
-0.060660772025585175,
0.13129253685474396,
-0.018545690923929214,
-0.07856907695531845,
-0.07864654809236526,
0.035522930324077606,
-0.047123078256845474,
-0.046737488359212875,
-0.07809607684612274,
-0.07349727302789688,
-0.03879403695464134,
0.02851429395377636,
-0.05960319563746452,
0.007004165556281805,
0.020127931609749794,
0.013248222880065441,
0.10794825106859207,
0.049216270446777344,
-0.02603350766003132,
0.05186145380139351,
-0.08138734102249146,
-0.01867159828543663,
-0.04795842617750168,
-0.016388267278671265,
0.025461861863732338,
-0.0023553534410893917,
0.023823216557502747,
0.047127559781074524,
0.049040913581848145,
0.042183492332696915,
0.0682058185338974,
0.0018889218335971236,
-0.0032463364768773317,
0.005753358360379934,
0.04387420043349266,
0.016766225919127464,
0.13429732620716095,
-0.051302775740623474,
-3.898203004837342e-8,
0.07078120857477188,
0.057492729276418686,
0.04059900715947151,
-0.008854390121996403,
0.04304536059498787,
0.05274120718240738,
-0.015602332539856434,
0.0547001026570797,
0.0037865606136620045,
-0.049340471625328064,
0.05139806866645813,
0.0361999049782753,
0.020552001893520355,
-0.00547025166451931,
-0.021100828424096107,
0.09340441226959229,
-0.07088726758956909,
-0.007509652059525251,
-0.05740268900990486,
0.01767805591225624,
-0.0021034441888332367,
0.09108604490756989,
-0.06069154292345047,
0.05939583480358124,
0.05988488718867302,
-0.10459014773368835,
0.0191821176558733,
-0.018760334700345993,
0.060949575155973434,
0.05202431231737137,
0.011842552572488785,
0.03161753714084625,
-0.0041872127912938595,
0.07576774060726166,
0.027438007295131683,
-0.011686420999467373,
-0.02311909757554531,
-0.027462443336844444,
-0.02999367006123066,
-0.08494780212640762,
0.021957622841000557,
0.05965263769030571,
-0.04904082044959068,
0.023739038035273552,
-0.00007235789234982803,
-0.034824296832084656,
-0.04025821387767792,
0.03592389449477196,
0.02051715925335884,
0.016830673441290855,
0.0026594160590320826,
-0.005483238026499748,
-0.03936001658439636,
-0.0903199166059494,
-0.03294875845313072,
0.04509010538458824,
-0.05205126851797104,
-0.006232650019228458,
0.01239039096981287,
-0.017561174929142,
-0.007858802564442158,
-0.012671689502894878,
0.05595686659216881,
-0.03938350826501846
] |
https://github.com/shiningsunnyday/synthesisnet
| 5 | 0 | 0 | 0 |
Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning
|
https://openreview.net/forum?id=44z7HL4mfX
|
[
"Simran Kaur",
"Simon Park",
"Anirudh Goyal",
"Sanjeev Arora"
] |
Poster
|
We introduce INSTRUCT-SKILLMIX, an automated approach for creating diverse, high quality SFT data for instruction-following. The pipeline involves two stages, each leveraging an existing powerful LLM: (1) Skill extraction: uses the LLM to extract core “skills” for instruction-following by directly prompting the model. This is inspired by “LLM metacognition” of (Didolkar et al., 2024); (2) Data generation: uses the powerful LLM to generate (instruction, response) data that
exhibit a randomly chosen pair of these skills. Here, the use of random skill combinations promotes diversity and difficulty. The estimated cost of creating the dataset is under $600.
Vanilla SFT (i.e., no PPO, DPO, or RL methods) on data generated from INSTRUCT-SKILLMIX leads to strong gains on instruction following benchmarks such as AlpacaEval 2.0, MT-Bench, and WildBench. With just 4K examples, LLaMA-3-8B-Base achieves 42.76% length-controlled win rate on AlpacaEval 2.0, a level similar to frontier models like Claude 3 Opus and LLaMA-3.1-405B-Instruct. Ablation studies also suggest plausible reasons for why creating open instruction-tuning datasets via naive crowd-sourcing has proved difficult. In our dataset,adding 20% low quality answers (“shirkers”) causes a noticeable degradation in performance.
The INSTRUCT-SKILLMIX pipeline seems flexible and adaptable to other settings.
|
instruction tuning, high quality synthetic data, diverse synthetic data
|
We introduce an automated approach for creating diverse, high quality SFT data for instruction-following.
| 11,917 | null |
[
-0.026374461129307747,
-0.0578056238591671,
-0.013165269047021866,
0.0030937481205910444,
-0.026345733553171158,
-0.0681244432926178,
-0.05729199945926666,
0.021413031965494156,
-0.04374350234866142,
-0.02297273837029934,
-0.08820150047540665,
-0.018274478614330292,
0.00013819491141475737,
-0.03859908878803253,
-0.011481810361146927,
0.01870918646454811,
0.031359393149614334,
0.05840279534459114,
-0.09491405636072159,
-0.06396552920341492,
0.05531521141529083,
0.012492679059505463,
0.050460003316402435,
-0.040381211787462234,
0.027508269995450974,
-0.002435117494314909,
-0.006898520048707724,
-0.005215918179601431,
0.051634155213832855,
-0.060843441635370255,
0.024009738117456436,
0.047701146453619,
-0.005221379920840263,
0.016173595562577248,
-0.028953755274415016,
0.013776411302387714,
-0.02974589914083481,
0.012845193967223167,
-0.03038662299513817,
-0.03318221494555473,
-0.04489796236157417,
0.039487797766923904,
0.046965379267930984,
0.026798658072948456,
0.06892218440771103,
-0.01931704580783844,
-0.02309996634721756,
-0.06240082159638405,
-0.03467509150505066,
0.08458954095840454,
-0.10465477406978607,
-0.0742793008685112,
-0.0074049183167517185,
0.01859756000339985,
-0.014682142063975334,
0.05126776918768883,
0.05664579197764397,
-0.01118515245616436,
-0.027489516884088516,
-0.00452979514375329,
-0.08383690565824509,
-0.11826544255018234,
-0.10085101425647736,
-0.014374291524291039,
0.008016795851290226,
0.002632658928632736,
0.01696111261844635,
0.03039618954062462,
0.09044408798217773,
-0.06021123379468918,
-0.05186356604099274,
0.03351525589823723,
-0.05915367230772972,
0.09309481084346771,
0.01061982661485672,
0.09760170429944992,
0.06269662082195282,
-0.0446406751871109,
0.04429560527205467,
-0.09712881594896317,
0.01007833145558834,
0.017779948189854622,
-0.023545658215880394,
-0.0027779587544500828,
0.03503766655921936,
-0.02773875929415226,
0.046883098781108856,
0.08685518801212311,
0.082618348300457,
0.034705668687820435,
0.0015770747559145093,
-0.008358649909496307,
-0.031148679554462433,
-0.028233293443918228,
-0.014052419923245907,
0.05738425999879837,
0.04457014426589012,
-0.08559393137693405,
-0.05294619873166084,
0.03766397014260292,
0.015270651318132877,
0.024318773299455643,
0.04022694379091263,
-0.031274620443582535,
-0.05096137151122093,
-0.026012873277068138,
0.034418392926454544,
0.0368715338408947,
-0.061562858521938324,
-0.09208481013774872,
0.0029326598159968853,
0.08330168575048447,
-0.021900104358792305,
0.018808959051966667,
0.02201140485703945,
0.024462057277560234,
-0.048209451138973236,
-0.055164821445941925,
-0.019854381680488586,
0.0752095952630043,
-0.05698298662900925,
-0.012266718782484531,
-0.023819362744688988,
0.042250022292137146,
0.030147159472107887,
-0.05260153114795685,
-0.10995593667030334,
5.5347743758864466e-33,
0.06635233014822006,
0.05779256671667099,
-0.03723689913749695,
0.00705415103584528,
0.08005183935165405,
-0.0638890489935875,
0.09322489798069,
0.05905580520629883,
-0.08847654610872269,
0.040340036153793335,
-0.04297976195812225,
-0.028453661128878593,
-0.027997776865959167,
0.06791047006845474,
0.1112309917807579,
-0.017251061275601387,
-0.0162536408752203,
0.01888708584010601,
-0.045155856758356094,
0.03850741684436798,
0.08845614641904831,
0.06344490498304367,
0.02903183549642563,
-0.07233629375696182,
0.017272595316171646,
0.014165899716317654,
0.04836330935359001,
-0.0008193600224331021,
-0.06735683232545853,
0.04089706763625145,
-0.0728265643119812,
-0.04372019320726395,
-0.012179839424788952,
0.05262770876288414,
0.08439438790082932,
-0.007610047236084938,
-0.007528357673436403,
-0.06609481573104858,
0.07492592930793762,
0.030100420117378235,
0.03461650013923645,
0.03580700233578682,
0.047483157366514206,
-0.024703629314899445,
-0.06299277395009995,
-0.03647970035672188,
0.014430760405957699,
0.020705707371234894,
-0.0008598494459874928,
0.042182113975286484,
0.0370413213968277,
0.04926486685872078,
-0.008836857974529266,
-0.04604298248887062,
0.00374775193631649,
0.02117515541613102,
0.015660898759961128,
0.06570711731910706,
0.039676640182733536,
0.1501023918390274,
-0.01887405291199684,
0.08188319951295853,
-0.04815459996461868,
0.13316422700881958,
0.012543330900371075,
0.010100416839122772,
-0.016554024070501328,
0.06554203480482101,
0.08129457384347916,
-0.043384142220020294,
-0.04087153449654579,
-0.049532148987054825,
0.006868566852062941,
-0.0598481222987175,
0.03835045546293259,
-0.01807943731546402,
0.09042128175497055,
-0.0730445608496666,
0.04594431072473526,
-0.05674472078680992,
0.0741804838180542,
0.11830837279558182,
-0.05590513348579407,
-0.07092452794313431,
-0.021398372948169708,
-0.0023719898890703917,
0.03441103547811508,
-0.044204626232385635,
-0.06084475666284561,
0.004032137803733349,
0.015381704084575176,
-0.06745235621929169,
-0.019634300842881203,
0.013035089708864689,
-0.03904129937291145,
-4.351946777024951e-33,
-0.01620015874505043,
-0.005266932304948568,
0.008350197225809097,
0.13314148783683777,
0.024267978966236115,
-0.02507350593805313,
0.05457567423582077,
-0.05994512140750885,
-0.011907514184713364,
-0.07593172788619995,
-0.04014485701918602,
-0.010999533347785473,
0.01734018325805664,
-0.031318362802267075,
0.021968601271510124,
-0.05726964771747589,
-0.08338173478841782,
-0.023006312549114227,
0.07873574644327164,
-0.022972354665398598,
0.01704580895602703,
0.07717239856719971,
-0.07593733817338943,
0.028237709775567055,
-0.04391065239906311,
0.0050505418330430984,
-0.05496075004339218,
0.05226153880357742,
0.005231218412518501,
0.05782260745763779,
0.00005929519102210179,
0.010105801746249199,
0.018750544637441635,
-0.01377329882234335,
-0.05062941834330559,
0.033377036452293396,
0.0663812905550003,
0.05900295823812485,
0.014365067705512047,
0.09437341243028641,
0.0244846660643816,
0.03070940636098385,
-0.038994286209344864,
-0.021273965016007423,
-0.025840185582637787,
0.026432102546095848,
-0.09564735740423203,
-0.049265552312135696,
0.05710703134536743,
0.030155980959534645,
0.047361742705106735,
0.0070816511288285255,
-0.047244105488061905,
-0.03212602436542511,
-0.013845846988260746,
-0.1171988695859909,
0.02146054431796074,
-0.05353216454386711,
-0.05512863025069237,
0.014963259920477867,
-0.041060928255319595,
-0.007965491153299809,
0.041869305074214935,
-0.04044516012072563,
0.09745386242866516,
0.024934614077210426,
-0.03559911996126175,
0.010450711473822594,
-0.08361779898405075,
0.04746541008353233,
0.06263107806444168,
0.00368137308396399,
0.020652256906032562,
-0.06336718052625656,
-0.004582096356898546,
0.007757033687084913,
-0.057423319667577744,
-0.03427780792117119,
-0.06435590982437134,
0.0626792386174202,
-0.027770278975367546,
0.007866671308875084,
0.001805850537493825,
0.07636203616857529,
-0.013339612632989883,
0.09707683324813843,
0.050660088658332825,
0.07174037396907806,
0.00584351597353816,
-0.008507205173373222,
-0.05157306790351868,
0.05677929148077965,
0.03934592381119728,
-0.0009255354525521398,
-0.016719741746783257,
-5.5646804497655467e-8,
0.030454210937023163,
-0.01882948912680149,
-0.028301166370511055,
0.09722676873207092,
0.06014785170555115,
0.016192467883229256,
-0.05587582290172577,
0.09326803684234619,
0.026842933148145676,
0.0039136167615652084,
0.11947867274284363,
-0.03466768190264702,
-0.04534660652279854,
-0.019879555329680443,
0.07216356694698334,
0.0337895043194294,
0.016425907611846924,
0.051508828997612,
-0.05213187262415886,
-0.040724411606788635,
0.03212117776274681,
0.04542506858706474,
0.04197957366704941,
-0.037636060267686844,
0.01293076854199171,
-0.07588132470846176,
-0.04054673761129379,
0.07183512300252914,
0.06018129736185074,
0.06510329991579056,
-0.019878992810845375,
-0.11433998495340347,
-0.010012982413172722,
-0.03480476513504982,
0.07481712102890015,
0.020246710628271103,
-0.05185778811573982,
0.0003200616338290274,
0.018208106979727745,
-0.006798131391406059,
0.019503220915794373,
0.03110773116350174,
-0.035190027207136154,
0.06716107577085495,
-0.004985279869288206,
-0.07717465609312057,
-0.11022061109542847,
-0.10403744131326675,
0.002645272994413972,
-0.04450340196490288,
0.012181789614260197,
-0.005475759040564299,
-0.10465976595878601,
0.008430473506450653,
0.06547309458255768,
0.09685959666967392,
-0.06925110518932343,
-0.0552762970328331,
-0.02616818994283676,
0.04953989014029503,
0.021868837997317314,
-0.03505565971136093,
-0.07437413930892944,
-0.03531506657600403
] | 0 | 0 | 0 | 0 |
|
P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS
|
https://openreview.net/forum?id=Sf4ep9Udjf
|
[
"Malyaban Bal",
"Abhronil Sengupta"
] |
Poster
|
Spiking neural networks (SNNs) are posited as a computationally efficient and biologically plausible alternative to conventional neural architectures, with their core computational framework primarily using the leaky integrate-and-fire (LIF) neuron model. However, the limited hidden state representation of LIF neurons, characterized by a scalar membrane potential, and sequential spike generation process, poses challenges for effectively developing scalable spiking models to address long-range dependencies in sequence learning tasks. In this study, we develop a scalable probabilistic spiking learning framework for long-range dependency tasks leveraging the fundamentals of state space models. Unlike LIF neurons that rely on the deterministic Heaviside function for a sequential process of spike generation, we introduce a SpikeSampler layer that samples spikes stochastically based on an SSM-based neuronal model while allowing parallel computations. To address non-differentiability of the spiking operation and enable effective training, we also propose a surrogate function tailored for the stochastic nature of the SpikeSampler layer. To enhance inter-neuron communication, we introduce the SpikeMixer block, which integrates spikes from neuron populations in each layer. This is followed by a ClampFuse layer, incorporating a residual connection to capture complex dependencies, enabling scalability of the model. Our models attain state-of-the-art performance among SNN models across diverse long-range dependency tasks, encompassing the Long Range Arena benchmark, permuted sequential MNIST, and the Speech Command dataset and demonstrate sparse spiking pattern highlighting its computational efficiency.
|
Spiking Neural Networks, Sequence Learning
|
Computationally efficient solution for long-range dependency tasks using spiking neural networks.
| 11,908 | null |
[
-0.10685671120882034,
-0.18077042698860168,
0.03789861127734184,
0.024317597970366478,
-0.015112284570932388,
0.0770164281129837,
0.022697286680340767,
-0.07804666459560394,
0.03977782651782036,
-0.04729130491614342,
-0.008227133192121983,
-0.059400252997875214,
-0.078760527074337,
0.0795229971408844,
-0.01927291788160801,
-0.02287890948355198,
-0.0031399319414049387,
0.007093580439686775,
-0.019954701885581017,
-0.004901327658444643,
0.09437042474746704,
-0.022461293265223503,
-0.004679081030189991,
0.06628993898630142,
0.010323012247681618,
0.026317590847611427,
0.03034256584942341,
-0.048568762838840485,
0.004144730046391487,
0.017190298065543175,
0.10172034054994583,
-0.06525594741106033,
-0.059005312621593475,
-0.04601414501667023,
-0.01887533999979496,
0.023866361007094383,
-0.09970953315496445,
-0.03537217527627945,
-0.02550794743001461,
-0.00427635433152318,
0.1125781238079071,
0.015909861773252487,
0.046628501266241074,
0.05342130735516548,
0.015481659211218357,
0.0043929265812039375,
0.07447565346956253,
-0.0718027651309967,
-0.032767802476882935,
-0.006242037285119295,
0.00031930545810610056,
0.026389483362436295,
0.020348623394966125,
0.02756396122276783,
0.01381224486976862,
0.07024180889129639,
0.05085314065217972,
0.030199162662029266,
-0.03771987184882164,
-0.003921080846339464,
-0.08818882703781128,
-0.06920488178730011,
0.04123566299676895,
-0.06825871765613556,
-0.07024775445461273,
-0.004944593645632267,
0.05144696310162544,
0.06434978544712067,
0.02853500470519066,
0.06308688968420029,
-0.022837989032268524,
0.12057516723871231,
-0.033868011087179184,
-0.020126085728406906,
0.007270650938153267,
0.0016161927487701178,
0.05467693507671356,
0.0021380120888352394,
0.03491413965821266,
-0.1348361074924469,
0.03591332212090492,
0.021520188078284264,
0.030713776126503944,
0.0024909654166549444,
0.12997570633888245,
0.030453301966190338,
-0.0009797221282497048,
0.059340424835681915,
-0.021230941638350487,
0.0015101052122190595,
-0.015511518344283104,
-0.047838423401117325,
0.030565327033400536,
-0.04050738736987114,
-0.02807333506643772,
-0.0005084235453978181,
0.03556593880057335,
0.0019213901832699776,
-0.024337278679013252,
0.036727458238601685,
-0.046748511493206024,
0.026489466428756714,
0.044542714953422546,
0.02285303920507431,
0.0821610689163208,
-0.0062886993400752544,
-0.039731115102767944,
0.0638231486082077,
-0.014168735593557358,
-0.04303302615880966,
0.05821770802140236,
0.04822594299912453,
0.03014673851430416,
0.06675592064857483,
0.08239887654781342,
-0.024794217199087143,
0.007568620145320892,
-0.03287988528609276,
0.06859003752470016,
0.044588591903448105,
-0.027367688715457916,
-0.005747449118643999,
-0.0718449205160141,
0.022165989503264427,
-0.031010035425424576,
0.01831110753118992,
-0.09243213385343552,
1.728078633109498e-33,
-0.019917789846658707,
-0.02190452441573143,
0.04122617840766907,
-0.032167598605155945,
0.020115366205573082,
-0.06123856455087662,
0.01699722558259964,
0.013117091730237007,
-0.036272257566452026,
0.04362652450799942,
-0.06676506251096725,
0.01793552003800869,
-0.01628142036497593,
0.07086853682994843,
0.04986730217933655,
-0.06450793892145157,
-0.06123613938689232,
-0.04726238548755646,
0.02808932214975357,
-0.14429426193237305,
0.031783122569322586,
-0.05704232305288315,
-0.01646457612514496,
-0.04652215912938118,
0.04191732034087181,
-0.0057153659872710705,
0.04762832820415497,
-0.05149073153734207,
-0.012111939489841461,
0.002345337299630046,
-0.04992819204926491,
0.06178533658385277,
-0.02371826395392418,
-0.07779999822378159,
0.07068528980016708,
-0.07684439420700073,
0.008748742751777172,
-0.03805048018693924,
-0.009916860610246658,
-0.09571324288845062,
0.041679542511701584,
-0.03754847124218941,
0.03454987332224846,
-0.007733146194368601,
-0.09900099784135818,
-0.081569604575634,
0.06122751906514168,
0.024182602763175964,
0.016924969851970673,
-0.0963725745677948,
0.059853728860616684,
-0.05608471482992172,
0.022602353245019913,
-0.01863129995763302,
0.02412465587258339,
-0.010023844428360462,
0.023590873926877975,
0.012196100316941738,
0.07683088630437851,
0.06214309111237526,
-0.006094224750995636,
-0.003554277354851365,
0.012162890285253525,
0.09850065410137177,
0.08193601667881012,
0.06829588860273361,
-0.011422054842114449,
0.04136193171143532,
0.0785837322473526,
-0.1014213040471077,
0.06051437556743622,
0.02724779210984707,
-0.018951354548335075,
-0.056446705013513565,
-0.010784248821437359,
0.04344486817717552,
0.02474820613861084,
-0.029948201030492783,
-0.10378662496805191,
0.0003322773554828018,
0.013469326309859753,
-0.12088310718536377,
-0.022263644263148308,
0.061163660138845444,
-0.024323061108589172,
-0.05295217037200928,
0.036688320338726044,
-0.04196566343307495,
-0.061394818127155304,
0.00841780100017786,
-0.05437324941158295,
-0.059440210461616516,
0.0734514519572258,
-0.06666576862335205,
-0.08061734586954117,
-2.3091179602102698e-33,
-0.07950308918952942,
0.019715847447514534,
-0.01748570054769516,
0.04717215150594711,
0.06314031779766083,
-0.023806922137737274,
0.03821476921439171,
-0.006780595984309912,
-0.041993558406829834,
-0.004484481178224087,
-0.12576444447040558,
-0.009208430536091328,
0.026613453403115273,
-0.017514893785119057,
0.06845215708017349,
-0.02298426441848278,
-0.055809520184993744,
0.07111655920743942,
0.01898760162293911,
0.005513019859790802,
-0.007064417004585266,
0.05431871861219406,
-0.10905587673187256,
0.031507641077041626,
-0.04000044986605644,
0.06340411305427551,
-0.008014031685888767,
0.0469963401556015,
-0.013170046731829643,
-0.01889895647764206,
-0.09090565890073776,
0.024279791861772537,
-0.05480532348155975,
-0.09173205494880676,
-0.035097889602184296,
0.030050231143832207,
0.054147105664014816,
-0.023187080398201942,
-0.032395366579294205,
-0.010022155940532684,
0.05746342986822128,
0.023735277354717255,
0.030752291902899742,
0.044419120997190475,
0.05288032442331314,
0.014188731089234352,
-0.06669020652770996,
0.056564562022686005,
-0.01795763522386551,
0.06390286237001419,
-0.07466200739145279,
0.04282143712043762,
-0.11930057406425476,
0.012248395010828972,
-0.10022891312837601,
-0.06326167285442352,
-0.02414545975625515,
-0.030027300119400024,
0.04211689904332161,
-0.04215617850422859,
0.0027364224661141634,
-0.07739675790071487,
0.015547327697277069,
-0.006561298854649067,
0.01699499413371086,
-0.05180699750781059,
-0.009997767396271229,
-0.030611561611294746,
0.05246412754058838,
-0.017663443461060524,
0.04915182292461395,
0.029308009892702103,
0.062078338116407394,
0.00951574370265007,
-0.06389947980642319,
-0.056786127388477325,
-0.0074808066710829735,
-0.034290723502635956,
0.04204339534044266,
-0.014554851688444614,
0.08613396435976028,
-0.03148254007101059,
-0.028226768597960472,
0.03260410204529762,
0.04185308516025543,
0.07444573938846588,
0.0649600699543953,
0.04846631735563278,
0.04843742772936821,
-0.07383454591035843,
-0.0011511199409142137,
0.027171598747372627,
-0.02233690768480301,
0.033864911645650864,
-0.0007137529319152236,
-3.8473043417752706e-8,
0.03749113902449608,
0.06037295609712601,
0.038918718695640564,
0.03562841936945915,
0.07299425452947617,
-0.032827723771333694,
0.03268207237124443,
-0.006815160624682903,
-0.024445699527859688,
-0.06957501173019409,
0.08727102726697922,
-0.05225251242518425,
0.039329249411821365,
-0.0562959685921669,
0.008769598789513111,
0.07136537879705429,
0.033066391944885254,
0.012322209775447845,
0.002839775988832116,
-0.0094613591209054,
0.008608649484813213,
0.05384308844804764,
-0.012404858134686947,
0.06968171894550323,
0.002227559220045805,
-0.050892747938632965,
-0.020643385127186775,
0.0625227689743042,
0.09847797453403473,
-0.09627419710159302,
0.07435448467731476,
0.01615416817367077,
0.059814538806676865,
0.02312644012272358,
0.0034863813780248165,
0.10331998765468597,
0.011668787337839603,
0.02476467192173004,
0.05105778947472572,
-0.012370034120976925,
-0.05830966308712959,
0.013535487465560436,
0.01763494312763214,
0.03167576715350151,
-0.11804021149873734,
-0.0268318522721529,
-0.020758436992764473,
-0.09339314699172974,
0.00697668083012104,
0.016578685492277145,
-0.00961587019264698,
0.026066796854138374,
-0.05079023540019989,
0.032343387603759766,
0.026343293488025665,
0.04266327992081642,
0.008661055006086826,
-0.03694920986890793,
-0.010993494652211666,
0.07270413637161255,
0.05260531231760979,
0.010642249137163162,
-0.09319137036800385,
0.051318347454071045
] | 0 | 0 | 0 | 0 |
|
Trained Transformer Classifiers Generalize and Exhibit Benign Overfitting In-Context
|
https://openreview.net/forum?id=jwsPS8yRe4
|
[
"Spencer Frei",
"Gal Vardi"
] |
Poster
|
Transformers have the capacity to act as supervised learning algorithms: by properly encoding a set of labeled training (''in-context'') examples and an unlabeled test example into an input sequence of vectors of the same dimension, the forward pass of the transformer can produce predictions for that unlabeled test example. A line of recent work has shown that when linear transformers are pre-trained on random instances for linear regression tasks, these trained transformers make predictions using an algorithm similar to that of ordinary least squares. In this work, we investigate the behavior of linear transformers trained on random linear classification tasks. Via an analysis of the implicit regularization of gradient descent, we characterize how many pre-training tasks and in-context examples are needed for the trained transformer to generalize well at test-time. We further show that in some settings, these trained transformers can exhibit ''benign overfitting in-context'': when in-context examples are corrupted by label flipping noise, the transformer memorizes all of its in-context examples (including those with noisy labels) yet still generalizes near-optimally for clean test examples.
|
in-context learning, generalization, benign overfitting, implicit regularization
| null | 11,902 |
2410.01774
|
[
-0.11355608701705933,
-0.0006820276030339301,
0.022299909964203835,
0.0651581659913063,
0.10687857866287231,
0.006409898400306702,
0.03579184412956238,
-0.024944163858890533,
-0.04169381037354469,
-0.0378190241754055,
-0.054801180958747864,
0.044457439333200455,
0.02175174094736576,
0.05702434480190277,
-0.0010122420499101281,
-0.021546056494116783,
0.0497969426214695,
0.07617233693599701,
-0.10578808188438416,
-0.020743023604154587,
0.021956725046038628,
0.043078020215034485,
0.016459763050079346,
0.04487709701061249,
0.04379632696509361,
-0.0229045357555151,
0.023130714893341064,
-0.04312225058674812,
-0.015990978106856346,
0.01727517507970333,
0.0018929862417280674,
-0.02740689553320408,
-0.0640571117401123,
0.05540843680500984,
-0.029265021905303,
0.015829522162675858,
-0.07027292996644974,
0.09285420179367065,
0.0352204293012619,
0.014431300573050976,
0.07405176758766174,
-0.04554424807429314,
0.028075270354747772,
0.030267760157585144,
0.02280333824455738,
-0.007311229128390551,
0.0276652704924345,
-0.11139463633298874,
-0.02055775560438633,
-0.0293078925460577,
0.0018891113577410579,
-0.048981960862874985,
-0.04360261932015419,
0.07278531789779663,
-0.06063759699463844,
0.007685516495257616,
0.11284384876489639,
0.01786203868687153,
0.024426132440567017,
0.03847704082727432,
-0.05593736842274666,
-0.11017953604459763,
-0.013767572119832039,
-0.03222593665122986,
-0.10353891551494598,
0.014901627786457539,
0.04185868427157402,
0.027446968480944633,
-0.0887637808918953,
-0.013550333678722382,
-0.06127874180674553,
0.11850302666425705,
-0.04733101278543472,
0.0793699100613594,
0.09779253602027893,
-0.009476655162870884,
0.02985375188291073,
-0.014694353565573692,
0.03261519968509674,
-0.031289469450712204,
0.027991604059934616,
0.022370291873812675,
0.030761167407035828,
-0.0036539833527058363,
0.14858397841453552,
0.013455153442919254,
-0.004065184388309717,
-0.02562999539077282,
-0.0076433513313531876,
0.014167305082082748,
-0.03465791046619415,
-0.0883217453956604,
-0.05232386291027069,
0.046995434910058975,
-0.0029464424587786198,
-0.013961700722575188,
0.01470456924289465,
-0.04348810389637947,
0.007292629219591618,
0.06607846915721893,
-0.02640787698328495,
-0.04254743084311485,
0.030049823224544525,
0.03423895314335823,
-0.01393874827772379,
0.009337032213807106,
-0.016361955553293228,
-0.02923760563135147,
0.059781961143016815,
-0.06508905440568924,
0.02977326326072216,
0.03192567825317383,
0.0009401921997778118,
-0.04396924749016762,
0.00883946381509304,
-0.03123694658279419,
0.0018299814546480775,
-0.022449661046266556,
-0.11646386981010437,
0.06630568206310272,
-0.029233569279313087,
0.04649229720234871,
0.0023121561389416456,
0.03214039281010628,
-0.001138019491918385,
0.03264437988400459,
-0.039722274988889694,
3.307538046076596e-33,
-0.031650006771087646,
0.018334070220589638,
0.024677375331521034,
0.012770572677254677,
-0.005086238496005535,
-0.016876710578799248,
0.009718877263367176,
0.06480266898870468,
0.10273753106594086,
0.04752740636467934,
0.0192122682929039,
0.04081171751022339,
-0.054536622017621994,
0.05681292712688446,
-0.03932174667716026,
0.053803350776433945,
-0.0034606116823852062,
0.045461516827344894,
-0.016388295218348503,
-0.031278662383556366,
0.027558093890547752,
-0.04088136553764343,
0.0008546846802346408,
-0.09975986182689667,
-0.05905507877469063,
-0.04299664497375488,
0.030358077958226204,
-0.034134574234485626,
0.010204914025962353,
-0.00107618048787117,
-0.07035987079143524,
0.011092506349086761,
0.019916705787181854,
0.04210437834262848,
0.01918724738061428,
-0.03466548025608063,
0.04932757839560509,
0.003003206104040146,
-0.04366815835237503,
0.0047713518142700195,
-0.029711328446865082,
0.02075672522187233,
0.0015453397063538432,
-0.003821172285825014,
0.03639231622219086,
-0.05483158677816391,
0.018215758726000786,
-0.0444190539419651,
0.004236262291669846,
-0.048809248954057693,
-0.007426480762660503,
0.011798230931162834,
-0.11231500655412674,
-0.009600907564163208,
0.041732821613550186,
0.10664859414100647,
0.038544751703739166,
-0.03411819785833359,
0.0037377355620265007,
0.025727050378918648,
-0.03369930014014244,
0.05513331666588783,
-0.007164234295487404,
0.02392142452299595,
-0.019785447046160698,
-0.025242524221539497,
0.08502382785081863,
0.004991341382265091,
0.01575690507888794,
-0.017759906128048897,
-0.01413500402122736,
-0.01668952777981758,
-0.09438710659742355,
-0.05314619094133377,
0.08044292777776718,
0.0010418187594041228,
0.09089509397745132,
0.029539810493588448,
-0.03076758235692978,
-0.016918916255235672,
0.03595716878771782,
-0.029547536745667458,
0.0467616505920887,
-0.029364587739109993,
-0.04814426973462105,
-0.035222481936216354,
-0.009353159926831722,
-0.131100133061409,
-0.012986946851015091,
-0.04847368597984314,
-0.08025391399860382,
0.01993885263800621,
-0.08924952894449234,
0.0052290684543550014,
-0.012392209842801094,
-3.3830378441780034e-33,
-0.0297189150005579,
0.061028845608234406,
-0.0479363352060318,
0.06711802631616592,
-0.018271906301379204,
-0.04965595901012421,
0.0035955028142780066,
0.04514461010694504,
-0.027507277205586433,
-0.008713673800230026,
-0.009323423728346825,
-0.05522748455405235,
-0.008322915993630886,
-0.04243958741426468,
0.02996913343667984,
-0.016361894086003304,
-0.049136172980070114,
0.05860154330730438,
-0.003445587120950222,
0.024825360625982285,
0.05379638448357582,
0.14942851662635803,
-0.05764451622962952,
0.029954828321933746,
-0.10437291860580444,
0.013211775571107864,
-0.13509589433670044,
0.13993337750434875,
0.10219272971153259,
-0.004428621847182512,
-0.00025957031175494194,
-0.0043138363398611546,
0.037775568664073944,
-0.001134518999606371,
-0.028666991740465164,
0.06647545844316483,
-0.0036392302718013525,
-0.012947489507496357,
0.0019913683645427227,
0.12205886840820312,
0.04134527966380119,
-0.000655314710456878,
-0.07278569042682648,
0.042223282158374786,
-0.05192058905959129,
-0.08351313322782516,
-0.009649014100432396,
-0.09430049359798431,
0.04845752567052841,
0.03328144550323486,
-0.05370388180017471,
0.05398046597838402,
-0.11578106135129929,
0.027083642780780792,
-0.04875342547893524,
-0.08379879593849182,
-0.019795890897512436,
-0.007940154522657394,
0.06767313927412033,
0.014281199313700199,
-0.029214389622211456,
0.0596749447286129,
0.055909622460603714,
-0.048698797821998596,
-0.001266292529180646,
-0.03849056363105774,
-0.011132732965052128,
-0.013861074112355709,
0.11180038750171661,
0.09841752797365189,
-0.010853931307792664,
0.0295871552079916,
-0.028920719400048256,
-0.051074881106615067,
-0.03603525087237358,
0.008132919669151306,
-0.011755373328924179,
-0.07794633507728577,
-0.003129752352833748,
-0.027393175289034843,
-0.010814463719725609,
-0.039139848202466965,
-0.024416305124759674,
0.06355180591344833,
0.014084922149777412,
0.052808601409196854,
0.056109409779310226,
0.05118512362241745,
0.07825917750597,
-0.04088861122727394,
0.016345040872693062,
0.02803918346762657,
-0.030067581683397293,
0.024224350228905678,
-0.04004089906811714,
-4.598340197503603e-8,
-0.11126672476530075,
0.06665121763944626,
-0.03427693620324135,
-0.032755713909864426,
0.02048914134502411,
-0.05250973999500275,
0.042780883610248566,
0.1588328778743744,
-0.08895538002252579,
0.0012803565477952361,
0.035414163023233414,
-0.01810905896127224,
-0.042466793209314346,
-0.03596675023436546,
-0.030149580910801888,
0.07015763968229294,
-0.02462722361087799,
0.07428910583257675,
0.017787933349609375,
0.04251015931367874,
-0.005581926088780165,
0.05949496850371361,
-0.0030610922258347273,
-0.06713293492794037,
0.02873135171830654,
-0.0936845913529396,
0.034460220485925674,
0.06508251279592514,
0.05081969127058983,
0.06413813680410385,
-0.03217647597193718,
0.030585983768105507,
-0.02168944850564003,
-0.04994184151291847,
-0.04071183502674103,
0.13778561353683472,
0.06320594996213913,
-0.08915536105632782,
0.004021556116640568,
-0.007279353681951761,
-0.011543470434844494,
0.06277652829885483,
-0.06596959382295609,
0.01591242104768753,
0.011340482160449028,
0.002386590465903282,
0.03693317249417305,
-0.10416251420974731,
-0.005777738988399506,
-0.08970598876476288,
0.040182847529649734,
-0.005639201495796442,
-0.036472778767347336,
0.033373236656188965,
0.07509371638298035,
0.002728625200688839,
-0.043007221072912216,
-0.06569397449493408,
-0.04990343004465103,
0.10575941950082779,
-0.0500151589512825,
0.052734702825546265,
-0.03778383508324623,
-0.0014492711052298546
] |
https://github.com/spencerfrei/icl_classification
| 0 | 0 | 0 | 0 |
PolyPythias: Stability and Outliers across Fifty Language Model Pre-Training Runs
|
https://openreview.net/forum?id=bmrYu2Ekdz
|
[
"Oskar van der Wal",
"Pietro Lesci",
"Max Müller-Eberstein",
"Naomi Saphra",
"Hailey Schoelkopf",
"Willem Zuidema",
"Stella Biderman"
] |
Poster
|
The stability of language model pre-training and its effects on downstream performance are still understudied. Prior work shows that the training process can yield significantly different results in response to slight variations in initial conditions, e.g., the random seed.
Crucially, the research community still lacks sufficient resources and tools to systematically investigate pre-training stability, particularly for decoder-only language models. We introduce the PolyPythias, a set of 45 new training runs for the Pythia model suite: 9 new seeds across 5 model sizes, from 14M to 410M parameters, resulting in about 7k new checkpoints that we release. Using these new 45 training runs, in addition to the 5 already available, we study the effects of different initial conditions determined by the seed---i.e., parameters' initialisation and data order---on (i) downstream performance, (ii) learned linguistic representations, and (iii) emergence of training phases. In addition to common scaling behaviours, our analyses generally reveal highly consistent training dynamics across both model sizes and initial conditions. Further, the new seeds for each model allow us to identify outlier training runs and delineate their characteristics.
Our findings show the potential of using these methods to predict training stability.
|
language models, training dynamics, interpretability, memorization, robustness, training stability
|
We introduce a set of 45 new training runs for the Pythia models suite to study stability across size and random seed.
| 11,899 |
2503.09543
|
[
-0.03487711772322655,
-0.14479467272758484,
0.0416145846247673,
0.08768787980079651,
0.01876729167997837,
0.014721068553626537,
-0.0480085089802742,
-0.019430262967944145,
0.06016362085938454,
-0.059496160596609116,
-0.008707097731530666,
0.038729604333639145,
-0.0336245633661747,
-0.00231066532433033,
-0.04241138696670532,
-0.04264074191451073,
0.0020661521703004837,
0.004203074146062136,
-0.10730928927659988,
-0.0235355943441391,
0.012738875113427639,
0.005282768979668617,
0.008037216030061245,
0.0628262385725975,
0.05202582851052284,
-0.03791821748018265,
-0.031632956117391586,
-0.015399254858493805,
0.0704166516661644,
-0.04392173886299133,
0.010736700147390366,
0.05601605772972107,
0.02004428394138813,
0.06173977628350258,
0.018215132877230644,
0.04920618236064911,
-0.0323028638958931,
-0.011961633339524269,
0.04295482859015465,
-0.0009584715589880943,
0.050997693091630936,
-0.013510342687368393,
-0.001455751247704029,
-0.013471821323037148,
0.03703586012125015,
-0.01998096890747547,
-0.006424264051020145,
-0.02691653184592724,
-0.08312863111495972,
-0.05434790998697281,
-0.06984250992536545,
-0.03750801831483841,
0.021852992475032806,
0.05675191432237625,
-0.019514214247465134,
-0.03458445519208908,
0.08703427761793137,
0.06306494772434235,
-0.013024283573031425,
-0.026939570903778076,
-0.10120289772748947,
-0.11102518439292908,
-0.06061873584985733,
-0.0015126687940210104,
-0.01986313983798027,
0.028868455439805984,
0.04289621114730835,
0.05774033069610596,
0.049002837389707565,
0.04699237644672394,
-0.0019942254293709993,
0.08360344171524048,
-0.03322065249085426,
0.03135961666703224,
0.04143616929650307,
0.010939841158688068,
0.06173699349164963,
-0.0008030172903090715,
-0.007895763963460922,
-0.07217779010534286,
0.02336997538805008,
0.010895962826907635,
0.03575904667377472,
-0.07112080603837967,
0.06348627060651779,
0.013054479844868183,
0.04194256290793419,
0.0493590272963047,
0.05404318869113922,
0.010385735891759396,
-0.03192576766014099,
0.0038142625708132982,
0.015671879053115845,
0.0283579733222723,
0.009240403771400452,
0.08327604830265045,
0.040561627596616745,
-0.03754434734582901,
0.04102560505270958,
0.04355666786432266,
0.06113249808549881,
0.041021645069122314,
0.038531478494405746,
0.02433224953711033,
0.030095016583800316,
-0.07244958728551865,
0.049408674240112305,
-0.016551213338971138,
0.03319134563207626,
0.003956385422497988,
0.02084595337510109,
0.04306859150528908,
0.03248042240738869,
0.04537440091371536,
-0.00029491170425899327,
0.06203595921397209,
-0.03812882676720619,
-0.01304322574287653,
-0.031950030475854874,
0.059298835694789886,
-0.042041223496198654,
0.0048784296959638596,
0.0509183444082737,
-0.011656460352241993,
-0.06981726735830307,
-0.019730782136321068,
-0.09565277397632599,
6.450826929742182e-34,
0.08890683948993683,
-0.031413450837135315,
-0.014680786989629269,
-0.006502058822661638,
0.021065298467874527,
-0.032656632363796234,
0.05058061331510544,
-0.05370126664638519,
-0.025261035189032555,
-0.05085594952106476,
-0.07957671582698822,
-0.011419061571359634,
-0.06765501201152802,
0.05516031011939049,
-0.01875760778784752,
-0.008963390253484249,
-0.03496405482292175,
0.0965396985411644,
-0.06071186810731888,
0.015386608429253101,
0.11095800250768661,
-0.05410567671060562,
-0.040722813457250595,
-0.08046852797269821,
0.01057369727641344,
0.032649241387844086,
0.018681135028600693,
-0.008445844985544682,
-0.12655866146087646,
0.044407352805137634,
-0.1483767181634903,
-0.02276505157351494,
-0.028344666585326195,
0.03239011391997337,
0.019609592854976654,
-0.06927868723869324,
0.01001134142279625,
-0.06059206277132034,
0.030161671340465546,
-0.10539475828409195,
-0.019556285813450813,
0.06022503972053528,
0.03290090337395668,
-0.06616424769163132,
-0.039782699197530746,
-0.08209380507469177,
-0.031037842854857445,
-0.004181546159088612,
-0.09721915423870087,
-0.029096323996782303,
-0.012799940072000027,
0.043597884476184845,
0.011852875351905823,
-0.005091928876936436,
-0.039502520114183426,
0.026292378082871437,
0.08236601948738098,
-0.013662504032254219,
-0.016308119520545006,
0.07368195056915283,
0.05936842039227486,
0.006615293677896261,
-0.037863314151763916,
0.036617763340473175,
0.06461206823587418,
0.04428230971097946,
-0.0716998279094696,
0.04753036051988602,
0.0599803663790226,
0.00009342128032585606,
-0.013711548410356045,
-0.08242521435022354,
-0.043128229677677155,
-0.005404775962233543,
0.1296079307794571,
-0.039943549782037735,
0.040804434567689896,
-0.050639815628528595,
-0.040764544159173965,
0.08416448533535004,
-0.010181507095694542,
-0.04216216877102852,
-0.034257080405950546,
-0.06861469149589539,
-0.08100136369466782,
-0.048135317862033844,
0.06805665045976639,
-0.08732250332832336,
-0.024497302249073982,
-0.043469738215208054,
0.006626875605434179,
-0.018497517332434654,
0.05865459516644478,
-0.006966080982238054,
-0.02953965775668621,
-1.557495117350991e-33,
-0.01876184716820717,
0.03366139903664589,
-0.046677976846694946,
0.03731326013803482,
-0.042261917144060135,
-0.07997709512710571,
0.027217470109462738,
0.10317449271678925,
-0.06817677617073059,
-0.09158363193273544,
0.00876692682504654,
-0.0486253947019577,
0.05856534466147423,
0.01609940268099308,
0.0856667309999466,
-0.017176343128085136,
0.052820514887571335,
0.0840839222073555,
0.12376523017883301,
0.07798395305871964,
-0.0007565698469989002,
0.02668432891368866,
-0.13544586300849915,
0.028293883427977562,
-0.017177240923047066,
-0.009145315736532211,
-0.040091343224048615,
0.08694086223840714,
-0.027231449261307716,
0.022904623299837112,
-0.028938574716448784,
0.04616468772292137,
-0.05300383269786835,
0.07119384407997131,
-0.041579876095056534,
-0.006070717703551054,
-0.004178888630121946,
-0.10698862373828888,
0.047173503786325455,
0.06382496654987335,
0.10267656296491623,
0.02527482993900776,
-0.061546243727207184,
0.06786902993917465,
-0.021373314782977104,
-0.042629122734069824,
-0.04781678318977356,
0.009065643884241581,
-0.007468744181096554,
-0.021302398294210434,
0.0009233394521288574,
0.07870017737150192,
0.026609472930431366,
0.030420921742916107,
-0.018033482134342194,
-0.0989169329404831,
-0.015034397132694721,
0.006521180737763643,
-0.04718490689992905,
0.0007743657333776355,
-0.05944259092211723,
-0.033258065581321716,
0.025522803887724876,
-0.045884303748607635,
-0.030061909928917885,
0.012819412164390087,
-0.015932852402329445,
0.0488358773291111,
0.023223990574479103,
-0.045939844101667404,
0.03363202512264252,
-0.00516887940466404,
-0.04297119006514549,
0.07174786925315857,
-0.07222720980644226,
-0.04035906121134758,
-0.07771797478199005,
-0.05819318816065788,
-0.021369777619838715,
-0.00018454751989338547,
-0.06047303229570389,
0.043281350284814835,
-0.030262324959039688,
0.034971654415130615,
0.058951251208782196,
0.04154330864548683,
0.04572388157248497,
0.06770674884319305,
0.007271874230355024,
-0.021042974665760994,
0.018783453851938248,
0.01110305730253458,
-0.021490097045898438,
0.11180084198713303,
0.003038441762328148,
-4.8434394273044745e-8,
-0.0687660351395607,
0.011786132119596004,
0.004942679312080145,
0.10637804865837097,
0.08259267359972,
-0.02028033323585987,
-0.06364795565605164,
0.06266426295042038,
-0.015571177005767822,
-0.022618820890784264,
0.060358576476573944,
0.004005242604762316,
-0.04089480638504028,
-0.056700900197029114,
0.011930244974792004,
0.0916183590888977,
-0.01686285063624382,
0.03874657303094864,
-0.03205942362546921,
-0.031820476055145264,
0.06137947365641594,
0.05161014944314957,
0.013976366259157658,
-0.03553510457277298,
-0.035704631358385086,
-0.042333491146564484,
0.018717113882303238,
0.09780429303646088,
0.05597290024161339,
-0.0655888244509697,
0.004512902814894915,
0.011116456240415573,
-0.011681094765663147,
-0.008894923143088818,
-0.02894682250916958,
0.17476937174797058,
-0.023985598236322403,
-0.060094673186540604,
0.05739007890224457,
0.04335809126496315,
0.0707317590713501,
0.0096480343490839,
-0.051017846912145615,
-0.00041151620098389685,
0.011872517876327038,
0.03379473835229874,
-0.1100243404507637,
-0.07532498985528946,
0.04448332637548447,
-0.070637546479702,
0.02844877913594246,
0.03347829729318619,
-0.012106966227293015,
0.04985630139708519,
0.0796327069401741,
0.0415661446750164,
-0.09638268500566483,
0.0036327196285128593,
-0.022263620048761368,
0.03902212902903557,
0.0508914478123188,
-0.004995147697627544,
0.025202780961990356,
0.009106526151299477
] | 0 | 0 | 0 | 0 |
|
Adapters for Altering LLM Vocabularies: What Languages Benefit the Most?
|
https://openreview.net/forum?id=KxQRHOre9D
|
[
"HyoJung Han",
"Akiko Eriguchi",
"Haoran Xu",
"Hieu Hoang",
"Marine Carpuat",
"Huda Khayrallah"
] |
Poster
|
Vocabulary adaptation, which integrates new vocabulary into pre-trained language models, enables expansion to new languages and mitigates token over-fragmentation. However, existing approaches are limited by their reliance on heuristics or external embeddings. We propose VocADT, a novel method for vocabulary adaptation using adapter modules that are trained to learn the optimal linear combination of existing embeddings while keeping the model’s weights fixed. VocADT offers a flexible and scalable solution without depending on external resources or language constraints. Across 11 languages—with diverse scripts, resource availability, and fragmentation—we demonstrate that VocADT outperforms the original Mistral model (Jiang et al., 2023) and other baselines across various multilingual tasks including natural language understanding and machine translation. We find that Latin-script languages and highly fragmented languages
benefit the most from vocabulary adaptation. We further fine-tune the adapted model on the generative task of machine translation and find that vocabulary adaptation is still beneficial after fine-tuning and that VocADT is the most effective.
|
Vocabulary Adaptation, Vocabulary Transfer, Tokenizer Transfer, Initializing Embedding, Adapter, Multilingual, Machine Translation
|
We propose VocADT, a novel method for vocabulary adaptation using adapter modules, and explore what language benefits the most among with various scripts, resource availability, and fragmentation.
| 11,894 |
2410.09644
|
[
-0.03059852123260498,
-0.10681547969579697,
0.01722382940351963,
0.005093382205814123,
0.06897976994514465,
-0.03743119537830353,
-0.05873047560453415,
0.031043779104948044,
0.03221157565712929,
-0.05134370177984238,
0.01416877843439579,
-0.05621350556612015,
0.0025496226735413074,
-0.03757024183869362,
0.00895446352660656,
-0.002215827349573374,
0.09865164756774902,
0.10817914456129074,
-0.061184126883745193,
-0.03619327396154404,
0.036787401884794235,
0.028084078803658485,
0.018666379153728485,
-0.017488976940512657,
0.10174592584371567,
-0.023045320063829422,
-0.035383302718400955,
-0.029724709689617157,
0.06120815500617027,
-0.06784545630216599,
0.07098710536956787,
0.0787455290555954,
-0.08156100660562515,
0.04524717852473259,
-0.012816101312637329,
0.041911546140909195,
-0.07309985905885696,
0.021341370418667793,
-0.01125282421708107,
-0.008026214316487312,
-0.0071759410202503204,
0.023961877450346947,
0.000029008917408646084,
0.03047363832592964,
0.028705522418022156,
0.006269458215683699,
0.006257638335227966,
-0.003210934344679117,
-0.04070137068629265,
0.07351688295602798,
-0.02308039925992489,
-0.03505920618772507,
-0.045018233358860016,
0.07758551090955734,
-0.04628724604845047,
0.005747531540691853,
0.021640941500663757,
0.018811414018273354,
-0.028398020192980766,
-0.07529015094041824,
-0.055475614964962006,
-0.030803410336375237,
0.027128927409648895,
0.021116115152835846,
-0.043669022619724274,
-0.10582179576158524,
0.030360154807567596,
0.11831920593976974,
-0.01579948700964451,
0.06728573888540268,
-0.028445089235901833,
0.08675889670848846,
-0.03739028796553612,
0.045912597328424454,
0.005653525702655315,
0.005452241748571396,
0.014092076569795609,
-0.03671976178884506,
0.11668819189071655,
-0.052973318845033646,
0.04184316471219063,
0.10020487755537033,
0.005666887387633324,
0.015690647065639496,
0.13236141204833984,
0.007757261395454407,
0.011773374862968922,
0.03635619580745697,
0.01650319993495941,
0.0023477727081626654,
-0.029110003262758255,
-0.04225553572177887,
0.07302191853523254,
0.023251552134752274,
-0.022255253046751022,
0.011259227059781551,
0.01925753243267536,
-0.04263637959957123,
-0.016143105924129486,
0.07249758392572403,
0.02333088591694832,
0.10834892094135284,
0.013047143816947937,
-0.03993863984942436,
-0.0905195027589798,
-0.03937586769461632,
-0.00021758314687758684,
0.03462604060769081,
0.03139086440205574,
-0.02612094022333622,
0.07276749610900879,
0.09057937562465668,
0.036190398037433624,
-0.0953206792473793,
0.007116810884326696,
-0.016241198405623436,
0.023611823096871376,
-0.051546160131692886,
0.0043867905624210835,
0.06005874276161194,
-0.07070507109165192,
0.07037168741226196,
0.0033001829870045185,
-0.007398697081953287,
-0.025177808478474617,
-0.03400982543826103,
-0.039495207369327545,
-4.2990937041116e-34,
0.04839324951171875,
0.04760785773396492,
-0.02333306334912777,
0.049231115728616714,
0.06064125522971153,
-0.015526789240539074,
0.06906206160783768,
0.08426854759454727,
-0.10891509801149368,
-0.05085362121462822,
0.022263025864958763,
0.06193232163786888,
-0.039048127830028534,
0.058751776814460754,
0.026807881891727448,
-0.011091300286352634,
-0.0051524098962545395,
0.04026450216770172,
-0.012442308478057384,
-0.021864674985408783,
0.0007539804209955037,
0.02003069967031479,
0.01541292853653431,
0.007196919992566109,
0.0010034091537818313,
0.04501491039991379,
0.07538383454084396,
-0.051796235144138336,
-0.043725915253162384,
0.023888658732175827,
-0.128638356924057,
-0.03161688894033432,
0.031745824962854385,
0.00027158018201589584,
0.04355363920331001,
-0.10725966095924377,
-0.04623556509613991,
-0.06364040076732635,
-0.013472164049744606,
-0.03188153728842735,
-0.024871457368135452,
0.029989292845129967,
0.014949630945920944,
-0.027758324518799782,
-0.014467018656432629,
-0.051206719130277634,
0.01878698356449604,
0.045451756566762924,
-0.04611264169216156,
-0.07072597742080688,
0.03857957571744919,
-0.0012789997272193432,
-0.05882499739527702,
0.04885459318757057,
0.029888225719332695,
0.04079659283161163,
-0.016810163855552673,
0.007019813172519207,
0.018265657126903534,
-0.007553424686193466,
0.014313841238617897,
0.04580222815275192,
0.04520659148693085,
0.1729140430688858,
0.11254573613405228,
0.0370699018239975,
0.007646547164767981,
0.06548330932855606,
0.004993067122995853,
0.009880918078124523,
-0.08210011571645737,
-0.0069214231334626675,
-0.03486688807606697,
0.019397076219320297,
0.03170705959200859,
-0.012807314284145832,
0.013847354799509048,
-0.10935722291469574,
0.004854443483054638,
0.008482184261083603,
0.03944229707121849,
-0.007888991385698318,
0.02475832961499691,
-0.08208168298006058,
-0.03441648557782173,
-0.08665543049573898,
0.07723212242126465,
-0.06446067243814468,
0.040006548166275024,
0.04497255012392998,
-0.006717891898006201,
-0.08130496740341187,
-0.0024650241248309612,
-0.03534742072224617,
0.05522100254893303,
4.06288316526615e-34,
-0.028395116329193115,
-0.0678141862154007,
-0.01601317897439003,
0.06563140451908112,
-0.07808490842580795,
-0.07902704924345016,
-0.014517838135361671,
0.07715937495231628,
-0.03072776459157467,
-0.07983395457267761,
-0.08104363828897476,
-0.0489002987742424,
-0.0027571276295930147,
0.012336942367255688,
0.03777390718460083,
-0.004776459652930498,
-0.035939350724220276,
0.04528200998902321,
-0.03106031008064747,
0.095861054956913,
-0.043092917650938034,
0.08117470890283585,
-0.055891118943691254,
0.033647820353507996,
0.01111183874309063,
0.016030820086598396,
-0.05008438974618912,
0.057971417903900146,
-0.03340420499444008,
-0.027900995686650276,
-0.06829099357128143,
0.031602222472429276,
-0.04714971035718918,
0.014316757209599018,
-0.07153721898794174,
0.07770167291164398,
0.059601254761219025,
0.015601913444697857,
-0.01968557760119438,
0.06589163839817047,
0.03641679883003235,
-0.009184978902339935,
0.0006280877278186381,
-0.025440746918320656,
-0.025540687143802643,
-0.004248495679348707,
-0.1006103977560997,
-0.03871515393257141,
0.013885959051549435,
-0.010218367911875248,
0.029324358329176903,
-0.07463715225458145,
-0.04583863168954849,
0.020672352984547615,
0.06558284908533096,
-0.13003897666931152,
-0.007967905141413212,
-0.09160634130239487,
-0.010543659329414368,
-0.09170371294021606,
-0.03441478684544563,
-0.0037075767759233713,
0.004379578400403261,
-0.03778531029820442,
0.04112866148352623,
-0.04215901717543602,
-0.07665269076824188,
-0.02387777715921402,
0.014059887267649174,
-0.02763294242322445,
0.08250253647565842,
-0.06400881707668304,
-0.008336247876286507,
-0.03935997933149338,
0.0033302512019872665,
-0.018022695556282997,
0.029574653133749962,
-0.05152403190732002,
-0.0380551740527153,
-0.09552772343158722,
-0.024881500750780106,
-0.06954702734947205,
0.05010184273123741,
-0.018488392233848572,
0.02427760884165764,
0.08151284605264664,
0.009255805984139442,
0.09299486875534058,
0.024472694844007492,
0.07398427277803421,
-0.06280364841222763,
0.04334031417965889,
0.001578355091623962,
0.05580306425690651,
-0.044105254113674164,
-3.508993984269182e-8,
-0.09349911659955978,
0.059263523668050766,
-0.05922717601060867,
0.04908085614442825,
0.030269503593444824,
-0.09518812596797943,
-0.050748683512210846,
0.08908241242170334,
0.012958921492099762,
0.023626932874321938,
0.07129117101430893,
0.014441423118114471,
0.044994648545980453,
-0.03427960351109505,
-0.026548460125923157,
0.09322354197502136,
0.01783326081931591,
0.01812254823744297,
-0.0008189204963855445,
0.0006887229392305017,
-0.022047551348805428,
0.12200485169887543,
0.03463691473007202,
-0.001639463473111391,
0.036248382180929184,
-0.09698908776044846,
-0.03203277289867401,
0.05090669170022011,
0.09329966455698013,
-0.05616625398397446,
-0.0539938248693943,
0.01945212483406067,
-0.035253580659627914,
-0.02445903606712818,
-0.0010578915243968368,
0.07458796352148056,
0.017840251326560974,
-0.02456817589700222,
0.012412361800670624,
-0.006133394781500101,
0.11553971469402313,
0.09910281747579575,
-0.07828691601753235,
0.029122475534677505,
0.01958128623664379,
-0.06822068244218826,
-0.0281390193849802,
-0.05901436135172844,
-0.010460530407726765,
-0.10968324542045593,
-0.005528260953724384,
-0.00871218927204609,
0.0025969843845814466,
0.016129223629832268,
0.01842234656214714,
0.029672283679246902,
-0.05542709305882454,
0.011719869449734688,
0.055534061044454575,
0.014381048269569874,
-0.04392878711223602,
-0.04284071549773216,
0.00932948850095272,
0.010670083574950695
] |
https://github.com/h-j-han/VocADT
| 4 | 0 | 0 | 0 |
DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search
|
https://openreview.net/forum?id=tn2mjzjSyR
|
[
"Murong Yue",
"Wenlin Yao",
"Haitao Mi",
"Dian Yu",
"Ziyu Yao",
"Dong Yu"
] |
Poster
|
Enhancing the capability of large language models (LLMs) in reasoning has gained significant attention in recent years. Previous studies have demonstrated the effectiveness of various prompting strategies in aiding LLMs in reasoning (called "reasoning actions"), such as step-by-step thinking, reflecting before answering, solving with programs, and their combinations. However, these approaches often applied static, predefined reasoning actions uniformly to all questions, without considering the specific characteristics of each question or the capability of the task-solving LLM. In this paper, we propose DOTS, an approach enabling LLMs to reason Dynamically via Optimal reasoning Trajectories Search, tailored to the specific characteristics of each question and the inherent capability of the task-solving LLM.
Our approach involves three key steps: i) defining atomic reasoning action modules that can be composed into various reasoning action trajectories; ii) searching for the optimal action trajectory for each training question through iterative exploration and evaluation for the specific task-solving LLM; and iii) using the collected optimal trajectories to train an LLM to plan for the reasoning trajectories of unseen questions. In particular, we propose two learning paradigms, i.e., fine-tuning an external LLM as a planner to guide the task-solving LLM, or directly fine-tuning the task-solving LLM with an internalized capability for reasoning actions planning. Our experiments across eight reasoning tasks show that our method consistently outperforms static reasoning techniques and the vanilla instruction tuning approach. Further analysis reveals that our method enables LLMs to adjust their computation based on problem complexity, allocating deeper thinking and reasoning to harder problems.
|
large language model, reasoning
|
We propose a method to dynamically select an optimal trajectory of reasoning actions tailored to the specific characteristics of questions and LLMs.
| 11,883 |
2410.03864
|
[
0.01043359562754631,
-0.023953871801495552,
0.013386248610913754,
-0.0005317702307365835,
0.07803872227668762,
-0.019915109500288963,
0.07644069939851761,
0.09056563675403595,
-0.002481238916516304,
-0.012126278132200241,
-0.0602741613984108,
-0.04848567396402359,
0.02389509603381157,
0.0996875911951065,
0.07887312024831772,
0.02700590342283249,
0.042229436337947845,
0.02360496297478676,
-0.058150727301836014,
-0.06246392801403999,
-0.018491452559828758,
-0.050858497619628906,
0.08677620440721512,
0.0002461156400386244,
-0.06074201315641403,
0.031205976381897926,
0.08372901380062103,
-0.054395511746406555,
0.003661953378468752,
0.05337958782911301,
0.06780871003866196,
0.03275177627801895,
-0.010028921067714691,
0.021702919155359268,
0.00257780565880239,
0.06724107265472412,
-0.0497647263109684,
0.05147269368171692,
-0.03226863592863083,
-0.11456575989723206,
-0.05148690566420555,
-0.030304238200187683,
0.012951948679983616,
-0.017667124047875404,
0.03708631545305252,
-0.004518487490713596,
0.00696401996538043,
0.04137058183550835,
-0.0038555702194571495,
0.03218884393572807,
-0.08174562454223633,
-0.043805334717035294,
-0.061053317040205,
-0.059609271585941315,
0.03907323628664017,
0.03115413710474968,
0.043399419635534286,
0.04595571756362915,
-0.009330173954367638,
-0.0858219712972641,
-0.046791572123765945,
-0.11931566148996353,
-0.032123640179634094,
0.017513934522867203,
-0.06321772187948227,
-0.01412293966859579,
0.014505882747471333,
0.04445229470729828,
-0.0033811740577220917,
0.09363799542188644,
-0.03267587721347809,
0.06764435768127441,
-0.023840127512812614,
0.04134434089064598,
-0.08054821938276291,
0.07852240651845932,
0.011683998629450798,
-0.050654880702495575,
-0.0047533209435641766,
-0.0594576857984066,
0.004812764935195446,
0.02534966915845871,
0.007296750787645578,
0.027753109112381935,
0.13196852803230286,
0.03409850224852562,
-0.0004899342311546206,
0.04603848606348038,
0.07209713011980057,
0.0643586590886116,
0.0008101806160993874,
-0.08219347149133682,
0.05856134369969368,
0.012514869682490826,
0.044340942054986954,
0.09296656399965286,
0.010022884234786034,
-0.06367921829223633,
-0.009828160516917706,
0.09157958626747131,
0.06703941524028778,
0.12104154378175735,
-0.036868736147880554,
-0.06468427181243896,
0.015919048339128494,
0.008124742656946182,
-0.029356060549616814,
-0.008282817900180817,
0.05518908426165581,
-0.02950911968946457,
0.0028090898413211107,
-0.015145218931138515,
0.09004121273756027,
0.03244439885020256,
0.027982966974377632,
-0.03493466600775719,
0.08982181549072266,
-0.004169784020632505,
0.018847953528165817,
0.008848571218550205,
0.03251808136701584,
0.003267931519076228,
-0.023057447746396065,
0.03667132183909416,
0.057168345898389816,
-0.0915774554014206,
-0.08026903867721558,
2.511581250105062e-33,
0.098793163895607,
0.024841738864779472,
0.11412211507558823,
0.03367311507463455,
0.016491364687681198,
0.012714309617877007,
0.0738721489906311,
0.025682127103209496,
-0.028312038630247116,
0.016648976132273674,
-0.01765807904303074,
-0.06311464309692383,
-0.030532367527484894,
0.012799018062651157,
0.031341344118118286,
-0.06393955647945404,
-0.06424625217914581,
0.028552593663334846,
-0.05054779723286629,
-0.07025304436683655,
0.016270088031888008,
-0.0612785778939724,
0.0008549057529307902,
-0.00027985102497041225,
0.021924102678894997,
0.007142011076211929,
0.05115184187889099,
-0.05380811542272568,
-0.015574141405522823,
-0.029919998720288277,
-0.08788551390171051,
-0.08240822702646255,
-0.0779736265540123,
0.08433482050895691,
0.057672321796417236,
-0.030108200386166573,
-0.07434435933828354,
-0.03089032880961895,
0.007789510302245617,
-0.074823297560215,
-0.02311703749001026,
0.01061314158141613,
0.02882036752998829,
0.059466492384672165,
-0.052673500031232834,
-0.03455023840069771,
-0.09787922352552414,
0.06015269458293915,
-0.0585327222943306,
0.013949600048363209,
0.036904171109199524,
-0.0025799544528126717,
-0.002358796540647745,
-0.11826857179403305,
0.004872814752161503,
-0.013651268556714058,
-0.028595291078090668,
0.015295484103262424,
0.008482587523758411,
0.023963376879692078,
0.0075355879962444305,
-0.015583010390400887,
-0.03149404376745224,
0.1222357302904129,
0.025604356080293655,
0.07310015708208084,
-0.09636345505714417,
0.04561908170580864,
0.08451396226882935,
-0.007813069969415665,
-0.0021971201058477163,
0.017801649868488312,
0.0680832639336586,
-0.05265387147665024,
0.03309384360909462,
0.023228110745549202,
0.016690222546458244,
-0.11396539211273193,
0.0384104922413826,
-0.005980634596198797,
0.013327362015843391,
-0.027750235050916672,
-0.06731507927179337,
-0.024516573175787926,
0.02561795525252819,
-0.0798674002289772,
0.0005557762924581766,
-0.11485649645328522,
-0.0021746361162513494,
-0.07503897696733475,
-0.01813449338078499,
-0.04856456443667412,
-0.023968325927853584,
-0.03212699294090271,
-0.035152748227119446,
-2.1716865075518206e-33,
0.018933838233351707,
-0.054792098701000214,
-0.05696551874279976,
0.03278624638915062,
0.0036558161955326796,
-0.04339217767119408,
-0.0667242556810379,
-0.04974876344203949,
0.004002288915216923,
-0.05061434954404831,
-0.09039997309446335,
-0.04089924320578575,
0.02895292080938816,
-0.006141694262623787,
-0.017710993066430092,
0.006148712709546089,
-0.018742438405752182,
0.012148999609053135,
0.020715447142720222,
0.11374411731958389,
-0.04312874376773834,
0.08430655300617218,
-0.08096175640821457,
-0.03317722678184509,
-0.0284854918718338,
0.03043985553085804,
0.05098969489336014,
0.04552389681339264,
0.005449413321912289,
0.07802063971757889,
-0.06084543094038963,
-0.06273873150348663,
-0.024695971980690956,
-0.06491360813379288,
-0.07971924543380737,
0.01979062892496586,
0.0673777237534523,
0.0027282647788524628,
-0.04048464074730873,
0.05089559033513069,
0.051116105169057846,
-0.08512787520885468,
0.012770667672157288,
0.05023486167192459,
-0.03452087193727493,
0.056472986936569214,
-0.08116690814495087,
0.08112535625696182,
-0.028033820912241936,
0.012883814983069897,
-0.01605614833533764,
-0.055319152772426605,
0.014240311458706856,
-0.007028431165963411,
0.03444802761077881,
-0.037749968469142914,
-0.02546163648366928,
-0.01546459086239338,
0.016529563814401627,
-0.028086936101317406,
-0.030726559460163116,
0.02475673146545887,
0.061456214636564255,
0.00041532644536346197,
0.01735188625752926,
-0.019166145473718643,
-0.03451614826917648,
-0.009683001786470413,
-0.07112035900354385,
-0.03912314400076866,
0.05624241381883621,
-0.009760599583387375,
-0.026453934609889984,
0.014720719307661057,
0.14784905314445496,
-0.016774345189332962,
0.0326969213783741,
-0.08488932996988297,
-0.059966880828142166,
-0.0464908704161644,
-0.030950332060456276,
-0.006068788468837738,
0.005735959392040968,
0.11622599512338638,
-0.07511206716299057,
-0.030844472348690033,
0.03204352781176567,
0.13738320767879486,
0.04425518959760666,
0.04237007349729538,
-0.027690701186656952,
0.03312582150101662,
0.04833376035094261,
0.05111625790596008,
-0.07947169989347458,
-3.964644079701429e-8,
-0.06903064250946045,
0.04522073641419411,
-0.05207277461886406,
0.04807204380631447,
0.08724240958690643,
-0.012634092010557652,
-0.039970722049474716,
0.040524303913116455,
0.009783981367945671,
-0.025434622541069984,
0.08677064627408981,
0.0010071698343381286,
-0.057742878794670105,
-0.010796446353197098,
-0.02116376906633377,
0.0558704137802124,
0.037265483289957047,
-0.053385667502880096,
-0.062006644904613495,
-0.028512457385659218,
0.04212639853358269,
0.002890503965318203,
-0.06842613965272903,
0.09404084086418152,
0.015940766781568527,
-0.0136402053758502,
-0.031197084113955498,
0.11670780926942825,
0.08898372203111649,
0.02753637358546257,
-0.02440519630908966,
-0.008764772675931454,
0.045005716383457184,
-0.0022017706651240587,
0.03411204367876053,
0.026135416701436043,
0.0030111537780612707,
-0.05315728485584259,
0.00007406480290228501,
-0.00665412750095129,
-0.008319870568811893,
0.07838509976863861,
-0.039928000420331955,
0.05592399463057518,
-0.022701933979988098,
-0.012622187845408916,
-0.07678121328353882,
-0.05222740396857262,
-0.021013177931308746,
-0.04287351295351982,
-0.09589628130197525,
-0.01628422737121582,
-0.0228689257055521,
0.014370354823768139,
0.03762966766953468,
0.07556858658790588,
-0.0065191322937607765,
0.0054722442291677,
0.016411490738391876,
0.023144427686929703,
0.05375298857688904,
0.08436337113380432,
-0.0009786695009097457,
0.03302113711833954
] |
https://github.com/MurongYue/DOTS
| 1 | 0 | 0 | 0 |
Learning Diagrams: A Graphical Language for Compositional Training Regimes
|
https://openreview.net/forum?id=dqyuCsBvn9
|
[
"Mason Lary",
"Richard Samuelson",
"Alexander Wilentz",
"Alina Zare",
"Matthew Klawonn",
"James Fairbanks"
] |
Poster
|
Motivated by deep learning regimes with multiple interacting yet distinct model components, we introduce learning diagrams, graphical depictions of training setups that capture parameterized learning as data rather than code. A learning diagram compiles to a unique loss function on which component models are trained. The result of training on this loss is a collection of models whose predictions ``agree" with one another. We show that a number of popular learning setups such as few-shot multi-task learning, knowledge distillation, and multi-modal learning can be depicted as learning diagrams. We further implement learning diagrams in a library that allows users to build diagrams of PyTorch and Flux.jl models. By implementing some classic machine learning use cases, we demonstrate how learning diagrams allow practitioners to build complicated models as compositions of smaller components, identify relationships between workflows, and manipulate models during or after training. Leveraging a category theoretic framework, we introduce a rigorous semantics for learning diagrams that puts such operations on a firm mathematical foundation.
|
ML Libraries, Training, Foundation Models, Multi-Task Learning
|
We rigorously model training setups diagrammatically and provide a library for powerful manipulations thereof.
| 11,881 | null |
[
-0.013992899097502232,
-0.0708148255944252,
0.058790918439626694,
0.031923528760671616,
0.024047497659921646,
-0.02796967700123787,
-0.029659995809197426,
-0.03979839012026787,
0.022756533697247505,
-0.089826799929142,
-0.04703051596879959,
-0.022068412974476814,
0.0011166572803631425,
0.09518859535455704,
0.034272387623786926,
0.010031516663730145,
0.004432353191077709,
0.10432028025388718,
-0.09545649588108063,
-0.10110173374414444,
0.004991925787180662,
-0.04445889592170715,
0.02718055434525013,
-0.0023826099932193756,
0.017254378646612167,
-0.0007338401628658175,
-0.018844127655029297,
-0.0023989244364202023,
0.021807976067066193,
-0.03827069699764252,
0.012912847101688385,
-0.04384041950106621,
-0.09212971478700638,
0.045120857656002045,
-0.054973945021629333,
0.10512768477201462,
0.0057634832337498665,
0.0342508926987648,
-0.02080320194363594,
-0.01288725808262825,
-0.05158129334449768,
0.04846157133579254,
0.01043249387294054,
0.05299394205212593,
0.0503694973886013,
0.008393138647079468,
-0.04355302453041077,
-0.08457489311695099,
-0.0021325363777577877,
0.042624879628419876,
-0.11148959398269653,
-0.03771425038576126,
-0.037362758070230484,
-0.0017994699301198125,
0.046587660908699036,
0.024788333103060722,
0.0400945208966732,
-0.08974511921405792,
-0.025602081790566444,
-0.05879458412528038,
-0.07821453362703323,
-0.045997120440006256,
-0.004798381123691797,
-0.002634720876812935,
-0.024308819323778152,
-0.0025706328451633453,
-0.007656255271285772,
0.11547720432281494,
0.0325804129242897,
-0.001200817758217454,
-0.03238325193524361,
0.010805434547364712,
-0.06415844708681107,
-0.012539499439299107,
0.04426880553364754,
0.0073979804292321205,
0.03832269832491875,
0.03322077542543411,
0.019081033766269684,
-0.0972777009010315,
-0.005027415696531534,
0.07296649366617203,
0.010614258237183094,
0.044443510472774506,
0.0680675208568573,
0.014605076983571053,
-0.029535893350839615,
0.002161231357604265,
0.029080379754304886,
-0.039279308170080185,
-0.05587819963693619,
-0.027041567489504814,
0.04246954992413521,
0.006904821842908859,
0.07217609882354736,
0.07320660352706909,
0.04883686080574989,
-0.08293233066797256,
-0.06527508795261383,
0.12151051312685013,
-0.028970807790756226,
0.005186465103179216,
0.08765636384487152,
-0.05409114062786102,
0.014838365837931633,
-0.04049242287874222,
-0.008155633695423603,
0.06177236884832382,
0.0553298220038414,
-0.10971825569868088,
0.052454978227615356,
0.03586864843964577,
-0.006726219318807125,
0.021962443366646767,
-0.028727898374199867,
0.033450935035943985,
-0.0389157310128212,
-0.03214561194181442,
0.013934819959104061,
0.08176730573177338,
-0.02516496740281582,
0.07866853475570679,
-0.03818006068468094,
0.05112971365451813,
-0.012265958823263645,
-0.09900698810815811,
-0.12025034427642822,
2.6782762672852474e-33,
0.12758123874664307,
0.0051567633636295795,
0.09679395705461502,
0.10189151763916016,
0.08419021964073181,
-0.042960163205862045,
-0.020983880385756493,
0.011067952029407024,
0.015401076525449753,
0.01568608358502388,
-0.019461078569293022,
0.026416806504130363,
-0.11012094467878342,
0.10627885907888412,
0.003720369655638933,
0.011689242906868458,
-0.037724029272794724,
0.020701637491583824,
0.0032565428409725428,
-0.014128311537206173,
0.014031331054866314,
-0.03154066577553749,
-0.08467844128608704,
0.03956114128232002,
-0.005841855891048908,
0.08926664292812347,
0.007913407869637012,
0.0543811209499836,
-0.0387243889272213,
0.017792250961065292,
-0.04748521000146866,
0.029253894463181496,
-0.010051884688436985,
0.028483007103204727,
0.021613720804452896,
-0.049130551517009735,
-0.0555279366672039,
-0.044153645634651184,
0.05877246335148811,
-0.043726734817028046,
-0.01186537928879261,
-0.042556196451187134,
-0.08749834448099136,
0.023379379883408546,
-0.04067987948656082,
-0.05739511176943779,
-0.013336395844817162,
-0.008628897368907928,
-0.04400644078850746,
-0.054496102035045624,
0.033357810229063034,
0.010659396648406982,
0.0425289049744606,
-0.07672739773988724,
0.011803684756159782,
0.016149528324604034,
-0.010445593856275082,
0.013043100945651531,
0.040911830961704254,
0.04404081776738167,
0.016021165996789932,
0.11331614851951599,
-0.014734968543052673,
0.09593141824007034,
-0.04326897859573364,
0.04968870431184769,
-0.06214668229222298,
0.06872384250164032,
0.08249109238386154,
0.015316973440349102,
-0.11204495280981064,
0.052424341440200806,
-0.06026090309023857,
-0.07068706303834915,
0.09463923424482346,
-0.05359113961458206,
0.003548351116478443,
-0.11254986375570297,
-0.03465818613767624,
0.05595209077000618,
-0.004704988095909357,
-0.021335456520318985,
-0.026284698396921158,
-0.02807890996336937,
0.01417144201695919,
-0.0642252042889595,
0.07052011042833328,
-0.07946424186229706,
-0.008738226257264614,
-0.004736389499157667,
-0.1095309928059578,
-0.03037262335419655,
-0.03640218451619148,
0.013589320704340935,
0.05443209782242775,
-3.788163890665243e-33,
-0.04944048076868057,
0.04742294177412987,
-0.050399769097566605,
0.053217485547065735,
-0.008425980806350708,
-0.0809055045247078,
-0.03418542444705963,
-0.007556313183158636,
-0.04209728538990021,
-0.004492863081395626,
-0.0034973956644535065,
-0.02580920420587063,
-0.0553644634783268,
-0.05198917165398598,
-0.0016936997417360544,
0.0013795148115605116,
-0.01619436778128147,
-0.00807225238531828,
-0.00471862219274044,
-0.0012513602850958705,
0.018015103414654732,
0.0858449637889862,
-0.04964320734143257,
-0.014985394664108753,
-0.03622594103217125,
0.029809700325131416,
-0.049070265144109726,
0.0567026287317276,
0.011385524645447731,
0.06554361432790756,
-0.01865873485803604,
-0.09243496507406235,
0.04038156196475029,
-0.05070579797029495,
-0.04917887970805168,
-0.03462795168161392,
0.037800777703523636,
-0.0723845511674881,
-0.017694054171442986,
0.013701441697776318,
0.03226961940526962,
-0.07072082906961441,
-0.04632388800382614,
0.03866397589445114,
-0.0343703031539917,
0.04942057654261589,
0.014189678244292736,
-0.019141459837555885,
0.04647437483072281,
-0.032924048602581024,
-0.0265242587774992,
-0.025176404044032097,
-0.03571578115224838,
-0.06934386491775513,
-0.03663083538413048,
-0.021987667307257652,
0.06672456860542297,
0.018173450604081154,
0.08037784695625305,
0.03641221299767494,
-0.09459890425205231,
-0.060446422547101974,
0.05124553665518761,
-0.010036204941570759,
-0.014726074412465096,
-0.05964629724621773,
-0.024059362709522247,
-0.004159394185990095,
0.042046524584293365,
0.01235300861299038,
0.06514531373977661,
0.05152207985520363,
0.01741170324385166,
0.03709009662270546,
0.061859097331762314,
-0.04760187119245529,
0.00852794200181961,
-0.0064209154807031155,
-0.03184749558568001,
-0.02827492542564869,
0.017215074971318245,
-0.08090753108263016,
0.0609334297478199,
0.07445428520441055,
0.09366902709007263,
0.016850095242261887,
0.01677241176366806,
0.12942297756671906,
0.02721867710351944,
0.025618646293878555,
-0.009257416240870953,
0.038151033222675323,
0.007318645715713501,
0.11175845563411713,
-0.011096425354480743,
-4.003204168157026e-8,
-0.06839004158973694,
0.09513948857784271,
0.037259820848703384,
-0.03596636652946472,
0.04814789071679115,
0.02249247021973133,
-0.0008056287770159543,
0.08737748861312866,
-0.025282999500632286,
0.041101690381765366,
0.11682447046041489,
0.06477498263120651,
-0.025699887424707413,
-0.08834768831729889,
0.012706924229860306,
0.12217994779348373,
0.05164133384823799,
-0.011836099438369274,
0.005583787336945534,
0.02498779632151127,
0.08423129469156265,
-0.03362498804926872,
0.07603069394826889,
0.048137422651052475,
0.024945106357336044,
-0.1410207599401474,
-0.029686512425541878,
0.11696931719779968,
0.022519610822200775,
0.00997964944690466,
-0.047324031591415405,
-0.04036900773644447,
0.03658415377140045,
0.00956082995980978,
-0.00876932218670845,
0.04899882152676582,
-0.028906038030982018,
-0.08727332949638367,
0.04313559830188751,
-0.05206619203090668,
-0.020607201382517815,
0.02006804384291172,
-0.02263784036040306,
0.03349694982171059,
0.011587036773562431,
0.04151204973459244,
0.010390954092144966,
-0.07934297621250153,
-0.044767383486032486,
0.04698467254638672,
-0.06857982277870178,
-0.005802071187645197,
-0.0030328226275742054,
0.05295878276228905,
-0.022408824414014816,
0.011929369531571865,
0.0071556298062205315,
-0.04208076745271683,
-0.006398188415914774,
0.026140939444303513,
-0.03306344524025917,
0.043075546622276306,
-0.024258175864815712,
-0.009532066062092781
] | 0 | 0 | 0 | 0 |
|
Learning Continually by Spectral Regularization
|
https://openreview.net/forum?id=Hcb2cgPbMg
|
[
"Alex Lewandowski",
"Michał Bortkiewicz",
"Saurabh Kumar",
"András György",
"Dale Schuurmans",
"Mateusz Ostaszewski",
"Marlos C. Machado"
] |
Poster
|
Loss of plasticity is a phenomenon where neural networks can become more difficult to train over the course of learning. Continual learning algorithms seek to mitigate this effect by sustaining good performance while maintaining network trainability. We develop a new technique for improving continual learning inspired by the observation that the singular values of the neural network parameters at initialization are an important factor for trainability during early phases of learning. From this perspective, we derive a new spectral regularizer for continual learning that better sustains these beneficial initialization properties throughout training. In particular, the regularizer keeps the maximum singular value of each layer close to one. Spectral regularization directly ensures that gradient diversity is maintained throughout training, which promotes continual trainability, while minimally interfering with performance in a single task. We present an experimental analysis that shows how the proposed spectral regularizer can sustain trainability and performance across a range of model architectures in continual supervised and reinforcement learning settings. Spectral regularization is less sensitive to hyperparameters while demonstrating better training in individual tasks, sustaining trainability as new tasks arrive, and achieving better generalization performance..
|
plasticity, neural networks, spectral regularization, continual learning
| null | 11,857 |
2406.06811
|
[
-0.028749674558639526,
-0.051979199051856995,
0.052639544010162354,
0.08488098531961441,
0.0323740653693676,
0.06910358369350433,
-0.0012339740060269833,
-0.13653986155986786,
-0.035484716296195984,
-0.08196063339710236,
-0.01965000107884407,
0.02059549279510975,
-0.017466751858592033,
0.03682803362607956,
-0.0520063154399395,
0.014105520211160183,
0.061904873698949814,
0.06117373704910278,
-0.10032150894403458,
-0.011469750665128231,
0.01983286440372467,
0.01745809242129326,
0.016534430906176567,
0.0529540479183197,
0.0028012744151055813,
-0.0014133232180029154,
-0.03390471264719963,
-0.05917131528258324,
0.120441734790802,
-0.05090022459626198,
0.06529561430215836,
-0.04584955796599388,
-0.07245009392499924,
0.03952851518988609,
-0.08562779426574707,
0.03092295490205288,
-0.039302099496126175,
0.06510160118341446,
-0.013665243051946163,
0.013842104002833366,
0.04781905189156532,
0.029509179294109344,
-0.019582998007535934,
0.05786404758691788,
0.017464101314544678,
0.006170519627630711,
0.03992303088307381,
-0.12136498093605042,
0.04482384771108627,
-0.023265838623046875,
-0.02075435221195221,
-0.04286208376288414,
0.022401094436645508,
0.037900160998106,
0.08417109400033951,
0.013021185994148254,
0.02024771086871624,
0.06132962927222252,
-0.06485006213188171,
-0.0018419325351715088,
-0.015337777324020863,
-0.09239771217107773,
0.0018554144771769643,
-0.07510654628276825,
0.01905490830540657,
-0.004546945448964834,
0.03164944797754288,
0.016516732051968575,
0.06886705011129379,
0.06065100431442261,
-0.030440617352724075,
0.09648466110229492,
-0.06617608666419983,
0.06844070553779602,
0.07449560612440109,
0.05602861940860748,
-0.013444693759083748,
-0.04185277596116066,
0.018687862902879715,
0.04153365269303322,
0.009089582599699497,
-0.014578677713871002,
0.0002864766283892095,
-0.0720444917678833,
0.09649667888879776,
0.01714632287621498,
0.031480077654123306,
-0.005181569140404463,
0.04588519409298897,
-0.08345207571983337,
-0.0431220680475235,
0.01678112894296646,
-0.028338279575109482,
-0.035380467772483826,
-0.012466014362871647,
0.0034569031558930874,
0.011409532278776169,
-0.1180848553776741,
-0.00477661332115531,
0.09395486861467361,
0.022447902709245682,
-0.03410391882061958,
-0.0417805053293705,
0.11407873779535294,
-0.000471633393317461,
-0.02244551107287407,
0.012528462335467339,
0.08236932754516602,
0.0134226493537426,
-0.010974555276334286,
0.06926760822534561,
0.013784406706690788,
-0.03940858319401741,
0.04180433601140976,
0.016087491065263748,
-0.001180847641080618,
0.02210463024675846,
0.003470309544354677,
0.01988624408841133,
0.1373174637556076,
0.0029386812821030617,
-0.01395674142986536,
-0.01692461222410202,
-0.0041183545254170895,
0.024250870570540428,
-0.0426495335996151,
-0.040535781532526016,
3.413766001192467e-33,
0.036619722843170166,
0.04935783892869949,
-0.014319753274321556,
-0.029262492433190346,
0.008259647525846958,
-0.03590952605009079,
-0.0032172834035009146,
0.05735747516155243,
0.043695978820323944,
-0.06252376735210419,
-0.023683344945311546,
-0.0027483832091093063,
-0.07177627086639404,
0.10610552877187729,
0.06675076484680176,
-0.05976448208093643,
-0.012677808292210102,
0.021884800866246223,
0.012609112076461315,
-0.16674581170082092,
0.035536106675863266,
0.02269333228468895,
0.002875949488952756,
-0.05646338686347008,
0.007223650813102722,
-0.03132403641939163,
0.1088799461722374,
0.006883098743855953,
-0.028925247490406036,
0.011861526407301426,
-0.02985827624797821,
0.021974626928567886,
0.005995899438858032,
0.05934661999344826,
0.031704485416412354,
-0.0019941101782023907,
0.04983953386545181,
0.02891133353114128,
0.08180971443653107,
-0.06868314743041992,
-0.05749959871172905,
0.03146745637059212,
0.011198258027434349,
-0.0006153166759759188,
-0.034698132425546646,
-0.14505955576896667,
0.006238754838705063,
0.004270058590918779,
-0.12045235931873322,
-0.07488687336444855,
-0.0409727543592453,
-0.0033870108891278505,
-0.017328158020973206,
-0.04052364081144333,
-0.08704420179128647,
-0.0028897044248878956,
-0.003946601413190365,
-0.011796107515692711,
-0.018346676602959633,
-0.0051626479253172874,
0.015541085973381996,
0.002415555529296398,
-0.024972129613161087,
0.08291839063167572,
0.010584804229438305,
0.06404589116573334,
0.02825639769434929,
0.047322019934654236,
0.01596369780600071,
-0.022915935143828392,
-0.0031244538258761168,
-0.044141899794340134,
0.04463465139269829,
-0.1326485425233841,
0.07756756991147995,
-0.038480762392282486,
0.03508308157324791,
-0.09406707435846329,
-0.03650263696908951,
0.05127706006169319,
-0.01353553868830204,
0.036672476679086685,
-0.03140712529420853,
-0.04139011353254318,
-0.02645033225417137,
-0.031944695860147476,
0.0609302781522274,
-0.11506292223930359,
0.044029269367456436,
-0.0073015522211790085,
-0.03421244025230408,
0.004613271448761225,
0.023140674456954002,
0.0039597866125404835,
-0.014467169530689716,
-4.1950873414832866e-33,
0.023695189505815506,
0.04514293745160103,
-0.005222449079155922,
0.05383320152759552,
0.04251756891608238,
-0.02417946606874466,
-0.06322546303272247,
0.08311668038368225,
-0.11425062268972397,
-0.09047084301710129,
0.006756324786692858,
-0.02055937424302101,
-0.019679469987750053,
0.031029032543301582,
-0.0615009069442749,
-0.009115878492593765,
-0.05432267114520073,
0.036361657083034515,
-0.008463125675916672,
0.03525391221046448,
-0.018699385225772858,
0.052399665117263794,
-0.08472087979316711,
0.04984375834465027,
0.022635765373706818,
0.011156774125993252,
-0.06639363616704941,
0.08352839946746826,
0.03778485953807831,
0.0013744729803875089,
-0.05724342539906502,
-0.04424184188246727,
-0.010504016652703285,
-0.08993235975503922,
0.02477460727095604,
0.09397386014461517,
0.01810949295759201,
0.06262405216693878,
-0.05451017990708351,
0.14783543348312378,
0.03065330721437931,
-0.08868329226970673,
0.005145153496414423,
0.009543432854115963,
0.05034240707755089,
-0.10692236572504044,
-0.0385618731379509,
-0.006812420208007097,
-0.07250719517469406,
0.006444606930017471,
0.030014827847480774,
0.008500839583575726,
-0.10491862893104553,
-0.012278052046895027,
-0.059418629854917526,
-0.02205697074532509,
0.01266475673764944,
0.03139989450573921,
0.043357931077480316,
0.016420841217041016,
-0.068902388215065,
-0.04653896391391754,
0.005276277661323547,
0.008261565119028091,
-0.00988858938217163,
-0.011007437482476234,
-0.00037424679612740874,
0.0068566151894629,
0.07313887029886246,
0.014921520836651325,
0.019238600507378578,
0.059332940727472305,
-0.024075130000710487,
0.03174252808094025,
-0.02223283424973488,
-0.03422224149107933,
-0.06160693243145943,
-0.06678806245326996,
-0.053929708898067474,
0.0614844374358654,
0.004935665987432003,
0.011182229034602642,
-0.041531212627887726,
0.010107116773724556,
0.07601102441549301,
0.07080511748790741,
-0.005660554859787226,
0.09435929358005524,
-0.0053976732306182384,
-0.043690841645002365,
0.030867023393511772,
-0.009898007847368717,
-0.02834773063659668,
-0.010111339390277863,
-0.03916016221046448,
-4.259581842802618e-8,
-0.12133369594812393,
0.05019167810678482,
0.012596153654158115,
0.04380042105913162,
0.07870727777481079,
-0.062114227563142776,
0.05450653284788132,
0.022655731067061424,
-0.021857433021068573,
-0.038985032588243484,
0.09296006709337234,
0.014897312968969345,
-0.013306008651852608,
-0.04292364791035652,
-0.029386037960648537,
0.10514945536851883,
0.031041394919157028,
-0.013476689346134663,
0.010466999374330044,
-0.009582177735865116,
-0.014906917698681355,
0.0006932854885235429,
0.022777646780014038,
-0.08404554426670074,
-0.01017661951482296,
-0.06983929872512817,
-0.014897488057613373,
0.12887084484100342,
0.04123905673623085,
0.06168501079082489,
0.02969719097018242,
0.006050148047506809,
0.020008457824587822,
0.02554217167198658,
-0.04945562779903412,
0.12122848629951477,
0.051725614815950394,
-0.06073424592614174,
-0.047572288662195206,
-0.059419047087430954,
-0.02836737222969532,
0.08430591225624084,
-0.013857012614607811,
0.04696555808186531,
0.02149290218949318,
0.0016650145407766104,
-0.003710238728672266,
-0.055175382643938065,
-0.028097443282604218,
-0.07606306672096252,
0.021137457340955734,
-0.029557226225733757,
-0.01609523594379425,
0.06978511810302734,
0.05451000854372978,
-0.036034274846315384,
-0.029213080182671547,
-0.05806968733668327,
-0.05801357701420784,
0.0706903487443924,
0.031820762902498245,
0.011944392696022987,
-0.03879930451512337,
0.02330445684492588
] | 0 | 0 | 0 | 0 |
|
No Equations Needed: Learning System Dynamics Without Relying on Closed-Form ODEs
|
https://openreview.net/forum?id=kbm6tsICar
|
[
"Krzysztof Kacprzyk",
"Mihaela van der Schaar"
] |
Poster
|
Data-driven modeling of dynamical systems is a crucial area of machine learning. In many scenarios, a thorough understanding of the model’s behavior becomes essential for practical applications. For instance, understanding the behavior of a pharmacokinetic model, constructed as part of drug development, may allow us to both verify its biological plausibility (e.g., the drug concentration curve is non-negative and decays to zero in the long term) and to design dosing guidelines (e.g., by looking at the peak concentration and its timing). Discovery of closed-form ordinary differential equations (ODEs) can be employed to obtain such insights by finding a compact mathematical equation and then analyzing it (a two-step approach). However, its widespread use is currently hindered because the analysis process may be time-consuming, requiring substantial mathematical expertise, or even impossible if the equation is too complex. Moreover, if the found equation's behavior does not satisfy the requirements, editing it or influencing the discovery algorithms to rectify it is challenging as the link between the symbolic form of an ODE and its behavior can be elusive. This paper proposes a conceptual shift to modeling low-dimensional dynamical systems by departing from the traditional two-step modeling process. Instead of first discovering a closed-form equation and then analyzing it, our approach, direct semantic modeling, predicts the semantic representation of the dynamical system (i.e., description of its behavior) directly from data, bypassing the need for complex post-hoc analysis. This direct approach also allows the incorporation of intuitive inductive biases into the optimization algorithm and editing the model's behavior directly, ensuring that the model meets the desired specifications. Our approach not only simplifies the modeling pipeline but also enhances the transparency and flexibility of the resulting models compared to traditional closed-form ODEs.
|
dynamical systems, interpretability, ordinary differential equations, symbolic regression
| null | 11,850 |
2501.18563
|
[
-0.0597316175699234,
-0.05881910398602486,
-0.018516264855861664,
0.050497934222221375,
-0.028784193098545074,
-0.04395228996872902,
-0.05797500163316727,
0.09885226935148239,
0.08495692908763885,
0.03685202822089195,
-0.005419956985861063,
0.02276875078678131,
0.03887191787362099,
0.06496395170688629,
-0.04123443365097046,
-0.005932047963142395,
-0.03064928762614727,
0.02655184641480446,
-0.02958720736205578,
-0.02006085030734539,
0.05775070935487747,
0.014408670365810394,
-0.05878394469618797,
0.06680712103843689,
-0.07644392549991608,
0.07913524657487869,
0.03910503908991814,
-0.004461868200451136,
-0.0590033121407032,
-0.0381302572786808,
0.07897114753723145,
0.01483225729316473,
-0.003208981128409505,
-0.07115219533443451,
-0.03072073683142662,
0.05809180438518524,
-0.023272482678294182,
0.03094315715134144,
0.0025730091147124767,
-0.013596928678452969,
-0.019138183444738388,
0.030915115028619766,
0.018532466143369675,
0.06667108088731766,
0.1370854675769806,
-0.08562273532152176,
-0.03526782616972923,
-0.06002431735396385,
-0.055713966488838196,
-0.008916211314499378,
-0.07387223094701767,
-0.0419219546020031,
-0.048337846994400024,
0.027003837749361992,
0.07976078242063522,
-0.05079983174800873,
-0.0028852096293121576,
-0.023974783718585968,
-0.06523758172988892,
0.017992883920669556,
-0.04386370629072189,
0.006378812249749899,
0.0070829554460942745,
-0.02106732875108719,
-0.012987880036234856,
0.02513962797820568,
0.009807656519114971,
0.07084667682647705,
0.05375993624329567,
0.046164605766534805,
-0.09964089095592499,
0.046281684190034866,
-0.04033433273434639,
-0.015908196568489075,
-0.006220399867743254,
-0.06465499103069305,
0.022965330630540848,
0.09386630356311798,
0.0035061442758888006,
-0.0496397539973259,
0.000677779724355787,
-0.03068123757839203,
-0.0784604549407959,
0.025356467813253403,
-0.0022286195307970047,
0.0075696492567658424,
0.01450535748153925,
0.06968865543603897,
0.1128152459859848,
-0.004822672810405493,
0.04426030069589615,
0.06191650778055191,
-0.029299551621079445,
-0.05972843989729881,
0.10586053878068924,
0.04454175755381584,
0.03306853771209717,
-0.004406460095196962,
0.007857083342969418,
0.08328981697559357,
-0.021429287269711494,
0.059467416256666183,
-0.006916824262589216,
0.023081397637724876,
0.065678671002388,
0.00923776812851429,
0.01560181938111782,
0.02862454392015934,
0.04594019427895546,
0.004327026661485434,
0.004944231826812029,
0.0014229943044483662,
-0.01324254646897316,
0.08844724297523499,
0.05625873804092407,
0.03571325168013573,
0.03680195286870003,
-0.08774882555007935,
-0.013452110812067986,
0.019026441499590874,
0.001847736188210547,
-0.042144645005464554,
-0.05821697786450386,
-0.007813701406121254,
0.025156857445836067,
0.05146941915154457,
-0.048503465950489044,
-6.90421647838678e-34,
-0.05261300131678581,
-0.047110140323638916,
0.08010049164295197,
0.04915839806199074,
0.017795870080590248,
-0.08677077293395996,
-0.02194908820092678,
-0.006036325357854366,
0.05100483447313309,
0.06252668052911758,
-0.0645815059542656,
0.024651581421494484,
-0.0446486659348011,
0.06358789652585983,
-0.015848200768232346,
-0.020023411139845848,
0.031295835971832275,
0.017673693597316742,
-0.021511901170015335,
-0.08797811716794968,
0.05028657615184784,
-0.09579908847808838,
-0.051459599286317825,
-0.03848159685730934,
-0.06591033190488815,
0.09543977677822113,
-0.016847966238856316,
0.07224690914154053,
0.022334661334753036,
-0.012968982569873333,
-0.008777864277362823,
0.021880604326725006,
-0.07974684238433838,
0.06422549486160278,
0.009856672957539558,
-0.0147877037525177,
0.0030661888886243105,
-0.022637510672211647,
0.05371068790555,
-0.008335964754223824,
-0.030551036819815636,
-0.022333597764372826,
-0.01163231860846281,
0.05490046367049217,
-0.0188986174762249,
-0.005473964847624302,
-0.02547888085246086,
-0.0231151282787323,
-0.04670754820108414,
0.009419337846338749,
-0.04010148346424103,
-0.03168820962309837,
-0.007654303684830666,
-0.09832660853862762,
-0.08671212941408157,
0.10698310285806656,
-0.02781037800014019,
-0.012798648327589035,
-0.03103627823293209,
0.08128009736537933,
-0.0041642701253294945,
-0.00732914824038744,
0.01755499839782715,
-0.016376180574297905,
0.02255675382912159,
0.10578830540180206,
-0.04990016296505928,
0.0065492866560816765,
0.02724648453295231,
-0.0003820767451543361,
0.0483514778316021,
0.06909332424402237,
-0.003363188588991761,
-0.04212420433759689,
0.11475183069705963,
-0.07049384713172913,
0.1048734039068222,
-0.0012052359525114298,
-0.000655856856610626,
0.0020409997086972,
0.02662459947168827,
-0.09110162407159805,
-0.023525329306721687,
0.036829642951488495,
-0.022399423643946648,
-0.059478964656591415,
0.0020860685035586357,
-0.04432189464569092,
-0.09857656061649323,
-0.051263391971588135,
-0.13691852986812592,
0.017980601638555527,
-0.07856065034866333,
0.03737799450755119,
0.029739849269390106,
-1.912802060211451e-33,
-0.010636596940457821,
-0.0006632678559981287,
-0.0033393357880413532,
-0.017724737524986267,
0.01958472467958927,
-0.0029611955396831036,
-0.07860242575407028,
0.01212976686656475,
0.0566614605486393,
-0.09776214510202408,
0.024032622575759888,
0.06963193416595459,
-0.02290385030210018,
0.03847062960267067,
-0.01768183335661888,
0.08016402274370193,
-0.03922149911522865,
-0.016850460320711136,
0.04681660607457161,
0.05065879225730896,
-0.09785307198762894,
0.01765270158648491,
-0.09867984056472778,
-0.05740824341773987,
0.02735021710395813,
0.031181495636701584,
-0.07071361690759659,
0.011201622895896435,
-0.020000163465738297,
0.05786750465631485,
-0.04704493656754494,
-0.044783782213926315,
-0.038601599633693695,
-0.032284315675497055,
-0.03663518652319908,
0.09995651245117188,
-0.017798515036702156,
-0.029299726709723473,
0.006866832729429007,
-0.011381334625184536,
0.010489760898053646,
0.000355472817318514,
-0.06721162796020508,
-0.07417438179254532,
0.09763403236865997,
-0.060787633061409,
-0.022029738873243332,
0.09336291998624802,
0.07405216991901398,
0.018187684938311577,
0.061620354652404785,
0.014976570382714272,
0.024633800610899925,
-0.010886219330132008,
-0.056035369634628296,
0.0796433836221695,
0.00960434228181839,
-0.014971313066780567,
0.007824813015758991,
-0.038650766015052795,
-0.04246528446674347,
-0.09809251129627228,
-0.008478924632072449,
-0.022067902609705925,
-0.10118129849433899,
-0.01903826929628849,
-0.03099069744348526,
0.0267485324293375,
0.007317773997783661,
-0.022145923227071762,
0.040176309645175934,
0.03654060885310173,
0.04515397176146507,
-0.00026826842804439366,
0.028808826580643654,
-0.12657679617404938,
-0.09757763147354126,
-0.07108257710933685,
-0.05950120463967323,
-0.03889060392975807,
0.041629474610090256,
0.05688716843724251,
0.010474400594830513,
-0.05101102963089943,
-0.07829306274652481,
-0.03261034935712814,
0.024470968171954155,
0.024371378123760223,
0.035171207040548325,
-0.04460173845291138,
-0.05129220336675644,
0.02888350374996662,
-0.0953066498041153,
0.12287648767232895,
-0.06232886388897896,
-5.341356512644779e-8,
0.004308354575186968,
0.027969585731625557,
0.08524267375469208,
-0.0019678170792758465,
0.15040168166160583,
0.09850268810987473,
-0.06032959744334221,
-0.015299246646463871,
-0.007735396735370159,
0.01805063709616661,
0.03879347816109657,
0.06525985896587372,
-0.015555838122963905,
0.005737117491662502,
-0.09735828638076782,
0.1233425885438919,
0.03580412268638611,
-0.0161379873752594,
-0.04539366438984871,
-0.0066983806900680065,
0.03048674575984478,
-0.017677009105682373,
-0.029517367482185364,
-0.016521546989679337,
0.14463816583156586,
-0.0975877121090889,
-0.037397369742393494,
-0.019931187853217125,
0.009583950974047184,
0.02412923239171505,
0.009075786918401718,
0.03879443183541298,
0.05269536003470421,
0.027882708236575127,
0.05176807940006256,
0.06072324141860008,
0.03933517634868622,
-0.06619533896446228,
-0.04338541254401207,
0.0027186416555196047,
-0.04886199161410332,
0.06596510857343674,
-0.039987221360206604,
0.003484770655632019,
0.05905219912528992,
0.034831345081329346,
0.003979375585913658,
-0.0025537745095789433,
0.005115716252475977,
0.04364136978983879,
0.07445578277111053,
0.07046093791723251,
-0.06678555905818939,
-0.0017462365794926882,
-0.009369844570755959,
0.0074761672876775265,
-0.03125154599547386,
0.008868617936968803,
-0.008140687830746174,
-0.053384311497211456,
0.0002833449107129127,
0.03876817971467972,
0.06105219945311546,
-0.0011975099332630634
] |
https://github.com/krzysztof-kacprzyk/SemanticODE
| 1 | 0 | 0 | 0 |
An Undetectable Watermark for Generative Image Models
|
https://openreview.net/forum?id=jlhBFm7T2J
|
[
"Sam Gunn",
"Xuandong Zhao",
"Dawn Song"
] |
Poster
|
We present the first undetectable watermarking scheme for generative image models.
_Undetectability_ ensures that no efficient adversary can distinguish between watermarked and un-watermarked images, even after making many adaptive queries.
In particular, an undetectable watermark does not degrade image quality under any efficiently computable metric.
Our scheme works by selecting the initial latents of a diffusion model using a pseudorandom error-correcting code (Christ and Gunn, 2024), a strategy which guarantees undetectability and robustness. We experimentally demonstrate that our watermarks are quality-preserving and robust using Stable Diffusion 2.1.
Our experiments verify that, in contrast to _every prior scheme_ we tested, our watermark does not degrade image quality.
Our experiments also demonstrate robustness: existing watermark removal attacks fail to remove our watermark from images without significantly degrading the quality of the images.
Finally, we find that we can robustly encode 512 bits in our watermark, and up to 2500 bits when the images are not subjected to watermark removal attacks.
Our code is available at https://github.com/XuandongZhao/PRC-Watermark.
|
Watermarking, AI Safety, Diffusion Models, Generative AI
|
We present a robust image watermarking scheme that is provably quality-preserving, and demonstrate that all prior schemes degrade quality.
| 11,846 |
2410.07369
|
[
-0.07812514901161194,
-0.025483379140496254,
0.054169945418834686,
0.0750853568315506,
0.10177506506443024,
-0.07583455741405487,
-0.02389627695083618,
-0.09562473744153976,
-0.009026313200592995,
-0.017688969150185585,
-0.0038840125780552626,
0.013846218585968018,
0.12097177654504776,
0.011630332097411156,
-0.07071595638990402,
-0.03169430419802666,
0.0056077465415000916,
0.11327306181192398,
-0.07127971202135086,
-0.0681738555431366,
-0.038071077316999435,
-0.06335365027189255,
-0.00038771459367126226,
-0.0402459055185318,
0.04528697207570076,
0.013429786078631878,
0.04357965663075447,
-0.04393106326460838,
0.022319722920656204,
-0.04909593611955643,
0.08248002082109451,
-0.0831422358751297,
-0.009155732579529285,
-0.0014835491310805082,
-0.02048461139202118,
0.05404428765177727,
0.010977853089571,
-0.03817739337682724,
0.05726954713463783,
-0.05856754630804062,
0.002555175917223096,
0.03629327192902565,
-0.06413225084543228,
-0.00710736308246851,
0.0023818309418857098,
0.02252216637134552,
0.038875121623277664,
0.05769410356879234,
-0.04331045225262642,
-0.023831700906157494,
0.03740885108709335,
-0.009186271578073502,
-0.09465890377759933,
0.02496984973549843,
-0.028868667781352997,
-0.06933842599391937,
-0.0038220814894884825,
-0.021092869341373444,
-0.040179185569286346,
0.02678685076534748,
-0.029038045555353165,
-0.03952840343117714,
-0.013335220515727997,
-0.006284639239311218,
0.06262991577386856,
-0.0005405431147664785,
0.061589185148477554,
0.010739435441792011,
0.03293907642364502,
0.04228745773434639,
0.012382499873638153,
0.07008370012044907,
-0.05380292981863022,
0.05300471931695938,
-0.024171186611056328,
0.07421551644802094,
0.0057930611073970795,
-0.012048245407640934,
0.012892510741949081,
-0.054919276386499405,
0.047106873244047165,
-0.0959845632314682,
0.08387397229671478,
0.027142612263560295,
0.10046983510255814,
-0.0571119487285614,
-0.03352274000644684,
0.023969726637005806,
0.054635174572467804,
-0.05003221333026886,
-0.06825758516788483,
0.07209473848342896,
-0.004483328666538,
0.0402505025267601,
0.0750170350074768,
0.012728610076010227,
0.031783170998096466,
-0.009024987928569317,
-0.013708784244954586,
0.0822601169347763,
-0.0715697854757309,
-0.042066484689712524,
-0.021471699699759483,
-0.06243116781115532,
0.15871167182922363,
0.03737878426909447,
0.0659211277961731,
0.05882301926612854,
0.05221942067146301,
-0.015434661880135536,
0.024253318086266518,
0.03531528636813164,
-0.024761255830526352,
-0.0006524116033688188,
0.017319444566965103,
-0.05836585536599159,
-0.03427045792341232,
0.021720582619309425,
-0.0200343020260334,
-0.024453578516840935,
0.07706783711910248,
0.008460377342998981,
-0.047647472470998764,
-0.04789802432060242,
-0.015974141657352448,
-0.05659274756908417,
-0.012665907852351665,
3.0086166284892037e-33,
-0.001928633777424693,
-0.00027530506486073136,
-0.01586970128118992,
-0.01929216831922531,
0.07728685438632965,
0.06971154361963272,
-0.028481556102633476,
-0.05004709213972092,
-0.05599207431077957,
-0.08242475986480713,
-0.009941394440829754,
-0.01625039242208004,
-0.1113038882613182,
0.12093256413936615,
0.0956624373793602,
0.004610794596374035,
-0.0014092643978074193,
-0.0033599224407225847,
0.039108168333768845,
0.006214553490281105,
0.05568060278892517,
-0.05020177736878395,
-0.07287281006574631,
-0.05431756377220154,
-0.01311819814145565,
0.010090089403092861,
0.10869608074426651,
-0.0651860162615776,
0.0776159018278122,
0.02485814318060875,
-0.1273268461227417,
-0.021944504231214523,
0.052978258579969406,
0.027163788676261902,
0.03628797456622124,
0.0127903763204813,
-0.013015097938477993,
0.0031522498466074467,
-0.02025415562093258,
0.013470317237079144,
0.0350925549864769,
0.08015847951173782,
-0.03609723225235939,
-0.021568724885582924,
-0.12572023272514343,
-0.01813509687781334,
-0.012564962729811668,
0.0027477729599922895,
-0.033650968223810196,
-0.03584727644920349,
0.041133128106594086,
-0.0036837703082710505,
-0.03001970611512661,
-0.08451731503009796,
-0.08970072120428085,
-0.052479296922683716,
0.047400541603565216,
-0.06712043285369873,
0.007784062065184116,
0.02275075577199459,
0.05562923103570938,
0.06004922837018967,
0.014298615045845509,
0.02367085963487625,
0.011940008029341698,
-0.02303582802414894,
0.017261359840631485,
0.012874709442257881,
-0.040246374905109406,
0.005051574669778347,
-0.04425675421953201,
0.022820310667157173,
0.0034672485198825598,
-0.02274065650999546,
-0.02207099460065365,
-0.05068459361791611,
0.1081278994679451,
-0.011597192846238613,
-0.0161220021545887,
-0.03019707091152668,
-0.05715961381793022,
-0.052222270518541336,
0.07174666225910187,
-0.03299478068947792,
-0.05890190973877907,
-0.04101511463522911,
-0.02841881848871708,
-0.03857053071260452,
-0.1269548237323761,
-0.033183127641677856,
0.02579279989004135,
-0.026351336389780045,
0.08463481068611145,
-0.023971721529960632,
-0.02811843529343605,
-2.8308003327242855e-33,
-0.043793946504592896,
0.09342411905527115,
0.008933371864259243,
0.11406777799129486,
-0.030783474445343018,
-0.04284248128533363,
0.020204948261380196,
0.07172863930463791,
0.05298465117812157,
-0.07219317555427551,
-0.0720989927649498,
0.05378519371151924,
-0.026823343709111214,
-0.054051391780376434,
-0.025846436619758606,
-0.060289766639471054,
0.0025055177975445986,
0.016493970528244972,
-0.05640408396720886,
0.01498168334364891,
0.037707794457674026,
0.042149826884269714,
-0.01646994985640049,
0.1032356396317482,
-0.038114018738269806,
0.11498789489269257,
-0.03907785192131996,
0.004495421424508095,
-0.00008637900464236736,
-0.023723512887954712,
-0.03311190754175186,
0.02733915112912655,
0.061479534953832626,
-0.10787860304117203,
-0.03757400065660477,
0.030535398051142693,
0.03595340996980667,
0.038097575306892395,
-0.05664524435997009,
0.10493983328342438,
0.0432308167219162,
-0.03053143434226513,
-0.04184572026133537,
0.06730838865041733,
0.04333906248211861,
-0.030541958287358284,
0.010862531140446663,
0.027320988476276398,
0.04073746129870415,
0.022689329460263252,
0.022742250934243202,
-0.01732007786631584,
-0.06571492552757263,
0.07616104930639267,
-0.0014941056724637747,
-0.034469809383153915,
-0.08164694160223007,
-0.005563411395996809,
0.05859644338488579,
0.051647670567035675,
-0.026423590257763863,
-0.06073460355401039,
-0.026994075626134872,
-0.025158969685435295,
0.008393494412302971,
-0.01321447268128395,
-0.052480220794677734,
0.011795640923082829,
-0.006900313310325146,
0.06207895651459694,
0.04598947986960411,
-0.027692211791872978,
-0.029540972784161568,
0.05082240700721741,
0.08225782960653305,
-0.03201496601104736,
-0.0038432932924479246,
0.03740062937140465,
0.002066223183646798,
0.0370357446372509,
0.03717717155814171,
0.02376178279519081,
0.02739100158214569,
-0.020518377423286438,
0.09210015088319778,
-0.020118366926908493,
-0.008172995410859585,
-0.09882495552301407,
-0.012227739207446575,
0.0906478688120842,
-0.003941873554140329,
0.011182201094925404,
0.018584007397294044,
0.025115324184298515,
-0.03667557239532471,
-4.111314666488397e-8,
-0.028210444375872612,
0.014701088890433311,
-0.011901306919753551,
-0.07520173490047455,
-0.0343925841152668,
-0.07331759482622147,
-0.009109602309763432,
0.07428745180368423,
0.04665789008140564,
-0.15482841432094574,
0.07281269878149033,
-0.02552116848528385,
-0.10166589915752411,
-0.08477938920259476,
-0.041471049189567566,
0.07297755032777786,
-0.032667629420757294,
0.030102627351880074,
-0.06214049831032753,
0.0004276269464753568,
-0.10327082127332687,
0.042387038469314575,
-0.05756593868136406,
-0.05924243479967117,
0.02565569244325161,
0.03478291630744934,
0.06762352585792542,
0.029145272448658943,
0.025662722066044807,
0.07391109317541122,
0.015437205322086811,
0.009260057471692562,
0.06098080798983574,
0.04390689358115196,
-0.013250317424535751,
0.0756646916270256,
-0.002693308051675558,
0.01683078706264496,
-0.05369081348180771,
-0.06162392720580101,
0.009829944930970669,
0.04308747127652168,
0.008725244551897049,
-0.03564746677875519,
0.03736620023846626,
-0.01072694268077612,
0.11139634996652603,
-0.0011370463762432337,
-0.021909723058342934,
0.03873615339398384,
0.02996656857430935,
-0.019971901550889015,
-0.057545676827430725,
0.053864430636167526,
0.014039277099072933,
-0.069880910217762,
0.03588201478123665,
-0.0036767360288649797,
0.0836939662694931,
0.04567105695605278,
0.0844927504658699,
-0.02268141321837902,
0.04973568022251129,
0.035097744315862656
] |
https://github.com/xuandongzhao/prc-watermark
| 42 | 0 | 0 | 0 |
Teaching LLMs How to Learn with Contextual Fine-Tuning
|
https://openreview.net/forum?id=FS2nukC2jv
|
[
"Younwoo Choi",
"Muhammad Adil Asif",
"Ziwen Han",
"John Willes",
"Rahul Krishnan"
] |
Poster
|
Prompting Large Language Models (LLMs), or providing context on the expected model of operation, is an effective way to steer the outputs of such models to satisfy human desiderata after they have been trained. But in rapidly evolving domains, there is often need to fine-tune LLMs to improve either the kind of knowledge in their memory or their abilities to perform open ended reasoning in new domains. When human's learn new concepts, we often do so by linking the new material that we are studying to concepts we have already learned before. To that end, we ask, "can prompting help us teach LLMs how to learn". In this work, we study a novel generalization of instruction tuning, called contextual fine-tuning, to fine-tune LLMs. Our method leverages instructional prompts designed to mimic human cognitive strategies in learning and problem-solving to guide the learning process during training, aiming to improve the model’s interpretation and understanding of domain-specific knowledge. We empirically demonstrate that this simple yet effective modification improves the ability of LLMs to be fine-tuned rapidly on new datasets both within the medical and financial domains.
|
Large Language Models
| null | 11,845 |
2503.09032
|
[
0.05605442076921463,
-0.04052911326289177,
0.029649920761585236,
0.01763034239411354,
0.01383167877793312,
-0.08661170303821564,
0.06745465844869614,
0.037813857197761536,
0.008116764947772026,
0.0017270420212298632,
-0.05801998823881149,
0.0052306088618934155,
0.03573622182011604,
0.01698705554008484,
0.05650290101766586,
-0.01974652148783207,
0.05364641174674034,
0.056605689227581024,
-0.11282193660736084,
-0.00651105772703886,
0.05309700220823288,
0.03061872161924839,
0.10802170634269714,
0.013151475228369236,
-0.06705301254987717,
-0.0056844437494874,
-0.0005552618531510234,
-0.04588215425610542,
0.0675494521856308,
-0.014397050254046917,
0.049936290830373764,
0.045535456389188766,
-0.027166277170181274,
0.026297086849808693,
-0.03820754960179329,
0.056681424379348755,
-0.04795406758785248,
0.10167638957500458,
-0.018709519878029823,
-0.021880079060792923,
-0.03122503124177456,
-0.029264453798532486,
0.02199395000934601,
0.037524785846471786,
0.06661700457334518,
-0.03065335936844349,
0.009846117347478867,
-0.004590312484651804,
0.018750496208667755,
0.022484371438622475,
-0.13877083361148834,
-0.051636260002851486,
-0.08728793263435364,
-0.02596728876233101,
0.007196532096713781,
0.06835747510194778,
0.05191719904541969,
0.03330760821700096,
-0.06500062346458435,
-0.054955389350652695,
-0.06705780327320099,
-0.1307101547718048,
-0.07569043338298798,
0.00614828011021018,
-0.06225713714957237,
-0.04147417098283768,
0.00346365780569613,
0.10503625124692917,
-0.01079631969332695,
0.029003432020545006,
-0.0419110432267189,
0.04116319492459297,
0.001917541492730379,
0.0818382278084755,
0.029662324115633965,
0.07322763651609421,
0.01031320821493864,
-0.010771216824650764,
0.03225328400731087,
-0.04624263942241669,
0.00721242418512702,
-0.01875603012740612,
0.006662623956799507,
-0.0002200242888648063,
0.08657464385032654,
0.029983604326844215,
0.04320259392261505,
0.03312981501221657,
0.03989459201693535,
-0.007239106576889753,
0.028807736933231354,
-0.05541171133518219,
-0.0036116030532866716,
0.0074442308396101,
0.05253947153687477,
0.08816897124052048,
0.027419662103056908,
-0.11863768100738525,
-0.040051646530628204,
0.06606247276067734,
0.03803021460771561,
0.10274828970432281,
-0.011079059913754463,
-0.04992416128516197,
-0.05821498483419418,
-0.007667705416679382,
-0.006687223445624113,
0.10797849297523499,
0.05845273286104202,
-0.12619777023792267,
0.03830321133136749,
0.09832525253295898,
0.040156010538339615,
0.026334647089242935,
-0.017952430993318558,
-0.008469698019325733,
-0.0023576286621391773,
-0.06684605032205582,
0.026481635868549347,
0.04569217190146446,
0.0021021857392042875,
0.06687968224287033,
-0.03942492604255676,
-0.04182397201657295,
0.028377575799822807,
-0.06562218815088272,
-0.07438188046216965,
-3.859879423476468e-34,
0.09893219918012619,
0.05615317076444626,
0.06116660684347153,
0.09715691208839417,
0.03720800206065178,
0.02743024006485939,
0.005799924023449421,
0.03846903517842293,
-0.027373313903808594,
0.07638005167245865,
0.024687688797712326,
0.021765664219856262,
-0.07587198913097382,
0.03714807331562042,
0.053653109818696976,
-0.0047532739117741585,
-0.10567566752433777,
0.021259013563394547,
-0.028403732925653458,
-0.005326475948095322,
0.030976930633187294,
-0.04252248629927635,
0.00333461188711226,
-0.08505839854478836,
0.03999793529510498,
0.042701080441474915,
0.05702681094408035,
-0.03323052078485489,
0.013027934357523918,
-0.0014326238306239247,
-0.11796524375677109,
-0.008109087124466896,
-0.018007876351475716,
0.0490027591586113,
-0.009082434698939323,
-0.04345007985830307,
-0.0076734041795134544,
-0.04075106233358383,
0.03963073343038559,
-0.05268217250704765,
0.03987438976764679,
0.014676651917397976,
0.053647030144929886,
0.08501408249139786,
-0.0610417015850544,
-0.06387554109096527,
-0.0005044104764238,
0.001414892147295177,
-0.07425181567668915,
-0.02681640535593033,
0.00867518875747919,
-0.01767474040389061,
-0.031102245673537254,
-0.10261889547109604,
-0.02728077955543995,
0.035151928663253784,
-0.006185102742165327,
0.01019657589495182,
-0.06508708000183105,
0.008645891211926937,
0.032502178102731705,
0.020816326141357422,
-0.014949844218790531,
0.10233337432146072,
-0.01666894368827343,
0.020213570445775986,
-0.03336580842733383,
0.10346193611621857,
0.031709037721157074,
-0.08297430723905563,
-0.06653659790754318,
-0.0035059379879385233,
-0.00219109863974154,
-0.026487892493605614,
0.0164019837975502,
-0.014045124873518944,
0.033912792801856995,
-0.0879187136888504,
0.030526766553521156,
-0.019465701654553413,
0.1562761515378952,
0.02282966859638691,
-0.03736410662531853,
0.016907652840018272,
-0.0050259847193956375,
-0.04081825539469719,
0.018193727359175682,
-0.08730266988277435,
-0.01473679393529892,
-0.03498932346701622,
-0.06669038534164429,
-0.03469196707010269,
-0.06847790628671646,
0.013444473966956139,
-0.012776895426213741,
-9.266561557486253e-34,
-0.007873279973864555,
-0.07109685242176056,
-0.029697123914957047,
0.07048455625772476,
0.027718566358089447,
-0.02947648987174034,
0.038380060344934464,
0.017194854095578194,
-0.031869783997535706,
-0.11472655832767487,
-0.047740887850522995,
-0.045261602848768234,
-0.038202375173568726,
0.018068116158246994,
-0.05256430804729462,
-0.06824656575918198,
-0.04170147329568863,
0.054763130843639374,
0.008808222599327564,
0.05540644749999046,
-0.05084475129842758,
0.07589681446552277,
-0.08451560884714127,
0.03776751458644867,
-0.011735782958567142,
0.02683679759502411,
-0.03902651369571686,
0.07991790771484375,
-0.04627751559019089,
0.07838775217533112,
-0.034461479634046555,
-0.06609053909778595,
0.07209590077400208,
-0.004086419008672237,
-0.055493421852588654,
0.048880435526371,
0.07853808999061584,
-0.0625990554690361,
-0.040694642812013626,
0.06515444070100784,
0.05951159819960594,
-0.03602692112326622,
-0.02231856994330883,
-0.01848144270479679,
0.006205335725098848,
-0.051461465656757355,
-0.059620097279548645,
-0.023544248193502426,
0.032524473965168,
0.011011198163032532,
0.006800017785280943,
-0.035173479467630386,
-0.02383151836693287,
-0.06635654717683792,
-0.015798334032297134,
-0.04523942992091179,
0.03260548785328865,
-0.10993820428848267,
0.05549087002873421,
-0.04622326418757439,
-0.07852209359407425,
0.003502256004139781,
0.027560370042920113,
-0.04828602075576782,
-0.023161398246884346,
-0.022298628464341164,
-0.02476891502737999,
-0.017128586769104004,
-0.023233206942677498,
0.022787757217884064,
-0.023778824135661125,
-0.027283482253551483,
-0.03476469963788986,
-0.07753886282444,
0.052043553441762924,
0.01971714198589325,
0.0012098134029656649,
-0.09230512380599976,
-0.05360368266701698,
-0.046400055289268494,
0.05726245045661926,
-0.08453840762376785,
-0.007789183873683214,
0.08587770164012909,
-0.012589452788233757,
0.028837574645876884,
0.08607952296733856,
0.14699088037014008,
0.00012084306945325807,
-0.005381572525948286,
-0.06307035684585571,
0.03948274254798889,
-0.02388252504169941,
0.03539877012372017,
-0.09880810976028442,
-4.5168654594363034e-8,
-0.06386326253414154,
0.04689287766814232,
0.024261288344860077,
0.09229955822229385,
0.06375979632139206,
-0.00518407765775919,
-0.0741112232208252,
0.08928830921649933,
-0.023131996393203735,
-0.02112908475100994,
0.028501834720373154,
0.03454858437180519,
-0.020525719970464706,
-0.016052300110459328,
0.05508206784725189,
0.07373516261577606,
0.054467324167490005,
0.01587715931236744,
-0.03976743668317795,
-0.046774305403232574,
0.05815811827778816,
0.006155326496809721,
0.025128187611699104,
0.015941863879561424,
0.07033839821815491,
-0.092316173017025,
-0.016241000965237617,
0.14218325912952423,
0.06745539605617523,
0.006867470685392618,
0.009392289444804192,
-0.032283034175634384,
-0.01418555248528719,
-0.007246504537761211,
0.012066175229847431,
0.006070994306355715,
0.026746908202767372,
-0.0587017796933651,
0.02267942950129509,
-0.07103465497493744,
-0.0118094552308321,
0.0685957595705986,
-0.006957490462809801,
0.09142804145812988,
-0.013246322982013226,
-0.030525706708431244,
-0.022154081612825394,
-0.028579972684383392,
-0.024469053372740746,
-0.007675115950405598,
-0.007024378515779972,
0.01994612067937851,
0.0013445550575852394,
-0.009430737234652042,
0.03122558631002903,
0.07391661405563354,
-0.01606370322406292,
0.011329004541039467,
-0.04806026071310043,
0.03167015686631203,
-0.007786672096699476,
0.06555324792861938,
-0.050948355346918106,
0.02246963046491146
] | 0 | 0 | 0 | 0 |
|
Modeling dynamic social vision highlights gaps between deep learning and humans
|
https://openreview.net/forum?id=wAXsx2MYgV
|
[
"Kathy Garcia",
"Emalie McMahon",
"Colin Conwell",
"Michael Bonner",
"Leyla Isik"
] |
Poster
|
Deep learning models trained on computer vision tasks are widely considered the most successful models of human vision to date. The majority of work that supports this idea evaluates how accurately these models predict behavior and brain responses to static images of objects and scenes. Real-world vision, however, is highly dynamic, and far less work has evaluated deep learning models on human responses to moving stimuli, especially those that involve more complicated, higher-order phenomena like social interactions. Here, we extend a dataset of natural videos depicting complex multi-agent interactions by collecting human-annotated sentence captions for each video, and we benchmark 350+ image, video, and language models on behavior and neural responses to the videos. As in prior work, we find that many vision models reach the noise ceiling in predicting visual scene features and responses along the ventral visual stream (often considered the primary neural substrate of object and scene recognition). In contrast, vision models poorly predict human action and social interaction ratings and neural responses in the lateral stream (a neural pathway theorized to specialize in dynamic, social vision), though video models show a striking advantage in predicting mid-level lateral stream regions. Language models (given human sentence captions of the videos) predict action and social ratings better than image and video models, but perform poorly at predicting neural responses in the lateral stream. Together, these results identify a major gap in AI's ability to match human social vision and provide insights to guide future model development for dynamic, natural contexts.
|
NeuroAI, vision, fMRI, deep learning, social perception
|
This study benchmarks 350+ AI models against human behavioral and neural responses to videos of social actions and highlights significant gaps in AI's ability to model dynamic social vision.
| 11,843 | null |
[
-0.011284167878329754,
-0.11801803112030029,
0.06317946314811707,
0.008128993213176727,
0.09130589663982391,
0.02208787202835083,
-0.034553855657577515,
-0.035041823983192444,
0.07523539662361145,
-0.11828859895467758,
-0.06939476728439331,
-0.11831246316432953,
-0.03472595289349556,
0.11458241939544678,
-0.02779814600944519,
-0.058066509664058685,
0.08468011766672134,
0.04494715481996536,
-0.09627046436071396,
-0.004524630028754473,
0.030092094093561172,
-0.014500449411571026,
0.0683365911245346,
0.010758564807474613,
0.010629639029502869,
-0.004687230568379164,
0.0014868805883452296,
-0.04165326803922653,
0.05127108097076416,
-0.04977506771683693,
0.06819016486406326,
-0.03679655119776726,
0.006364851724356413,
0.09072113037109375,
-0.08202996850013733,
0.0834740400314331,
-0.023923588916659355,
0.004947036970406771,
0.022547977045178413,
-0.030014917254447937,
-0.013693037442862988,
-0.0058516631834208965,
0.07448964565992355,
0.014214796014130116,
0.04872924089431763,
0.0676669105887413,
0.007031810469925404,
-0.10142417997121811,
-0.02835894003510475,
-0.0038138709496706724,
-0.11038517951965332,
0.0026278540026396513,
0.05038921907544136,
0.10443268716335297,
-0.041406769305467606,
0.01783384010195732,
0.05250788480043411,
-0.017788589000701904,
0.08119002729654312,
-0.026743607595562935,
0.04864223301410675,
-0.1360982209444046,
0.07853823900222778,
-0.022571634501218796,
-0.05360257625579834,
0.033755216747522354,
-0.02879115752875805,
0.014089240692555904,
-0.05504214018583298,
-0.00947478599846363,
-0.02853197045624256,
0.0491270013153553,
0.013401155360043049,
-0.03745655715465546,
0.0021575370337814093,
0.00966413039714098,
0.036963850259780884,
-0.08414682000875473,
0.0673377588391304,
-0.07911030203104019,
0.04456523805856705,
-0.023409167304635048,
0.03950770944356918,
-0.016114000231027603,
0.13346385955810547,
-0.014712338335812092,
-0.013180200010538101,
0.007615470793098211,
-0.07499223947525024,
0.016002288088202477,
-0.10204166918992996,
-0.029344933107495308,
0.000825017224997282,
-0.06591932475566864,
0.03275765851140022,
0.04781807214021683,
-0.04385073110461235,
-0.11639261990785599,
-0.011055766604840755,
0.04498416930437088,
-0.0039743478409945965,
0.044432803988456726,
0.014338193461298943,
0.009936776012182236,
0.05733189731836319,
0.04224589467048645,
0.07866412401199341,
0.03655746579170227,
0.007032007910311222,
-0.039215151220560074,
-0.03249790146946907,
0.0060407682321965694,
-0.03507174178957939,
0.06571172922849655,
0.031143566593527794,
-0.08102579414844513,
-0.016846759244799614,
-0.008443600498139858,
-0.019942665472626686,
0.007171636912971735,
0.03875969722867012,
0.020528890192508698,
0.021716170012950897,
-0.019628673791885376,
0.05795019492506981,
-0.03410671278834343,
-0.023129740729928017,
1.0983993701526155e-33,
0.02216796763241291,
-0.03744426742196083,
0.057363927364349365,
0.052200306206941605,
0.08083386719226837,
-0.028244419023394585,
-0.06110117584466934,
0.03481366112828255,
-0.041082099080085754,
0.007732466794550419,
-0.04132596775889397,
-0.047830209136009216,
-0.02898062951862812,
0.14376351237297058,
-0.0156883355230093,
0.025678854435682297,
-0.03482406586408615,
-0.03621679171919823,
-0.036495961248874664,
0.03306921571493149,
0.004494888707995415,
0.046478334814310074,
-0.005399726331233978,
-0.025867922231554985,
0.03103618696331978,
-0.009464452974498272,
0.04427493363618851,
-0.08628205209970474,
0.02751256711781025,
-0.006611877121031284,
-0.09361227601766586,
0.0651874840259552,
-0.003976012580096722,
0.04378519579768181,
0.07412555813789368,
-0.03078477829694748,
-0.005631912965327501,
0.008022112771868706,
0.0634855106472969,
-0.0026350996922701597,
-0.01906416192650795,
0.06228068843483925,
-0.058003511279821396,
-0.043744057416915894,
-0.01326452475041151,
-0.04030323028564453,
0.044067222625017166,
0.0978052169084549,
-0.0900518000125885,
0.006845547817647457,
0.04301414638757706,
0.03441620245575905,
-0.027795353904366493,
-0.052700262516736984,
-0.006927086506038904,
0.022865356877446175,
0.05240403860807419,
0.018938397988677025,
-0.03854818269610405,
-0.026739267632365227,
0.01599792391061783,
0.013213373720645905,
0.008307593874633312,
0.03842690587043762,
0.08420931547880173,
0.025645149871706963,
0.0037053267005831003,
0.10030806809663773,
0.0317002534866333,
-0.020025931298732758,
0.04170585423707962,
0.1102992594242096,
-0.026123641058802605,
-0.05917176231741905,
0.04585525020956993,
0.03787580877542496,
0.07869236916303635,
-0.11673689633607864,
-0.00030353633337654173,
0.05570905655622482,
-0.0316762737929821,
-0.06652870029211044,
0.0364021360874176,
-0.020840592682361603,
-0.02681753598153591,
0.007737726904451847,
0.025124600157141685,
-0.10797172784805298,
0.022500723600387573,
0.025699786841869354,
-0.03237350285053253,
-0.04291531816124916,
-0.026223639026284218,
0.012517831288278103,
-0.03792247921228409,
-1.0558427093975062e-33,
-0.022780435159802437,
-0.0007629706524312496,
-0.06801729649305344,
0.0032220338471233845,
-0.034851010888814926,
-0.0728902816772461,
-0.017004525288939476,
0.04686092957854271,
0.06748289614915848,
-0.0118558956310153,
0.00829237885773182,
-0.06860233098268509,
-0.005605942104011774,
0.00016673712525516748,
0.07273095846176147,
-0.0596039742231369,
0.04346609488129616,
-0.008389852941036224,
-0.004861293360590935,
-0.03524957224726677,
0.03205801919102669,
0.08934566378593445,
-0.007658233866095543,
0.00464239064604044,
-0.036475542932748795,
0.03491930663585663,
-0.012877540662884712,
0.03564846143126488,
-0.032928310334682465,
-0.007769898045808077,
-0.030158953741192818,
0.0038253609091043472,
-0.044725529849529266,
-0.020276011899113655,
-0.05365101248025894,
0.06250110268592834,
-0.037539042532444,
-0.08202573657035828,
-0.033308979123830795,
0.01599946990609169,
0.12454002350568771,
-0.048239950090646744,
-0.046993374824523926,
-0.007041882257908583,
-0.0832330733537674,
-0.018916398286819458,
-0.07112380862236023,
0.00049329933244735,
-0.035118430852890015,
0.06251765042543411,
-0.042608316987752914,
0.037996333092451096,
-0.10024246573448181,
0.03940104693174362,
-0.06795452535152435,
-0.05801374092698097,
0.009694577194750309,
0.018889309838414192,
0.08471916615962982,
0.006557622458785772,
-0.024097034707665443,
0.07912969589233398,
-0.05728223919868469,
-0.010485728271305561,
-0.01587512716650963,
-0.007626269478350878,
-0.10513603687286377,
-0.015545738860964775,
0.0699821189045906,
-0.005790901370346546,
0.076166532933712,
-0.06453505903482437,
0.004340509418398142,
0.07413703948259354,
-0.00315756700001657,
-0.018002469092607498,
0.016018465161323547,
-0.026619378477334976,
0.026839492842555046,
-0.04853600263595581,
0.01934909261763096,
-0.041745852679014206,
-0.039050038903951645,
0.010417947545647621,
0.024876827374100685,
0.09945188462734222,
0.00024785997811704874,
0.034081656485795975,
0.006696346681565046,
0.013720565475523472,
-0.03518788143992424,
-0.024263855069875717,
-0.05539972707629204,
-0.0005598338902927935,
-0.09070777893066406,
-3.870446718678977e-8,
-0.10017530620098114,
0.0538691021502018,
0.025455867871642113,
0.0013434295542538166,
0.007403350900858641,
-0.0071515110321342945,
0.020947663113474846,
0.08809321373701096,
-0.025156423449516296,
-0.06450213491916656,
0.07988909631967545,
0.04298153892159462,
0.04739995673298836,
-0.033874958753585815,
-0.03220133110880852,
0.09223653376102448,
0.02621442824602127,
-0.03289862349629402,
0.01814868673682213,
0.01845274679362774,
0.031550660729408264,
-0.0200251005589962,
0.02505415678024292,
0.0004333896504249424,
-0.02312544919550419,
-0.024992363527417183,
-0.05969797447323799,
0.09454626590013504,
-0.0009912492241710424,
-0.030654625967144966,
0.10067770630121231,
0.05836284160614014,
0.006644035689532757,
-0.07211154699325562,
0.027644861489534378,
0.1109858974814415,
0.013327355496585369,
-0.04638155177235603,
0.053088221698999405,
-0.01328563503921032,
0.0009300601668655872,
0.045338474214076996,
0.007377671543508768,
0.01764582097530365,
0.020603764802217484,
0.04550308734178543,
-0.009749140590429306,
-0.11767514050006866,
0.11527029424905777,
0.009401490911841393,
-0.05752383545041084,
-0.009692687541246414,
-0.010669457726180553,
0.15587039291858673,
0.006462123245000839,
-0.02815166302025318,
0.0029733674600720406,
-0.03202930837869644,
0.08107585459947586,
0.041824668645858765,
0.05022645369172096,
0.04129287973046303,
-0.060646917670965195,
-0.04330156370997429
] | 0 | 0 | 0 | 0 |
|
Statistical Tractability of Off-policy Evaluation of History-dependent Policies in POMDPs
|
https://openreview.net/forum?id=Qja5s0K3VX
|
[
"Yuheng Zhang",
"Nan Jiang"
] |
Poster
|
We investigate off-policy evaluation (OPE), a central and fundamental problem
in reinforcement learning (RL), in the challenging setting of Partially Observable
Markov Decision Processes (POMDPs) with large observation spaces. Recent
works of Uehara et al. (2023a); Zhang & Jiang (2024) developed a model-free
framework and identified important coverage assumptions (called belief and outcome coverage) that enable accurate OPE of memoryless policies with polynomial sample complexities, but handling more general target policies that depend on
the entire observable history remained an open problem. In this work, we prove
information-theoretic hardness for model-free OPE of history-dependent policies in
several settings, characterized by additional assumptions imposed on the behavior
policy (memoryless vs. history-dependent) and/or the state-revealing property of
the POMDP (single-step vs. multi-step revealing). We further show that some hardness can be circumvented by a natural model-based algorithm—whose analysis has surprisingly eluded the literature despite the algorithm’s simplicity—demonstrating
provable separation between model-free and model-based OPE in POMDPs.
|
Partially Observable Markov Decision Process; Offline Policy Evaluation; Reinforcement Learning Theory
| null | 11,834 |
2503.01134
|
[
-0.09145521372556686,
0.01007384154945612,
0.030800458043813705,
-0.002190501196309924,
0.11019285768270493,
0.03143110126256943,
-0.011246377602219582,
-0.008437356911599636,
0.01905260607600212,
0.11279356479644775,
-0.06706120818853378,
0.029950372874736786,
0.04743575677275658,
0.04674883186817169,
0.013269736431539059,
0.03127598017454147,
0.023985445499420166,
-0.013278531841933727,
-0.03618430346250534,
-0.030752534046769142,
0.02859281189739704,
0.015030096285045147,
0.008085746318101883,
0.0012176649179309607,
-0.0030059777200222015,
0.00006291660247370601,
0.06495622545480728,
-0.005389798432588577,
0.020092900842428207,
0.05473174899816513,
0.038302674889564514,
-0.08624772727489471,
0.026351574808359146,
-0.11492612212896347,
-0.009046319872140884,
0.04578026011586189,
-0.08289926499128342,
-0.01385896559804678,
-0.06661602109670639,
0.012042095884680748,
0.04714665934443474,
0.042565666139125824,
-0.013110815547406673,
0.01526834536343813,
0.020839685574173927,
-0.006186691112816334,
0.06956086307764053,
-0.05192967504262924,
-0.0707743689417839,
-0.040721435099840164,
0.0296266321092844,
0.005983555223792791,
0.02470357157289982,
-0.053840454667806625,
0.031699180603027344,
-0.03525641933083534,
-0.0042782435193657875,
-0.0028665601275861263,
-0.07994679361581802,
-0.07390391826629639,
-0.0688139796257019,
-0.0836937204003334,
-0.09282723814249039,
-0.020227668806910515,
-0.05905919894576073,
0.04423493891954422,
0.056338462978601456,
0.03048013150691986,
0.026706522330641747,
0.05975102260708809,
-0.06794438511133194,
0.06675641238689423,
-0.0035610953345894814,
-0.023582754656672478,
-0.0016873597633093596,
0.01229812204837799,
0.04792720824480057,
-0.0013435297878459096,
-0.025170007720589638,
-0.036310069262981415,
-0.023580044507980347,
-0.05236847698688507,
0.007020600140094757,
0.021192045882344246,
0.032591380178928375,
-0.055377453565597534,
-0.005700160749256611,
0.03253236785531044,
0.10283749550580978,
-0.0013139898655936122,
-0.014988880604505539,
0.006151551380753517,
-0.011803464964032173,
0.00557157676666975,
0.054734960198402405,
-0.025719836354255676,
-0.000010027403732237872,
-0.056056950241327286,
-0.04036545008420944,
0.09766999632120132,
0.0023120474070310593,
0.06214240938425064,
0.053261786699295044,
-0.04146569222211838,
-0.012271207757294178,
-0.03582802414894104,
-0.03694761544466019,
-0.014142777770757675,
0.020614800974726677,
-0.017455248162150383,
0.030891617760062218,
-0.011615466326475143,
0.09750732779502869,
0.030514858663082123,
-0.02372765727341175,
-0.028372637927532196,
-0.02866915985941887,
-0.012952215038239956,
-0.076062873005867,
-0.005239653401076794,
0.006922246888279915,
-0.08513534069061279,
0.03696660324931145,
0.05881589278578758,
0.008009228855371475,
-0.04109922796487808,
0.033198051154613495,
3.268321074974768e-34,
0.043474745005369186,
-0.019815489649772644,
0.01714370958507061,
-0.01960926502943039,
0.041280873119831085,
0.05630233511328697,
0.0030160564929246902,
-0.0535317026078701,
0.06614022701978683,
0.023715980350971222,
0.037848129868507385,
0.027270864695310593,
-0.04901636019349098,
-0.017527986317873,
0.057625990360975266,
0.03776858001947403,
-0.053549282252788544,
0.11256939172744751,
-0.006210759747773409,
-0.03441441059112549,
0.042676571756601334,
0.007174345199018717,
-0.01449876930564642,
-0.10947778075933456,
0.06312718242406845,
0.013523788191378117,
0.03620369732379913,
-0.009292957372963428,
-0.019556820392608643,
-0.009997849352657795,
0.004650393035262823,
0.03808474540710449,
-0.002354573691263795,
0.08139505237340927,
0.012314504943788052,
0.0023024273104965687,
-0.10012327879667282,
-0.06992954015731812,
-0.006140719633549452,
-0.05593079701066017,
0.023160098120570183,
-0.05301642417907715,
0.06569597870111465,
0.009434843435883522,
-0.07693670690059662,
-0.12330940365791321,
-0.06463874131441116,
-0.09010036289691925,
-0.0357721745967865,
-0.04002704098820686,
0.08375764638185501,
0.0058510797098279,
-0.04361318051815033,
-0.10797925293445587,
-0.09024342149496078,
-0.006529063452035189,
-0.0583527497947216,
0.07566577941179276,
0.0021983031183481216,
0.08623698353767395,
0.04994303733110428,
0.03646215423941612,
0.08521071821451187,
0.025538334622979164,
-0.022304663434624672,
0.091062530875206,
-0.12597735226154327,
0.022001080214977264,
0.08741733431816101,
0.0037736045196652412,
0.032759226858615875,
0.0451689288020134,
0.030821140855550766,
-0.10868912935256958,
0.11614074558019638,
-0.040403421968221664,
0.08757627010345459,
-0.03251487389206886,
0.023356344550848007,
0.009022283367812634,
-0.006263961549848318,
-0.1019158735871315,
-0.03967363387346268,
-0.0088201854377985,
0.012535947374999523,
-0.01774822548031807,
-0.025045601651072502,
0.03234487026929855,
-0.0751226469874382,
-0.051708318293094635,
-0.07444247603416443,
0.0012143413769081235,
-0.012294433079659939,
0.015506141819059849,
0.03599055856466293,
-2.522124883418961e-33,
-0.05735710635781288,
-0.07680726796388626,
0.07193296402692795,
-0.04203610494732857,
-0.022134989500045776,
-0.01712106354534626,
-0.07302908599376678,
-0.053943533450365067,
0.021919015794992447,
-0.07249545305967331,
-0.09050196409225464,
0.028068246319890022,
0.061039332300424576,
-0.022768957540392876,
-0.028166258707642555,
-0.0028458365704864264,
-0.006190305110067129,
-0.013351437635719776,
-0.024981630966067314,
0.035856395959854126,
0.01836157590150833,
0.03623903542757034,
-0.09137596935033798,
0.09208125621080399,
0.0014703868655487895,
0.037892431020736694,
0.013470175676047802,
0.03450390696525574,
0.045076627284288406,
0.0440542958676815,
-0.0844516009092331,
-0.0726681724190712,
-0.0845295712351799,
0.01727142371237278,
-0.07587207108736038,
0.051644217222929,
0.09332071244716644,
0.0603555329144001,
-0.09408913552761078,
0.09116747975349426,
0.019543591886758804,
-0.04462593421339989,
-0.08633681386709213,
-0.020968671888113022,
0.01139826886355877,
-0.033665504306554794,
-0.041502695530653,
0.0568232424557209,
0.043495435267686844,
-0.045983944088220596,
-0.06422670930624008,
0.07494590431451797,
0.010322910733520985,
0.03945036232471466,
-0.0778159573674202,
-0.018274642527103424,
-0.061476342380046844,
-0.006547695491462946,
0.018768124282360077,
0.029438816010951996,
-0.09587139636278152,
0.08113060146570206,
-0.04775296151638031,
0.10627707839012146,
0.018841058015823364,
0.05286700278520584,
0.02883925288915634,
-0.0031820328440517187,
0.0861896499991417,
-0.004049734678119421,
0.01577896624803543,
0.02763652428984642,
-0.07283143699169159,
-0.004120837897062302,
0.1722778081893921,
0.07799185812473297,
-0.0034798607230186462,
-0.004200424067676067,
-0.036519113928079605,
0.022579926997423172,
0.054594431072473526,
-0.019492095336318016,
0.06564708799123764,
-0.006457888055592775,
0.02970912866294384,
0.04241776093840599,
0.05243454873561859,
0.027468334883451462,
0.04721440002322197,
-0.05773202329874039,
-0.037721313536167145,
0.01614898443222046,
0.01805240288376808,
0.06843612343072891,
-0.06881843507289886,
-4.665562514105659e-8,
0.006437791045755148,
-0.011051872745156288,
0.12216092646121979,
0.004352320451289415,
0.06385556608438492,
-0.021996086463332176,
-0.07491116970777512,
-0.0222402885556221,
-0.042324721813201904,
-0.04155454784631729,
0.03956916928291321,
0.079576276242733,
-0.04517315700650215,
-0.04424052685499191,
-0.06138277053833008,
0.021418729797005653,
-0.0006159520125947893,
0.0028325561434030533,
0.0020300676114857197,
-0.007897832430899143,
-0.028492385521531105,
-0.031335361301898956,
-0.03328410163521767,
0.007312229834496975,
0.039856672286987305,
0.008564839139580727,
-0.04817184805870056,
0.05900491401553154,
0.06794857978820801,
0.10198921710252762,
0.04973197355866432,
-0.01117992028594017,
0.04203085973858833,
0.04140491783618927,
0.05626130849123001,
0.03002919815480709,
0.059317708015441895,
-0.0040725041180849075,
-0.0003856232506223023,
-0.09280330687761307,
-0.0178616214543581,
0.07887466996908188,
0.00749070942401886,
0.04392613470554352,
0.006201650016009808,
0.04189711809158325,
-0.06320331990718842,
-0.1361483931541443,
0.003058915724977851,
-0.06371796131134033,
0.05942969396710396,
0.03771437704563141,
-0.030213983729481697,
0.053227443248033524,
-0.0021631286945194006,
0.031328652054071426,
0.029387300834059715,
-0.07786798477172852,
-0.03116699308156967,
0.034242287278175354,
0.04142465069890022,
0.018184904009103775,
-0.006850798614323139,
0.031647246330976486
] | 0 | 0 | 0 | 0 |
|
Large Language Models can Become Strong Self-Detoxifiers
|
https://openreview.net/forum?id=jY5oml9fe9
|
[
"Ching-Yun Ko",
"Pin-Yu Chen",
"Payel Das",
"Youssef Mroueh",
"Soham Dan",
"Georgios Kollias",
"Subhajit Chaudhury",
"Tejaswini Pedapati",
"Luca Daniel"
] |
Poster
|
Reducing the likelihood of generating harmful and toxic output is an essential task when aligning large language models (LLMs). Existing methods mainly rely on training an external reward model (i.e., another language model) or fine-tuning the LLM using self-generated data to influence the outcome. In this paper, we show that LLMs have the capability of self-detoxification without external reward model learning or retraining of the LM. We propose \textit{Self-disciplined Autoregressive Sampling (SASA)}, a lightweight controlled decoding algorithm for toxicity reduction of LLMs. SASA leverages the contextual representations from an LLM to learn linear subspaces from labeled data characterizing toxic v.s. non-toxic output in analytical forms. When auto-completing a response token-by-token, SASA dynamically tracks the margin of the current output to steer the generation away from the toxic subspace, by adjusting the autoregressive sampling strategy. Evaluated on LLMs of different scale and nature, namely Llama-3.1-Instruct (8B), Llama-2 (7B), and GPT2-L models with the RealToxicityPrompts, BOLD, and AttaQ benchmarks, SASA markedly enhances the quality of the generated sentences relative to the original models and attains comparable performance to state-of-the-art detoxification techniques, significantly reducing the toxicity level by only using the LLM's internal representations.
|
detoxification; LLM
| null | 11,831 | null |
[
-0.04019877687096596,
-0.07316159456968307,
0.020175766199827194,
0.03713217377662659,
0.05208168178796768,
0.005293902941048145,
0.049744538962841034,
-0.014824191108345985,
0.09059757739305496,
-0.010890739038586617,
-0.06325051188468933,
-0.04848429933190346,
0.08406458050012589,
-0.018591860309243202,
-0.022084904834628105,
0.023511027917265892,
0.05563916638493538,
0.07208167016506195,
-0.0678214281797409,
-0.03696483373641968,
0.07889838516712189,
0.059775788336992264,
0.043877918273210526,
0.060849521309137344,
0.07455752044916153,
0.012848385609686375,
-0.030615616589784622,
-0.023990362882614136,
0.04326178878545761,
-0.020188692957162857,
0.1001778393983841,
0.02912179008126259,
0.03508249297738075,
-0.06026419252157211,
-0.06249041110277176,
-0.012663734145462513,
-0.10516386479139328,
0.011021412909030914,
0.00761419627815485,
-0.04908376559615135,
-0.03050924278795719,
0.04498274251818657,
0.01469702459871769,
0.004874958656728268,
0.08395560830831528,
-0.06904448568820953,
-0.0276426300406456,
-0.08337820321321487,
-0.00989286694675684,
0.02610268071293831,
-0.02903044782578945,
0.026943985372781754,
-0.06096577271819115,
0.09644230455160141,
-0.026007357984781265,
-0.10114984214305878,
0.05254252254962921,
0.02315378561615944,
-0.024472128599882126,
0.0186197217553854,
-0.012076693587005138,
-0.09144941717386246,
-0.014207115396857262,
-0.003645031014457345,
0.024461137130856514,
0.010809441097080708,
0.0038453233428299427,
0.15383239090442657,
-0.04677118360996246,
0.019626906141638756,
-0.053672969341278076,
-0.020305631682276726,
-0.004744453355669975,
-0.00437732320278883,
0.002411959692835808,
0.10032124072313309,
-0.03028680384159088,
-0.041586268693208694,
0.10700305551290512,
0.02167356386780739,
0.002907454501837492,
0.06107316166162491,
0.09916861355304718,
0.045975979417562485,
0.023587362840771675,
0.009853475727140903,
0.010793102905154228,
0.0051817214116454124,
0.10860238969326019,
0.033744510263204575,
-0.00011638110299827531,
-0.00016837986186146736,
0.06849313527345657,
0.02225816249847412,
-0.02288859896361828,
-0.002171954372897744,
0.00600552698597312,
-0.054831959307193756,
0.002579960273578763,
0.10216309875249863,
0.03546994552016258,
0.08968356996774673,
-0.06039285287261009,
-0.05836818739771843,
-0.00040519211324863136,
-0.035622488707304,
0.0016867900267243385,
0.042691249400377274,
0.11290743201971054,
-0.062381502240896225,
-0.0067533524706959724,
0.07538457959890366,
-0.03692974895238876,
0.03903578221797943,
0.035978980362415314,
0.04409889504313469,
0.06312040239572525,
-0.09419869631528854,
0.05778936296701431,
0.050472699105739594,
-0.06125286594033241,
-0.03504427149891853,
-0.012047268450260162,
0.012455497868359089,
0.05202561989426613,
-0.03089183382689953,
-0.034094277769327164,
1.2456386701297458e-33,
0.04615079611539841,
0.07167546451091766,
0.011699855327606201,
0.03200984001159668,
0.01895337551832199,
0.004906394053250551,
0.012711219489574432,
0.046696510165929794,
0.025873221457004547,
0.009007873013615608,
-0.019691335037350655,
0.019298838451504707,
-0.031399164348840714,
0.0746435672044754,
0.06878150254487991,
-0.009554142132401466,
-0.04708458483219147,
0.0034477540757507086,
-0.011035998351871967,
-0.11859474331140518,
0.03253903239965439,
-0.03181124106049538,
0.03080153837800026,
-0.10509159415960312,
0.019028039649128914,
0.0869518518447876,
0.057464826852083206,
0.005702783819288015,
-0.12118946015834808,
0.03093886189162731,
-0.0750139057636261,
-0.009318523108959198,
-0.01680050604045391,
-0.005174464546144009,
0.02066255360841751,
-0.06356432288885117,
-0.04386233910918236,
0.04938243329524994,
-0.014900824055075645,
-0.034598611295223236,
0.01195121556520462,
0.06685163080692291,
0.043612491339445114,
0.010830912739038467,
-0.03314647078514099,
-0.09563399851322174,
-0.04456448182463646,
-0.0403510220348835,
-0.018691400066018105,
-0.06047079339623451,
0.04934777319431305,
-0.00043199717765673995,
0.002543536014854908,
-0.045035868883132935,
-0.0992395281791687,
0.07377202063798904,
0.0027209145482629538,
0.000721480930224061,
0.025913218036293983,
0.0464620366692543,
-0.008122969418764114,
0.02968534454703331,
-0.016745703294873238,
0.0528828501701355,
0.07861177623271942,
0.0498587004840374,
-0.03719337284564972,
-0.0040844520553946495,
0.03805125132203102,
-0.042711291462183,
-0.015109600499272346,
0.02225164696574211,
-0.06115582212805748,
-0.06528298556804657,
0.004322525579482317,
-0.05419708788394928,
0.08329346030950546,
-0.03760790824890137,
-0.03350164741277695,
0.047042880207300186,
0.05451851710677147,
-0.04780206456780434,
-0.061385270208120346,
-0.05768224596977234,
-0.022409377619624138,
-0.08399676531553268,
0.02965780720114708,
-0.14214755594730377,
0.013016069307923317,
-0.019937405362725258,
-0.023937197402119637,
-0.040420494973659515,
-0.010405337437987328,
0.005302037112414837,
-0.05548199638724327,
-5.803701218402685e-34,
-0.013136536814272404,
0.006683853920549154,
-0.018173977732658386,
0.00385083956643939,
0.008700692094862461,
-0.09204506129026413,
0.02372385375201702,
0.08639096468687057,
0.014801162295043468,
-0.00641700578853488,
-0.08036588132381439,
-0.015072112902998924,
0.06459520012140274,
0.05196718871593475,
0.06815372407436371,
-0.020057586953043938,
-0.001036869827657938,
0.04593665897846222,
-0.008060364983975887,
-0.001492409035563469,
-0.040378637611866,
0.08809786289930344,
-0.07462862879037857,
0.020831769332289696,
-0.016864893957972527,
0.04425503686070442,
-0.025126833468675613,
0.049751315265893936,
0.05710337683558464,
-0.008794616907835007,
-0.04276346415281296,
0.06500997394323349,
-0.027910416945815086,
0.003904913552105427,
-0.10949878394603729,
0.04609551653265953,
-0.013281029649078846,
-0.0091879116371274,
-0.007145307492464781,
0.028880532830953598,
0.07385897636413574,
0.010363678447902203,
-0.09540647268295288,
0.03318897262215614,
-0.0015617756871506572,
-0.0953337773680687,
-0.14945700764656067,
-0.03977227956056595,
0.10003326088190079,
-0.0033068787306547165,
-0.005502695683389902,
0.03392963111400604,
-0.1053648367524147,
0.06491117179393768,
0.010327101685106754,
-0.11626461893320084,
-0.04044044390320778,
-0.04543839022517204,
0.028206905350089073,
-0.05626596882939339,
-0.05676420032978058,
-0.019689185544848442,
0.055345822125673294,
-0.08329657465219498,
-0.005192325450479984,
-0.07241035252809525,
-0.023007480427622795,
0.025368046015501022,
-0.0062791286036372185,
-0.044521164149045944,
0.05883035808801651,
0.014846334233880043,
-0.0009098874870687723,
-0.030169976875185966,
-0.04887709394097328,
-0.0462198406457901,
-0.060157179832458496,
-0.08936535567045212,
-0.02200925722718239,
-0.07727637141942978,
0.03836476057767868,
0.026186132803559303,
0.025105470791459084,
0.06971530616283417,
0.0008788375998847187,
-0.02274695783853531,
0.01954793930053711,
0.09609463065862656,
-0.0007250917260535061,
-0.0020559001713991165,
0.01504561584442854,
-0.00865104142576456,
0.002225922653451562,
0.09524983167648315,
-0.02129199728369713,
-4.5134623150033804e-8,
-0.07653243094682693,
-0.018119489774107933,
0.005646880250424147,
0.13422246277332306,
0.0067553226836025715,
0.03940419480204582,
-0.050982601940631866,
0.04314257204532623,
0.001510846195742488,
-0.04816972091794014,
0.0727352574467659,
0.010709968395531178,
-0.0851081907749176,
-0.06735716015100479,
-0.04211484640836716,
0.05238154157996178,
0.02700800634920597,
0.06497157365083694,
-0.020823853090405464,
0.016477735713124275,
0.00754169886931777,
0.026701267808675766,
0.0138466851785779,
-0.060526467859745026,
0.06121627986431122,
-0.0783294215798378,
-0.020921478047966957,
0.03423665836453438,
0.07911834865808487,
-0.0181684922426939,
0.03201908618211746,
0.022460337728261948,
-0.03414866700768471,
-0.006003877613693476,
-0.04871479421854019,
0.06659584492444992,
0.021494893357157707,
-0.06651540100574493,
-0.05485343560576439,
0.023927897214889526,
0.018497714772820473,
0.09672573208808899,
-0.09696616977453232,
0.026904214173555374,
-0.010095926001667976,
0.00396372564136982,
0.03793134167790413,
-0.0870688259601593,
0.07260745018720627,
-0.05070266127586365,
0.07635639607906342,
-0.07220994681119919,
-0.018076734617352486,
0.01418075617402792,
0.037729982286691666,
0.014925455674529076,
-0.02079918421804905,
0.037832506000995636,
0.03337500989437103,
0.03547342121601105,
0.03937826305627823,
0.021574832499027252,
0.003806901630014181,
-0.019845684990286827
] | 0 | 0 | 0 | 0 |
|
STRAP: Robot Sub-Trajectory Retrieval for Augmented Policy Learning
|
https://openreview.net/forum?id=4VHiptx7xe
|
[
"Marius Memmel",
"Jacob Berg",
"Bingqing Chen",
"Abhishek Gupta",
"Jonathan Francis"
] |
Poster
|
Robot learning is witnessing a significant increase in the size, diversity, and complexity of pre-collected datasets, mirroring trends in domains such as natural language processing and computer vision. Many robot learning methods treat such datasets as multi-task expert data and learn a multi-task, generalist policy by training broadly across them. Notably, while these generalist policies can improve the average performance across many tasks, the performance of generalist policies on any one task is often suboptimal due to negative transfer between partitions of the data, compared to task-specific specialist policies. In this work, we argue for the paradigm of training policies during deployment given the scenarios they encounter: rather than deploying pre-trained policies to unseen problems in a zero-shot manner, we non-parametrically retrieve and train models directly on relevant data at test time. Furthermore, we show that many robotics tasks share considerable amounts of low-level behaviors and that retrieval at the "sub"-trajectory granularity enables significantly improved data utilization, generalization, and robustness in adapting policies to novel problems. In contrast, existing full-trajectory retrieval methods tend to underutilize the data and miss out on shared cross-task content. This work proposes STRAP, a technique for leveraging pre-trained vision foundation models and dynamic time warping to retrieve sub-sequences of trajectories from large training corpora in a robust fashion. STRAP outperforms both prior retrieval algorithms and multi-task learning methods in simulated and real experiments, showing the ability to scale to much larger offline datasets in the real world as well as the ability to learn robust control policies with just a handful of real-world demonstrations.
|
dynamic time warping, few-shot imitation learning, retrieval, foundation models
|
Subsequence-DTW for sub-trajectory retrieval to augment few-shot policy learning
| 11,830 |
2412.15182
|
[
0.01128373108804226,
-0.08804477751255035,
0.04889686033129692,
0.017753079533576965,
0.0740862563252449,
-0.015459151938557625,
0.040088508278131485,
-0.03157128766179085,
-0.0477025993168354,
0.022767841815948486,
0.01904892548918724,
-0.037316370755434036,
0.03492802008986473,
0.07644841074943542,
0.04941879212856293,
0.03964993730187416,
0.1162760853767395,
0.040081821382045746,
-0.11306130141019821,
-0.00006837154796812683,
0.027065858244895935,
0.0005884792190045118,
0.08605556935071945,
0.05014032870531082,
-0.04372299835085869,
0.010244281031191349,
0.030649198219180107,
-0.10688547044992447,
-0.045976050198078156,
-0.010192139074206352,
-0.01723775453865528,
-0.056438278406858444,
-0.08596410602331161,
0.01764674298465252,
0.028468791395425797,
0.07824023067951202,
-0.1102982759475708,
0.06465354561805725,
0.0029192117508500814,
-0.010129066184163094,
0.00454350933432579,
0.01913287490606308,
0.03423187881708145,
0.020694775506854057,
0.07819610089063644,
0.08949674665927887,
0.011995486915111542,
-0.10150092840194702,
0.039485324174165726,
-0.030382031574845314,
-0.037825148552656174,
-0.04355350136756897,
0.0379803404211998,
-0.013118762522935867,
-0.0033662570640444756,
0.03017251566052437,
0.032707568258047104,
-0.02787013165652752,
0.04931991174817085,
-0.021528419107198715,
-0.005880627315491438,
-0.12884466350078583,
-0.051480744034051895,
-0.0578402616083622,
-0.03326527029275894,
-0.013981442898511887,
-0.019318604841828346,
-0.015110364183783531,
-0.0016848340164870024,
0.04471633955836296,
-0.05714859813451767,
0.09922119230031967,
0.024133862927556038,
0.023448659107089043,
0.008791955187916756,
0.022195329889655113,
0.01629745215177536,
-0.003952991683036089,
-0.00503375893458724,
-0.10161099582910538,
0.025575513020157814,
-0.015018468722701073,
0.023940756916999817,
0.11860872805118561,
0.15269385278224945,
-0.006964311935007572,
-0.01719004102051258,
0.019328493624925613,
0.03619015961885452,
0.0006264044786803424,
-0.021101579070091248,
-0.03896588459610939,
0.05794139951467514,
-0.046445537358522415,
-0.02192908339202404,
0.0542798787355423,
0.020849835127592087,
-0.07974746823310852,
-0.035539139062166214,
0.07209333777427673,
0.022189361974596977,
0.004351467359811068,
0.024199295789003372,
0.04446679353713989,
-0.04055890440940857,
0.011325573548674583,
0.012800325639545918,
0.013591887429356575,
0.05366130545735359,
-0.02675589546561241,
0.036377158015966415,
-0.032933492213487625,
0.03195139765739441,
-0.03150857239961624,
-0.012105288915336132,
0.03599931672215462,
-0.053158823400735855,
0.005051261745393276,
-0.023018110543489456,
0.12333866208791733,
-0.002764090895652771,
-0.04924718663096428,
0.06981328129768372,
0.0336931012570858,
0.0572347454726696,
0.027280442416667938,
-0.017828255891799927,
7.800704489976513e-34,
0.09789196401834488,
0.002131885848939419,
0.039331164211034775,
-0.007057140581309795,
0.03404128924012184,
-0.024694181978702545,
0.02016780711710453,
0.0258942861109972,
-0.012541989795863628,
-0.014219576492905617,
0.0036965948529541492,
0.03577090799808502,
-0.06594628095626831,
0.0969083309173584,
-0.02354487217962742,
0.01945684477686882,
-0.03094138763844967,
0.05938104912638664,
0.0089567257091403,
0.05539816990494728,
0.028964951634407043,
-0.03523264825344086,
-0.047452662140131,
-0.026839466765522957,
0.05768435820937157,
-0.02799355424940586,
0.047521524131298065,
-0.015374338254332542,
-0.014398513361811638,
0.010576804168522358,
-0.13374833762645721,
0.06104525923728943,
0.024470996111631393,
0.0659131407737732,
0.05376160517334938,
-0.03050130605697632,
-0.029908757656812668,
-0.016933994367718697,
0.0091358358040452,
0.022170361131429672,
0.004944497719407082,
0.013086186721920967,
0.03811167553067207,
-0.058571476489305496,
-0.0756450667977333,
-0.125459685921669,
-0.009273333474993706,
0.052573829889297485,
-0.07462963461875916,
-0.036731041967868805,
0.03303833678364754,
0.013415944762527943,
-0.06296876817941666,
-0.1142675131559372,
-0.0009144282666966319,
-0.047470226883888245,
-0.0063330442644655704,
0.01072325836867094,
0.0003310607571620494,
-0.05172618478536606,
0.061721526086330414,
0.004531677812337875,
0.004471079912036657,
0.10635928064584732,
-0.03087395615875721,
0.04355500638484955,
0.026489412412047386,
0.07615753263235092,
0.13130100071430206,
-0.0022388126235455275,
-0.020702270790934563,
0.006751330103725195,
-0.023353107273578644,
-0.08576863259077072,
0.12028919905424118,
0.015669016167521477,
0.04313096031546593,
-0.09522391855716705,
0.027523312717676163,
0.032146889716386795,
-0.011787742376327515,
-0.05804898962378502,
0.005169131327420473,
-0.022987913340330124,
-0.03950284421443939,
-0.07046183198690414,
0.033122751861810684,
-0.085879385471344,
0.005468146409839392,
-0.0379130095243454,
-0.08292367309331894,
-0.052928224205970764,
-0.17492346465587616,
0.01835956610739231,
-0.02384953945875168,
-1.3101582489685195e-33,
0.053496748208999634,
-0.01012846827507019,
-0.00519167073071003,
0.006419749464839697,
-0.019609712064266205,
0.007337850984185934,
0.009627592749893665,
-0.0460820272564888,
0.01008143275976181,
-0.0474405400454998,
-0.11365029960870743,
-0.05261225625872612,
0.015997450798749924,
-0.004882704932242632,
-0.0006574185681529343,
0.009375877678394318,
-0.06262101978063583,
0.06177089363336563,
0.023424236103892326,
-0.022543665021657944,
0.01092473790049553,
0.05463916063308716,
-0.05331040546298027,
0.03778801113367081,
-0.020971247926354408,
0.009528973139822483,
-0.05619597062468529,
0.04356861114501953,
-0.01909765973687172,
0.04064948111772537,
-0.02970476821064949,
-0.07058365643024445,
0.020993368700146675,
-0.03880000114440918,
-0.0789194330573082,
0.10571689158678055,
0.0082908496260643,
0.03341243043541908,
-0.06054724380373955,
0.09804078936576843,
0.036474235355854034,
-0.1194029375910759,
0.004547120537608862,
0.023461030796170235,
0.003893065731972456,
-0.08781154453754425,
-0.05112066864967346,
0.024257224053144455,
0.013385406695306301,
0.044578712433576584,
-0.08768418431282043,
-0.016137614846229553,
-0.07838112860918045,
-0.012960880063474178,
-0.06346943974494934,
-0.0581432469189167,
-0.03072385862469673,
-0.00953879114240408,
0.031041795387864113,
-0.015555787831544876,
-0.017828170210123062,
-0.010077562183141708,
-0.024975715205073357,
0.06407273560762405,
-0.017470868304371834,
0.0031838363502174616,
-0.02796570397913456,
0.0009090578532777727,
-0.0084304828196764,
0.020063722506165504,
-0.020559996366500854,
-0.024831054732203484,
0.0026234725955873728,
-0.04023832827806473,
0.021966969594359398,
0.0420323982834816,
0.04993855953216553,
-0.027033844962716103,
-0.0042483070865273476,
0.017427967861294746,
0.0032713976688683033,
0.0034994096495211124,
-0.0025721730198711157,
0.11052883416414261,
-0.021966785192489624,
0.12018904834985733,
0.023300547152757645,
0.07682095468044281,
0.011825406923890114,
-0.019142022356390953,
-0.01832433044910431,
-0.01633657142519951,
-0.0481080636382103,
0.03185022622346878,
-0.05342384800314903,
-4.203323555884708e-8,
-0.06311358511447906,
0.10700744390487671,
0.03490559011697769,
0.030617427080869675,
0.030711445957422256,
0.027670906856656075,
-0.06064106151461601,
0.07840824872255325,
-0.04104215279221535,
-0.023339087143540382,
0.05781272053718567,
0.03735363855957985,
-0.021719153970479965,
-0.052732959389686584,
-0.00037218810757622123,
0.04613802954554558,
0.023502133786678314,
0.019155729562044144,
-0.02935667149722576,
0.03253527730703354,
0.007656204979866743,
0.0028618737123906612,
0.03941470757126808,
-0.07509391754865646,
-0.037272222340106964,
-0.03588588908314705,
-0.08140095323324203,
0.05108840763568878,
0.08310566842556,
0.0234354417771101,
-0.0007168951560743153,
-0.04185666888952255,
0.0532243475317955,
-0.052036985754966736,
0.04878171160817146,
0.0387057326734066,
-0.009524539113044739,
-0.0484146811068058,
0.037905409932136536,
-0.057510823011398315,
0.008744745515286922,
0.10590280592441559,
0.055466897785663605,
0.044168028980493546,
0.01228789146989584,
0.010552709922194481,
-0.06720943003892899,
-0.11494111269712448,
0.02635064721107483,
-0.00964427087455988,
-0.016295623034238815,
-0.037388596683740616,
0.003680853871628642,
0.1452372819185257,
0.09319630265235901,
0.003722002264112234,
0.0618324801325798,
-0.10898962616920471,
0.0065566981211304665,
0.09387694299221039,
0.013825030997395515,
0.028442153707146645,
-0.10476820915937424,
0.052752215415239334
] | 0 | 0 | 0 | 0 |
|
PortLLM: Personalizing Evolving Large Language Models with Training-Free and Portable Model Patches
|
https://openreview.net/forum?id=gyHoR6uFhU
|
[
"Rana Shahroz",
"Pingzhi Li",
"Sukwon Yun",
"Zhenyu Wang",
"Shahriar Nirjon",
"Chau-Wai Wong",
"Tianlong Chen"
] |
Poster
|
As large language models (LLMs) increasingly shape the AI landscape, fine-tuning pretrained models has become more popular than in the pre-LLM era for achieving optimal performance in domain-specific tasks. However, pretrained LLMs such as ChatGPT are periodically evolved (i.e., model parameters are frequently updated), making it challenging for downstream users with limited resources to keep up with fine-tuning the newest LLMs for their domain application. Even though fine-tuning costs have nowadays been reduced thanks to the innovations of parameter-efficient fine-tuning such as LoRA, not all downstream users have adequate computing for frequent personalization. Moreover, access to fine-tuning datasets, particularly in sensitive domains such as healthcare, could be time-restrictive, making it crucial to retain the knowledge encoded in earlier fine-tuned rounds for future adaptation. In this paper, we present PORTLLM, a training-free framework that (i) creates an initial lightweight model update patch to capture domain-specific knowledge, and (ii) allows a subsequent seamless plugging for the continual personalization of evolved LLM at minimal cost. Our extensive experiments cover seven representative datasets, from easier question-answering tasks {BoolQ, SST2} to harder reasoning tasks {WinoGrande, GSM8K}, and models including {Mistral-7B,Llama2, Llama3.1, and Gemma2}, validating the portability of our designed model patches and showcasing the effectiveness of our proposed framework. For instance, PORTLLM achieves comparable performance to LoRA fine-tuning with reductions of up to 12.2× in GPU memory usage. Finally, we provide theoretical justifications to understand the portability of our model update patches, which offers new insights into the theoretical dimension of LLMs’ personalization.
|
Large Language Models, NLP, Efficiency, Fine-tuning, Efficient Fine-tuning, Portability
| null | 11,824 |
2410.10870
|
[
-0.0755133256316185,
-0.08754827082157135,
0.016546154394745827,
0.047659631818532944,
0.042363815009593964,
0.00602306192740798,
0.048388779163360596,
0.04725200682878494,
-0.007609877269715071,
-0.062411099672317505,
-0.06441964209079742,
-0.06195949390530586,
0.03677217662334442,
0.011595998890697956,
0.03381829336285591,
-0.002307678572833538,
0.08745789527893066,
0.008326660841703415,
-0.09967470914125443,
-0.036177121102809906,
-0.013540397398173809,
0.06441351771354675,
0.0493595190346241,
0.002958517288789153,
-0.00015137217997107655,
-0.0677485466003418,
0.03097005933523178,
-0.02279515564441681,
0.046277184039354324,
0.0047732992097735405,
0.037084586918354034,
0.0902494415640831,
-0.02003953978419304,
0.04636192321777344,
-0.06686356663703918,
0.004921311512589455,
-0.10865150392055511,
0.047516774386167526,
-0.020873893052339554,
-0.0497669018805027,
-0.03482085093855858,
-0.08061892539262772,
0.02028176560997963,
0.04592490196228027,
0.06734967976808548,
0.024354372173547745,
0.006994063965976238,
-0.007409514393657446,
0.01777992583811283,
0.09110454469919205,
-0.09330311417579651,
-0.1020616665482521,
0.032839518040418625,
0.044433873146772385,
-0.020765217021107674,
0.02037627622485161,
0.014483830891549587,
0.028457393869757652,
-0.05958154425024986,
-0.05312899872660637,
-0.07787609845399857,
-0.09972193837165833,
-0.034147996455430984,
-0.014558147639036179,
-0.07777246832847595,
-0.018183493986725807,
0.012456395663321018,
0.06961306184530258,
0.03268918767571449,
0.024235287681221962,
-0.024136383086442947,
0.1172301173210144,
-0.01182433869689703,
0.043328408151865005,
-0.000041773251723498106,
0.05483824014663696,
-0.055144570767879486,
-0.0297244880348444,
0.050086840987205505,
-0.05845482647418976,
-0.0385344922542572,
-0.0030232619028538465,
0.021804742515087128,
-0.002430237829685211,
0.05200136825442314,
0.0001278273412026465,
0.037573106586933136,
-0.011670430190861225,
0.03556789085268974,
-0.005226915702223778,
0.011479008942842484,
-0.010080812498927116,
0.10612262785434723,
-0.036644767969846725,
0.010658303275704384,
0.0747467651963234,
0.028895001858472824,
-0.04737887904047966,
-0.07351890206336975,
0.09027016907930374,
-0.04031691327691078,
0.08705683052539825,
0.020201222971081734,
-0.0003556687443051487,
-0.022345134988427162,
0.0068415189161896706,
-0.005924963392317295,
0.056052953004837036,
0.011554962955415249,
-0.04432035610079765,
-0.0071136318147182465,
0.07793866842985153,
0.04557372257113457,
-0.049235403537750244,
0.05295538902282715,
-0.0023402017541229725,
-0.012872731313109398,
-0.09017028659582138,
0.010631314478814602,
0.13557039201259613,
0.0012598964385688305,
0.04587854817509651,
-0.0145472576841712,
0.000709070242010057,
-0.019575146958231926,
0.07281719148159027,
-0.044193193316459656,
2.4790588111666632e-33,
0.09861674159765244,
0.08855611830949783,
-0.010949810966849327,
0.05137336254119873,
0.001767966547049582,
-0.04402652382850647,
0.008685505017638206,
0.04087057337164879,
-0.04266414791345596,
-0.0562976635992527,
-0.03257709741592407,
0.0490371510386467,
-0.04290534555912018,
0.0501890555024147,
0.04711022973060608,
0.004546390380710363,
-0.04244629666209221,
0.06890201568603516,
0.022312017157673836,
-0.014459710568189621,
0.06662775576114655,
-0.021037882193922997,
0.007068938575685024,
-0.023295646533370018,
0.057878680527210236,
0.046660415828228,
0.11692816019058228,
-0.033651623874902725,
0.025926539674401283,
-0.0019520239438861609,
-0.04111391305923462,
-0.03275114670395851,
0.056445278227329254,
0.020470354706048965,
0.03269714489579201,
-0.07240983843803406,
-0.009030478075146675,
-0.023941848427057266,
0.02904023975133896,
0.004141547717154026,
0.013642353937029839,
0.008060886524617672,
0.03180459514260292,
-0.02569769136607647,
-0.0705198422074318,
-0.11821509897708893,
-0.019261786714196205,
0.01982341893017292,
-0.03651665151119232,
-0.04142215847969055,
0.08075427263975143,
-0.008142410777509212,
-0.06639783829450607,
0.011334274895489216,
-0.06876358389854431,
-0.04682764783501625,
-0.025237834081053734,
0.02326015755534172,
-0.006533868610858917,
-0.024134887382388115,
0.033736731857061386,
0.023382795974612236,
0.023631710559129715,
0.097260020673275,
0.02362968400120735,
-0.06188042089343071,
0.020257247611880302,
0.059115830808877945,
0.0417574979364872,
-0.017097948119044304,
-0.007644954603165388,
-0.004507198929786682,
-0.03054129146039486,
-0.03073379583656788,
0.02823183313012123,
0.002441458171233535,
0.05950623005628586,
-0.07771194726228714,
-0.013506903313100338,
-0.02067660540342331,
0.06516321748495102,
0.026790734380483627,
-0.039629947394132614,
-0.028299914672970772,
0.028071042150259018,
-0.07875309139490128,
0.04368867352604866,
-0.06894690543413162,
-0.04382454231381416,
-0.04512305557727814,
-0.04551928862929344,
-0.006380031816661358,
-0.010616065934300423,
0.014266209676861763,
-0.00474900659173727,
-9.877806681299342e-34,
-0.02833549678325653,
-0.06190374121069908,
-0.041949376463890076,
0.06148523837327957,
0.046618957072496414,
-0.0779104083776474,
-0.009160637855529785,
0.06495156139135361,
-0.032613322138786316,
-0.0892932340502739,
-0.08648227900266647,
-0.06044697016477585,
0.029259173199534416,
-0.03882817178964615,
0.022369876503944397,
0.007885468192398548,
-0.04885029047727585,
0.006994171999394894,
-0.022740894928574562,
0.0821039155125618,
-0.036963216960430145,
0.07848046720027924,
-0.15149910748004913,
0.046856846660375595,
-0.04379022493958473,
0.028131701052188873,
-0.041339702904224396,
0.07667727023363113,
0.012814833782613277,
0.006614270154386759,
-0.08359778672456741,
0.01679965667426586,
-0.07161486893892288,
-0.06778267025947571,
-0.07007713615894318,
0.04709133878350258,
0.06693313270807266,
0.009241238236427307,
-0.03975048288702965,
0.12395049631595612,
0.07194951176643372,
-0.03373551368713379,
-0.1004818007349968,
-0.02605537697672844,
-0.026179051026701927,
0.0019151050364598632,
-0.14422976970672607,
-0.047065094113349915,
0.014467509463429451,
-0.017592908814549446,
0.0009484585025347769,
-0.02416190132498741,
0.034822989255189896,
-0.027150651440024376,
-0.014966854825615883,
-0.08700726181268692,
0.030085541307926178,
-0.07014922052621841,
-0.019335774704813957,
0.006102442275732756,
-0.09739838540554047,
-0.031882207840681076,
0.08254685252904892,
-0.08266545832157135,
-0.006377922371029854,
-0.04231511056423187,
-0.011214951053261757,
0.00998785998672247,
-0.052358098328113556,
-0.019585534930229187,
0.01121351309120655,
-0.06194556877017021,
-0.07571429759263992,
-0.02309497818350792,
-0.02431611716747284,
-0.020447248592972755,
0.04052263870835304,
-0.033118780702352524,
0.04664703086018562,
-0.025394106283783913,
0.019406836479902267,
-0.001305346260778606,
0.02941746637225151,
-0.002691269852221012,
0.07672041654586792,
0.049438681453466415,
0.054485466331243515,
0.13420012593269348,
0.028113916516304016,
0.03056555800139904,
-0.032834362238645554,
0.004387860186398029,
-0.0429379977285862,
0.0451139397919178,
-0.07529156655073166,
-4.370917139340236e-8,
-0.05859515070915222,
0.044120557606220245,
-0.007653134874999523,
0.1416819542646408,
0.028156163170933723,
-0.035190433263778687,
-0.038589637726545334,
0.12457051873207092,
0.019416656345129013,
0.02860187739133835,
0.02064783126115799,
0.021971445530653,
0.0053255301900208,
-0.03869708254933357,
0.04298827052116394,
0.11843382567167282,
-0.019163964316248894,
0.03888697922229767,
-0.04667084291577339,
-0.04386257752776146,
0.06103406846523285,
0.06219431385397911,
0.026365192607045174,
-0.0034039095044136047,
0.09861510246992111,
-0.04572567343711853,
-0.014524783939123154,
0.09081390500068665,
0.06274107098579407,
0.057702235877513885,
-0.03468605503439903,
-0.0041374522261321545,
0.006287129130214453,
-0.028404375538229942,
0.021527813747525215,
0.016229581087827682,
-0.02587033621966839,
-0.05130651593208313,
-0.0019245198927819729,
-0.022821135818958282,
0.09961548447608948,
0.06191690266132355,
-0.04890372231602669,
0.013708015903830528,
0.028817428275942802,
-0.050162624567747116,
-0.011160530149936676,
-0.10534675419330597,
0.03701694309711456,
-0.03495508432388306,
-0.036473553627729416,
-0.030375773087143898,
0.06716589629650116,
0.07023077458143234,
0.06956023722887039,
0.05265558138489723,
0.028595484793186188,
0.027534887194633484,
0.11619410663843155,
0.03676225617527962,
-0.0058232685551047325,
-0.014141043648123741,
-0.04520159587264061,
0.06310964375734329
] | 0 | 0 | 0 | 0 |
|
Robust Transfer of Safety-Constrained Reinforcement Learning Agents
|
https://openreview.net/forum?id=rvXdGL4pCJ
|
[
"Markel Zubia",
"Thiago D. Simão",
"Nils Jansen"
] |
Poster
|
Reinforcement learning (RL) often relies on trial and error, which may cause undesirable outcomes. As a result, standard RL is inappropriate for safety-critical applications. To address this issue, one may train a safe agent in a controlled environment (where safety violations are allowed) and then transfer it to the real world (where safety violations may have disastrous consequences). Prior work has made this transfer safe as long as the new environment preserves the safety-related dynamics. However, in most practical applications, differences or shifts in dynamics between the two environments are inevitable, potentially leading to safety violations after the transfer. This work aims to guarantee safety even when the new environment has different (safety-related) dynamics. In other words, we aim to make the process of safe transfer robust. Our methodology (1) robustifies an agent in the controlled environment and (2) provably provides---under mild assumption---a safe transfer to new environments. The empirical evaluation shows that this method yields policies that are robust against changes in dynamics, demonstrating safety after transfer to a new environment.
|
Reinforcement Learning, Safe Transfer, Adversarial Training, Robustness
|
This paper trains agents under action disturbances for a safe and robust transfer between environments with different dynamics, preventing safety violations after the transfer.
| 11,822 | null |
[
-0.057922475039958954,
-0.028803573921322823,
-0.02348119392991066,
0.018426155671477318,
0.09701371937990189,
0.049169402569532394,
0.0044447206892073154,
0.001294113346375525,
-0.026850420981645584,
0.06697157025337219,
-0.011330902576446533,
-0.031200163066387177,
0.07781984657049179,
0.060157522559165955,
-0.023359743878245354,
-0.010974686592817307,
0.07373982667922974,
0.07166452705860138,
-0.1261582374572754,
0.02629144862294197,
0.004241349641233683,
0.008793101646006107,
0.052501704543828964,
0.052369896322488785,
-0.07887856662273407,
-0.014424342662096024,
-0.018851837143301964,
0.030607808381319046,
0.007415820378810167,
-0.07258101552724838,
0.011202082969248295,
-0.09729725867509842,
-0.07077959179878235,
-0.10070489346981049,
-0.030918318778276443,
0.0726495161652565,
-0.1020781546831131,
-0.01708936132490635,
0.002657720586284995,
0.0034489899408072233,
-0.06294596940279007,
-0.00788918137550354,
-0.03255404904484749,
-0.030617527663707733,
-0.04655797779560089,
0.003967985976487398,
0.012804877012968063,
-0.039956241846084595,
-0.026954054832458496,
-0.05560965836048126,
-0.0366138257086277,
0.009367470629513264,
0.0177315566688776,
0.05043715611100197,
0.028202228248119354,
-0.004718174692243338,
0.08813288062810898,
-0.03417328745126724,
-0.04742632433772087,
-0.026349874213337898,
0.021700605750083923,
-0.04271654412150383,
-0.01839544251561165,
-0.020991025492548943,
-0.0738164409995079,
0.025912297889590263,
-0.12090271711349487,
0.07005160301923752,
0.061461012810468674,
0.044914186000823975,
-0.04886134713888168,
-0.0014281617477536201,
-0.05120575428009033,
-0.06363289803266525,
0.04882512614130974,
0.03184543177485466,
-0.04972046986222267,
0.022807471454143524,
0.027193473652005196,
0.007062152959406376,
0.0646994486451149,
-0.02996985614299774,
-0.03954264894127846,
0.019946686923503876,
0.05363335832953453,
0.01630736142396927,
-0.04306194186210632,
0.03712473064661026,
0.12851659953594208,
0.019034121185541153,
0.039203137159347534,
0.07121783494949341,
0.1086520105600357,
0.035311177372932434,
0.04395922273397446,
-0.027195192873477936,
-0.024417659267783165,
-0.040617071092128754,
-0.003722885623574257,
0.05351807549595833,
0.007416100241243839,
0.034373827278614044,
0.02069055661559105,
0.06926975399255753,
0.04973502829670906,
-0.03300127759575844,
0.01939670741558075,
-0.0625530257821083,
0.03421934321522713,
-0.009897736832499504,
0.0367475263774395,
0.033653050661087036,
0.035864219069480896,
0.06908931583166122,
-0.039042696356773376,
0.04666715860366821,
-0.0893656387925148,
-0.0028378504794090986,
-0.05379170551896095,
0.04158272221684456,
0.08036284148693085,
-0.07404680550098419,
0.11157901585102081,
0.01748000457882881,
0.05222059413790703,
-0.02701260708272457,
0.040083855390548706,
8.58337617940284e-34,
0.06803521513938904,
0.017698660492897034,
-0.035009607672691345,
0.020235002040863037,
0.008139455690979958,
-0.03849300742149353,
-0.03783135488629341,
-0.03467726334929466,
0.006928739137947559,
-0.017609870061278343,
0.020895790308713913,
0.003887125290930271,
-0.025910234078764915,
0.0008627638453617692,
0.01545064989477396,
0.02150627411901951,
-0.050300780683755875,
0.036909498274326324,
0.01498663891106844,
-0.06775788962841034,
0.050162963569164276,
-0.09298990666866302,
-0.0016360660083591938,
-0.05184312164783478,
-0.014935476705431938,
0.0618378221988678,
0.008149215020239353,
0.03420085832476616,
0.07575470209121704,
0.02463742531836033,
-0.04020146280527115,
0.07330764830112457,
-0.04167947173118591,
0.1007254421710968,
-0.0011380539508536458,
0.05392041802406311,
-0.1291850060224533,
-0.052491601556539536,
-0.02250014618039131,
0.032799478620290756,
-0.02837110124528408,
-0.04383053258061409,
0.07838214933872223,
0.05651567503809929,
0.0855644941329956,
-0.035489100962877274,
0.008348705247044563,
-0.020122922956943512,
-0.11295294761657715,
-0.05587054044008255,
-0.06230026111006737,
0.01695946976542473,
0.028888508677482605,
-0.08853115141391754,
-0.08913113921880722,
0.041080329567193985,
0.033287134021520615,
0.06620486080646515,
-0.053108688443899155,
0.02248840220272541,
0.002967023989185691,
0.027704861015081406,
0.015348779037594795,
0.043405406177043915,
0.04221957549452782,
0.055698029696941376,
-0.030405530706048012,
-0.004729988984763622,
0.019764354452490807,
-0.02849259041249752,
-0.012492505833506584,
0.029016006737947464,
-0.009676051326096058,
-0.038512468338012695,
0.017119919881224632,
-0.08812019228935242,
0.06422296911478043,
-0.023670611903071404,
-0.005056207533925772,
-0.0559813566505909,
-0.002962755970656872,
0.027298076078295708,
-0.060705699026584625,
0.0655972883105278,
-0.024113427847623825,
-0.08715692907571793,
-0.0687088742852211,
-0.01105513982474804,
-0.06271356344223022,
0.08178828656673431,
-0.0239794310182333,
-0.03873936086893082,
0.027809174731373787,
0.09377369284629822,
0.058412980288267136,
-1.4680437625087573e-33,
0.00421175267547369,
-0.08240360021591187,
-0.04356444999575615,
0.04456451162695885,
-0.06151802837848663,
-0.02507207915186882,
-0.037897899746894836,
-0.011711470782756805,
0.03166534751653671,
-0.05535207688808441,
-0.10131898522377014,
0.048099443316459656,
0.012370657175779343,
-0.009296744130551815,
-0.04109007120132446,
-0.11644842475652695,
-0.008894405327737331,
0.006795703433454037,
0.002996685216203332,
-0.00011738339526345953,
0.08916227519512177,
0.027060717344284058,
-0.09811437875032425,
0.1097676083445549,
0.007224519737064838,
0.038730017840862274,
-0.04696362838149071,
0.03530098497867584,
0.04650309681892395,
0.06273674964904785,
-0.030857659876346588,
-0.03277336806058884,
-0.006178221665322781,
0.060714274644851685,
-0.03342588245868683,
0.05755091458559036,
0.07581814378499985,
-0.025584427639842033,
-0.056952137500047684,
0.03983674198389053,
0.014938168227672577,
-0.013423921540379524,
-0.034716710448265076,
-0.03500625118613243,
0.07234442979097366,
-0.018943436443805695,
-0.04222869500517845,
-0.014564360491931438,
-0.012849674560129642,
-0.037218235433101654,
-0.01674729958176613,
0.019186770543456078,
-0.05986855924129486,
-0.04749341309070587,
0.03439229354262352,
0.03360098600387573,
0.06748323142528534,
0.040129076689481735,
0.019696813076734543,
0.003395423060283065,
-0.07559031248092651,
0.03387327492237091,
-0.0665917620062828,
0.0966903418302536,
-0.05005887150764465,
0.08109257370233536,
-0.07269664853811264,
-0.006454610265791416,
0.08889368921518326,
0.003942104987800121,
0.09142947196960449,
0.04773489385843277,
-0.007196300197392702,
-0.08138302713632584,
0.0868692696094513,
-0.09792184829711914,
-0.01960049569606781,
-0.057376883924007416,
-0.04261501878499985,
-0.008584026247262955,
0.002176631009206176,
0.002562387613579631,
0.06345707178115845,
0.09446261823177338,
-0.015797434374690056,
0.01673595979809761,
-0.012817232869565487,
-0.003599481424316764,
-0.014197323471307755,
-0.06310112029314041,
-0.0641108825802803,
0.02583642676472664,
-0.03824583813548088,
-0.0023833950981497765,
-0.08480305969715118,
-4.525223928908417e-8,
-0.012511124834418297,
0.036265745759010315,
0.05512136593461037,
0.0010532671585679054,
0.0896260216832161,
0.006256373133510351,
-0.04571941867470741,
-0.08784724771976471,
-0.030141368508338928,
0.00029484869446605444,
-0.018375368788838387,
0.021317649632692337,
0.08383587002754211,
-0.02508842572569847,
-0.12049313634634018,
0.05656496062874794,
0.027380814775824547,
-0.003927409183233976,
-0.030635269358754158,
-0.029593341052532196,
-0.005583095829933882,
0.06111231446266174,
0.029188038781285286,
-0.03381592035293579,
0.03371139615774155,
-0.06797108054161072,
-0.00974698830395937,
0.025718363001942635,
0.057539377361536026,
0.02739659696817398,
0.006896836217492819,
-0.0513821579515934,
0.07982394844293594,
0.06979167461395264,
-0.009933209046721458,
0.12158620357513428,
0.03612378612160683,
-0.04009387269616127,
0.00683259591460228,
-0.0325038842856884,
-0.038895443081855774,
0.11348069459199905,
-0.024515274912118912,
-0.0013324491446837783,
-0.020840603858232498,
-0.0011193434474989772,
-0.038539282977581024,
-0.08218571543693542,
0.06832780689001083,
-0.01750945672392845,
-0.006667808163911104,
-0.011896690353751183,
0.000005558243628911441,
0.11021851748228073,
0.08678142726421356,
-0.008227606303989887,
-0.008268317207694054,
-0.09877078980207443,
0.08261875808238983,
0.11161503195762634,
-0.018135398626327515,
0.047484204173088074,
0.015484128147363663,
-0.013622470200061798
] | 0 | 0 | 0 | 0 |
|
Residual Stream Analysis with Multi-Layer SAEs
|
https://openreview.net/forum?id=XAjfjizaKs
|
[
"Tim Lawson",
"Lucy Farnik",
"Conor Houghton",
"Laurence Aitchison"
] |
Poster
|
Sparse autoencoders (SAEs) are a promising approach to interpreting the internal representations of transformer language models. However, SAEs are usually trained separately on each transformer layer, making it difficult to use them to study how information flows across layers. To solve this problem, we introduce the multi-layer SAE (MLSAE): a single SAE trained on the residual stream activation vectors from every transformer layer. Given that the residual stream is understood to preserve information across layers, we expected MLSAE latents to 'switch on' at a token position and remain active at later layers. Interestingly, we find that individual latents are often active at a single layer for a given token or prompt, but the layer at which an individual latent is active may differ for different tokens or prompts. We quantify these phenomena by defining a distribution over layers and considering its variance. We find that the variance of the distributions of latent activations over layers is about two orders of magnitude greater when aggregating over tokens compared with a single token. For larger underlying models, the degree to which latents are active at multiple layers increases, which is consistent with the fact that the residual stream activation vectors at adjacent layers become more similar. Finally, we relax the assumption that the residual stream basis is the same at every layer by applying pre-trained tuned-lens transformations, but our findings remain qualitatively similar. Our results represent a new approach to understanding how representations change as they flow through transformers. We release our code to train and analyze MLSAEs at https://github.com/tim-lawson/mlsae.
|
sparse autoencoders, mechanistic interpretability, language models
|
We train a single SAE on the residual stream activations from every transformer layer and find latents active at multiple layers for different tokens.
| 11,815 |
2409.04185
|
[
-0.08431844413280487,
-0.11680865287780762,
0.04518307372927666,
0.06044325232505798,
0.11882583051919937,
0.03686429560184479,
-0.004432230722159147,
-0.007336567156016827,
0.1161736473441124,
-0.0667969137430191,
-0.041404951363801956,
-0.06859426200389862,
-0.023853225633502007,
0.03345886617898941,
-0.0160288754850626,
-0.055158715695142746,
-0.02624395489692688,
0.08454225212335587,
-0.13672944903373718,
-0.10523783415555954,
0.02044246532022953,
0.02188270539045334,
-0.042444318532943726,
0.0074140941724181175,
0.09918329119682312,
0.01196100004017353,
-0.02175135724246502,
-0.08233439922332764,
0.03270578756928444,
-0.012025433592498302,
0.04575219005346298,
0.02618972398340702,
-0.05425234138965607,
0.08399350196123123,
-0.015551066026091576,
0.047745656222105026,
-0.07331912219524384,
0.010605763643980026,
0.01954096369445324,
-0.0008962327265180647,
0.05272525176405907,
-0.018032822757959366,
-0.0009716215427033603,
0.04290087893605232,
0.013543525710701942,
-0.007372342050075531,
0.014002054929733276,
-0.04982590675354004,
-0.03981632739305496,
-0.0040832748636603355,
0.008233502507209778,
-0.03435291722416878,
0.006337678991258144,
0.08831720054149628,
-0.05382522568106651,
-0.04460393637418747,
0.015383048914372921,
0.010009133256971836,
0.01256563514471054,
0.00981868989765644,
-0.04493188112974167,
-0.07131961733102798,
-0.026371005922555923,
-0.021123437210917473,
-0.033700063824653625,
0.01406546775251627,
0.00931573286652565,
0.07920140773057938,
0.004001135937869549,
0.02718474343419075,
0.005547248292714357,
0.047734584659338,
-0.10975991189479828,
0.05257750302553177,
0.09336084872484207,
-0.01797134056687355,
0.024858087301254272,
-0.0312248133122921,
0.03761450946331024,
-0.09285695105791092,
0.06009296700358391,
0.053770363330841064,
0.05184531584382057,
-0.02358744852244854,
0.11317557096481323,
-0.006330872420221567,
-0.0062249805778265,
0.019289864227175713,
-0.0030341465026140213,
0.03288859501481056,
-0.08078907430171967,
-0.06502014398574829,
0.02566337026655674,
0.029595864936709404,
0.05051467940211296,
0.0282434169203043,
0.00567818945273757,
-0.025737568736076355,
-0.009532528929412365,
0.09575530141592026,
-0.006707282736897469,
0.01051614060997963,
0.06386303901672363,
0.017843371257185936,
0.0218537375330925,
-0.04635803401470184,
0.042663510888814926,
0.06415940076112747,
-0.03125810995697975,
-0.03722911328077316,
-0.08005695790052414,
0.0335410013794899,
-0.001405550166964531,
-0.049713581800460815,
0.04294703155755997,
-0.020572742447257042,
0.005621652584522963,
-0.05765077844262123,
-0.04826585575938225,
0.04687478765845299,
-0.04364869371056557,
-0.001078234170563519,
0.0023275981657207012,
0.006441534496843815,
0.07193811237812042,
0.029741359874606133,
-0.023940807208418846,
3.518337611299633e-33,
0.003426408628001809,
0.058984532952308655,
-0.0040223910473287106,
0.04931415989995003,
-0.024814479053020477,
-0.029530875384807587,
0.047985512763261795,
-0.019323378801345825,
0.03055942803621292,
-0.0325990654528141,
-0.03948666527867317,
0.055008888244628906,
-0.06288743764162064,
0.07985740154981613,
-0.04361436888575554,
-0.005270715337246656,
-0.0011050243629142642,
0.029249463230371475,
-0.0045488933101296425,
-0.03544309735298157,
0.008943251334130764,
0.005003706086426973,
-0.017045816406607628,
-0.039822664111852646,
-0.045669592916965485,
-0.06058317795395851,
0.04513191431760788,
-0.06739671528339386,
-0.09862364828586578,
0.020620722323656082,
-0.12101916968822479,
-0.000051777871703961864,
0.0943160355091095,
-0.009729482233524323,
0.027180079370737076,
-0.01403129193931818,
0.07789148390293121,
-0.039984773844480515,
-0.007596890442073345,
-0.04929332807660103,
0.020386813208460808,
0.039219848811626434,
-0.026626240462064743,
-0.051523350179195404,
-0.06151298061013222,
-0.061308786273002625,
-0.03470732644200325,
-0.019219188019633293,
-0.027263633906841278,
-0.053583044558763504,
0.09071239084005356,
0.041929375380277634,
-0.012708785012364388,
0.010190429165959358,
0.07279542088508606,
0.03439674898982048,
-0.011018767952919006,
0.009750504978001118,
0.04495421424508095,
0.0010964561952278018,
0.030761322006583214,
0.06063312292098999,
-0.022041000425815582,
0.06347959488630295,
0.09979337453842163,
0.03465011715888977,
0.027540909126400948,
0.049300651997327805,
0.06243622303009033,
-0.04600124806165695,
-0.07549692690372467,
-0.008256434462964535,
-0.026410076767206192,
-0.06604643166065216,
0.027898957952857018,
0.004881573840975761,
0.030522674322128296,
-0.10332880914211273,
0.017295828089118004,
0.07624577730894089,
-0.02294474095106125,
-0.04899897053837776,
0.045337263494729996,
-0.0023388515692204237,
-0.07490058988332748,
0.052813079208135605,
0.03270481526851654,
-0.10849738121032715,
-0.07368301600217819,
0.0213470496237278,
0.025341732427477837,
-0.06728368997573853,
0.025053519755601883,
-0.0433478020131588,
0.012310481630265713,
-4.2130536706096514e-33,
-0.03759590536355972,
0.061182960867881775,
-0.04942181706428528,
0.051257241517305374,
-0.04430781304836273,
-0.0369923859834671,
0.030118534341454506,
0.06856511533260345,
-0.03259746730327606,
-0.03364016115665436,
0.011250095441937447,
-0.046650901436805725,
-0.005397818051278591,
0.0006751964101567864,
0.07614342868328094,
0.006438429467380047,
0.017756594344973564,
0.016873124986886978,
0.1194748803973198,
0.03844260796904564,
0.04053447023034096,
0.024470334872603416,
-0.11399854719638824,
0.059789884835481644,
-0.036574311554431915,
-0.007815729826688766,
-0.07515491545200348,
0.10972335189580917,
0.008304993622004986,
-0.0263848677277565,
0.012971119955182076,
-0.01747487485408783,
-0.012339090928435326,
-0.031059112399816513,
-0.0509108267724514,
0.04397856071591377,
0.002737009432166815,
0.03306615352630615,
-0.02316112630069256,
0.050683822482824326,
0.05271035060286522,
-0.07929159700870514,
-0.02343294396996498,
-0.015655644237995148,
-0.026155846193432808,
-0.010746851563453674,
-0.038156431168317795,
0.007212762720882893,
-0.04320908337831497,
-0.022960929200053215,
0.009472915902733803,
0.051834579557180405,
-0.0798012763261795,
0.009037275798618793,
-0.03340661898255348,
-0.03432116657495499,
0.04033784568309784,
-0.06494423002004623,
0.04937450960278511,
-0.04359558969736099,
-0.0146280312910676,
-0.007206371054053307,
0.001224225852638483,
-0.16604669392108917,
0.01608673483133316,
-0.08413640409708023,
-0.010266048833727837,
-0.027268560603260994,
0.01374333817511797,
-0.016695821657776833,
0.1478150635957718,
-0.030275598168373108,
-0.05697939917445183,
-0.006396132986992598,
0.041380640119314194,
-0.0225613322108984,
0.005403077695518732,
-0.0500488243997097,
-0.021350160241127014,
-0.046709123998880386,
-0.08329425752162933,
0.037289515137672424,
-0.013340719975531101,
0.010110820643603802,
0.037615157663822174,
0.11201488971710205,
0.016770506277680397,
-0.0042489697225391865,
0.06335140019655228,
0.021637126803398132,
-0.024023758247494698,
0.00022750982316210866,
-0.01488080807030201,
0.021997833624482155,
-0.08878812938928604,
-4.736984138276057e-8,
-0.05790429189801216,
0.07133806496858597,
-0.01502128504216671,
-0.035665277391672134,
0.020954815670847893,
-0.04469380900263786,
0.0045146397314965725,
0.08824115246534348,
-0.015602441504597664,
-0.019370442256331444,
0.06453731656074524,
-0.05827818065881729,
-0.011164768598973751,
-0.08996956795454025,
0.03116954118013382,
0.07579072564840317,
-0.013179796747863293,
0.07316000014543533,
-0.02557365410029888,
-0.03643270954489708,
0.06901036947965622,
0.1334310919046402,
-0.0013531048316508532,
-0.055360618978738785,
-0.008689961396157742,
-0.01851997897028923,
0.041193895041942596,
0.0754045769572258,
0.02801070548593998,
0.005661685951054096,
0.0073756990022957325,
0.025438811630010605,
0.009114416316151619,
-0.04267452657222748,
0.010685197077691555,
0.1557493358850479,
0.1057656854391098,
-0.08440759032964706,
-0.06489673256874084,
0.002493636915460229,
0.005702021066099405,
-0.017792994156479836,
-0.07057228684425354,
0.003652166575193405,
-0.028054989874362946,
0.047805119305849075,
-0.01066469494253397,
-0.09922458976507187,
0.01230626180768013,
-0.09253976494073868,
0.018601736053824425,
-0.04726006090641022,
-0.05415115877985954,
0.15652430057525635,
0.05655067041516304,
-0.015589464455842972,
0.0021449497435241938,
-0.07738129794597626,
0.01466951984912157,
0.017575694248080254,
0.005199492443352938,
0.06250561773777008,
-0.07101268321275711,
0.04414130747318268
] |
https://github.com/tim-lawson/mlsae
| 22 | 0 | 0 | 0 |
The Same but Different: Structural Similarities and Differences in Multilingual Language Modeling
|
https://openreview.net/forum?id=NCrFA7dq8T
|
[
"Ruochen Zhang",
"Qinan Yu",
"Matianyu Zang",
"Carsten Eickhoff",
"Ellie Pavlick"
] |
Poster
|
We employ new tools from mechanistic interpretability to ask whether the internal structure of large language models (LLMs) shows correspondence to the linguistic structures which underlie the languages on which they are trained. In particular, we ask (1) when two languages employ the same morphosyntactic processes, do LLMs handle them using shared internal circuitry? and (2) when two languages require different morphosyntactic processes, do LLMs handle them using different internal circuitry? In a focused case study on English and Chinese multilingual and monolingual models, we analyze the internal circuitry involved in two tasks. We find evidence that models employ the same circuit to handle the same syntactic process independently of the language in which it occurs, and that this is the case even for monolingual models trained completely independently. Moreover, we show that multilingual models employ language-specific components (attention heads and feed-forward networks) when needed to handle linguistic processes (e.g., morphological marking) that only exist in some languages. Together, our results are revealing about how LLMs trade off between exploiting common structures and preserving linguistic differences when tasked with modeling multiple languages simultaneously, opening the door for future work in this direction.
|
multilinguality, interpretability
| null | 11,814 |
2410.09223
|
[
-0.011769022792577744,
-0.08145508170127869,
0.0740785002708435,
-0.020294539630413055,
-0.008021143265068531,
-0.05815070867538452,
-0.03809358552098274,
-0.01229577325284481,
0.15259896218776703,
-0.03353649005293846,
-0.016643645241856575,
-0.11223593354225159,
0.07148627191781998,
0.06019536778330803,
0.06650853157043457,
-0.052034541964530945,
-0.016228465363383293,
0.0734817162156105,
-0.09723498672246933,
-0.041592225432395935,
0.09140200912952423,
-0.0005515213124454021,
-0.017072705551981926,
0.02741643413901329,
-0.014844045974314213,
-0.029184559360146523,
-0.08393184095621109,
-0.022216079756617546,
0.02154865674674511,
-0.07507152855396271,
0.010845395736396313,
0.04117894172668457,
-0.031512949615716934,
0.10724791139364243,
-0.04328257963061333,
0.08838877081871033,
-0.029494628310203552,
-0.04103869944810867,
0.012974247336387634,
-0.061266519129276276,
-0.07683898508548737,
0.03668524697422981,
0.02076130360364914,
-0.0246585626155138,
0.09211412072181702,
0.0417625792324543,
-0.015232555568218231,
-0.009700051508843899,
-0.11751234531402588,
-0.0375685952603817,
-0.11115297675132751,
-0.03338789567351341,
0.02813492715358734,
0.07973232120275497,
0.035063039511442184,
0.005626533180475235,
0.07673577964305878,
0.07443738728761673,
-0.04458927735686302,
-0.050567418336868286,
-0.10982174426317215,
-0.03863377869129181,
0.00448486115783453,
0.055975768715143204,
-0.015078302472829819,
0.02642042376101017,
0.011651359498500824,
0.040112048387527466,
-0.05283016711473465,
0.019879279658198357,
0.013590967282652855,
0.018164850771427155,
-0.021659288555383682,
-0.023953158408403397,
0.054421789944171906,
0.028610333800315857,
0.00675678625702858,
0.007871978916227818,
-0.010389953851699829,
-0.1020357683300972,
-0.005017205607146025,
0.03512657806277275,
0.03989310562610626,
-0.05395037680864334,
0.10811911523342133,
0.027708858251571655,
-0.027303174138069153,
0.019238559529185295,
0.002900532679632306,
-0.028166357427835464,
-0.0765003189444542,
-0.07820688188076019,
0.011495874263346195,
0.032459910959005356,
0.03590252250432968,
0.017451021820306778,
0.05387888476252556,
0.055359408259391785,
-0.0038944121915847063,
0.06328281760215759,
0.08379220962524414,
0.04143814742565155,
0.041607704013586044,
-0.003009189385920763,
-0.023639719933271408,
-0.024420488625764847,
0.009873543865978718,
-0.013656014576554298,
0.05295953154563904,
-0.112852081656456,
-0.0008463922422379255,
0.06314844638109207,
-0.018561121076345444,
0.06476863473653793,
0.003209219081327319,
-0.057367149740457535,
0.028404539451003075,
0.00013023384963162243,
0.08753085136413574,
0.027467060834169388,
-0.08151279389858246,
0.004445718601346016,
0.041141707450151443,
-0.034604262560606,
0.02675716020166874,
-0.004302906803786755,
-0.012519839219748974,
2.0001910653164646e-33,
0.07507863640785217,
0.026578756049275398,
-0.02999447099864483,
-0.017696840688586235,
0.04808785393834114,
-0.02256614901125431,
0.012998965568840504,
0.02775166556239128,
0.022005043923854828,
-0.05558782443404198,
0.01820492371916771,
0.08455506712198257,
-0.05584735423326492,
0.02632649429142475,
0.03848079591989517,
0.03177071735262871,
-0.004671783186495304,
-0.002521960064768791,
-0.01805824041366577,
-0.0153061393648386,
0.08451640605926514,
0.038439322263002396,
0.030196409672498703,
-0.023993847891688347,
0.007819466292858124,
0.01081257127225399,
0.05131658539175987,
-0.07226728647947311,
-0.09664954245090485,
0.02147919498383999,
-0.091596320271492,
0.01620200276374817,
0.0296720489859581,
0.07286902517080307,
-0.029143067076802254,
-0.06757813692092896,
0.042304836213588715,
-0.05204132944345474,
0.051477499306201935,
-0.07687356323003769,
0.029114078730344772,
0.02800043858587742,
-0.018840929493308067,
0.021515564993023872,
-0.08256537467241287,
-0.020568115636706352,
-0.0692279189825058,
0.03574077785015106,
-0.049932338297367096,
-0.07300890982151031,
0.02547803521156311,
-0.0013104446697980165,
0.061684995889663696,
-0.07227012515068054,
0.03586119785904884,
0.09801998734474182,
0.007306999992579222,
-0.02577938884496689,
0.03009122796356678,
0.021124372258782387,
-0.02968035824596882,
0.07457141578197479,
0.020305098965764046,
0.062227025628089905,
0.09036730974912643,
0.04499799385666847,
-0.08506754785776138,
0.043544620275497437,
0.04916704073548317,
-0.012276765890419483,
-0.09161633253097534,
-0.011527922935783863,
-0.1280413419008255,
-0.025005729869008064,
0.046137209981679916,
-0.049475349485874176,
-0.013976960442960262,
-0.12479974329471588,
0.04973670840263367,
0.09560265392065048,
0.012627394869923592,
-0.03184032440185547,
0.019606009125709534,
-0.035335514694452286,
-0.005486062727868557,
-0.003711557015776634,
0.07510501146316528,
-0.08081606030464172,
0.07778138667345047,
0.03534885495901108,
-0.004072810523211956,
-0.02919173054397106,
-0.005127856507897377,
-0.017070308327674866,
0.035055011510849,
-3.4127117296965735e-33,
-0.05905604735016823,
0.009095597080886364,
-0.06639041751623154,
0.03594903647899628,
-0.05288667604327202,
-0.08952020108699799,
0.02713826671242714,
-0.018746670335531235,
-0.048842549324035645,
-0.04559063911437988,
-0.0467676967382431,
-0.03023276850581169,
0.06230982765555382,
0.04360751062631607,
0.0024334120098501444,
-0.05677170306444168,
0.08146470785140991,
0.08343058079481125,
0.07673288136720657,
0.05936143547296524,
0.014602754265069962,
0.04359549656510353,
-0.14732925593852997,
0.02026665024459362,
-0.055766403675079346,
0.03618405759334564,
-0.13201400637626648,
0.015872744843363762,
-0.060706935822963715,
-0.028791002929210663,
-0.05365738272666931,
0.01560208573937416,
0.004306586924940348,
-0.04526035487651825,
-0.04460990056395531,
0.06312248110771179,
-0.08007214963436127,
-0.03448121249675751,
0.020942162722349167,
0.011838801205158234,
-0.0007951502338983119,
-0.04178953543305397,
-0.012870097532868385,
0.01913980394601822,
0.008073830977082253,
-0.02936650440096855,
-0.11943937838077545,
0.011833978816866875,
-0.03601202368736267,
-0.0023134436924010515,
-0.023915933445096016,
0.01098383404314518,
-0.041795097291469574,
-0.05907028168439865,
-0.04753636568784714,
-0.07466943562030792,
0.03080770932137966,
-0.07953518629074097,
-0.02094115875661373,
-0.08977851271629333,
-0.023719027638435364,
-0.01842179149389267,
0.029286500066518784,
-0.026897817850112915,
0.01820320449769497,
-0.01087892521172762,
0.020762033760547638,
-0.07554904371500015,
0.11038706451654434,
-0.07500404864549637,
0.06886804848909378,
-0.019954653456807137,
-0.026526926085352898,
0.018110930919647217,
0.004563674330711365,
-0.010257751680910587,
-0.08891861140727997,
-0.06693543493747711,
-0.006374862045049667,
-0.06611742079257965,
-0.006715675815939903,
0.02987092360854149,
-0.0023589450865983963,
-0.05643953010439873,
-0.0031076723244041204,
0.01385952066630125,
0.013862107880413532,
0.08227024227380753,
0.0192310381680727,
-0.04862283915281296,
0.03330254554748535,
0.08299358934164047,
-0.029245460405945778,
0.07855301350355148,
-0.005138492211699486,
-4.6283957999548875e-8,
-0.04258839413523674,
-0.004783932119607925,
0.012761806137859821,
0.056666988879442215,
-0.017546435818076134,
-0.11387481540441513,
-0.0312604084610939,
0.023980457335710526,
-0.00002388563734712079,
-0.011427943594753742,
0.022370532155036926,
0.03354188799858093,
-0.00966863427311182,
-0.04076562449336052,
0.0020891076419502497,
0.0720221996307373,
0.043683040887117386,
0.05395693704485893,
-0.01750382035970688,
-0.01249286625534296,
0.03286871314048767,
0.01987784169614315,
-0.002669782843440771,
0.0552598312497139,
0.035649627447128296,
-0.031462591141462326,
-0.08411100506782532,
0.02949104644358158,
0.013750961981713772,
-0.01725980080664158,
-0.003738687140867114,
-0.0005811316659674048,
-0.07169708609580994,
0.007603513542562723,
0.031463395804166794,
0.06869985908269882,
0.004813142120838165,
0.016218526288866997,
0.05356912314891815,
0.009469958022236824,
0.084665946662426,
0.028660167008638382,
-0.06640391051769257,
0.08905589580535889,
0.0903184786438942,
-0.05012804642319679,
-0.06587810069322586,
-0.08021999895572662,
0.024785714223980904,
-0.007727215997874737,
0.03097178414463997,
-0.03907909244298935,
-0.01448267325758934,
0.06052719056606293,
0.0008504580473527312,
0.03635190799832344,
0.03456077352166176,
-0.03304474055767059,
0.04275595769286156,
0.07581828534603119,
0.041977956891059875,
0.02288457378745079,
0.07989305257797241,
-0.03585580736398697
] | 0 | 0 | 0 | 0 |
|
Disentangling Representations through Multi-task Learning
|
https://openreview.net/forum?id=yVGGtsOgc7
|
[
"Pantelis Vafidis",
"Aman Bhargava",
"Antonio Rangel"
] |
Poster
|
Intelligent perception and interaction with the world hinges on internal representations that capture its underlying structure ("disentangled" or "abstract" representations). Disentangled representations serve as world models, isolating latent factors of variation in the world along approximately orthogonal directions, thus facilitating feature-based generalization. We provide experimental and theoretical results guaranteeing the emergence of disentangled representations in agents that optimally solve multi-task evidence accumulation classification tasks, canonical in the neuroscience literature. The key conceptual finding is that, by producing accurate multi-task classification estimates, a system implicitly represents a set of coordinates specifying a disentangled representation of the underlying latent state of the data it receives. The theory provides conditions for the emergence of these representations in terms of noise, number of tasks, and evidence accumulation time, when the classification boundaries are affine in the latent space. Surprisingly, the theory also produces closed-form expressions for extracting the disentangled representation from the model's latent state $\mathbf Z(t)$. We experimentally validate these predictions in RNNs trained on multi-task classification, which learn disentangled representations in the form of continuous attractors, leading to zero-shot out-of-distribution (OOD) generalization in predicting latent factors. We demonstrate the robustness of our framework across autoregressive architectures, decision boundary geometries and in tasks requiring classification confidence estimation. We find that transformers are particularly suited for disentangling representations, which might explain their unique world understanding abilities. Overall, our framework establishes a formal link between competence at multiple tasks and the formation of disentangled, interpretable world models in both biological and artificial systems, and helps explain why ANNs often arrive at human-interpretable concepts, and how they both may acquire exceptional zero-shot generalization capabilities.
|
zero-shot generalization, disentanglement, interpretability, world models, multi-task learning, computational neuroscience, neuroAI, evidence accumulation, cognitive maps, continuous attractors, RNNs, transformers
|
We theoretically prove multi-task learning is guaranteed to lead to disentangled, generalizable representations in autoregressive models, and validate our theory on RNNs and transformers performing cognitive neuroscience evidence accumulation tasks.
| 11,804 |
2407.11249
|
[
-0.07849586755037308,
-0.14557801187038422,
0.044497050344944,
0.03420168533921242,
0.09939604997634888,
0.007629345171153545,
0.05735114589333534,
-0.07191581279039383,
0.08435708284378052,
0.0011952647473663092,
-0.10313360393047333,
-0.07173590362071991,
0.027153780683875084,
0.12371634691953659,
-0.08566319942474365,
-0.022434784099459648,
0.09110013395547867,
0.022273322567343712,
-0.11779285967350006,
-0.01698032021522522,
0.04885824769735336,
0.012782708741724491,
0.06750477105379105,
0.06539857387542725,
0.012580540962517262,
0.018678739666938782,
0.06962444633245468,
-0.014425824396312237,
0.03218380734324455,
0.0020132693462073803,
0.12274450063705444,
0.009596861898899078,
-0.049744900315999985,
0.043922796845436096,
-0.06340499967336655,
0.05418848618865013,
-0.08157962560653687,
0.032591313123703,
0.05047747865319252,
0.023492177948355675,
0.06940515339374542,
0.10790511965751648,
-0.026750216260552406,
0.07000377029180527,
0.022702006623148918,
0.03118518367409706,
0.01438424177467823,
-0.054375454783439636,
-0.020472606644034386,
-0.02302829548716545,
-0.013964616693556309,
0.07683665305376053,
-0.015200422145426273,
0.04503580555319786,
-0.001576353912241757,
0.03455550596117973,
0.07434884458780289,
0.0334932841360569,
-0.08735507726669312,
-0.01971200481057167,
0.030081456527113914,
-0.08494462817907333,
-0.016790514811873436,
-0.07525099813938141,
0.05654595419764519,
0.055225636810064316,
-0.0506993867456913,
0.02930680476129055,
-0.07774621248245239,
-0.019868992269039154,
-0.04532409459352493,
0.0972062274813652,
-0.029252106323838234,
-0.03232163190841675,
0.0448925606906414,
-0.037919364869594574,
0.05151787027716637,
-0.021080100908875465,
0.034612271934747696,
-0.0683467909693718,
0.0127412723377347,
0.050077252089977264,
0.00869822222739458,
0.024504005908966064,
0.1718825101852417,
0.02711568959057331,
-0.05286480113863945,
0.03890123590826988,
-0.013192851096391678,
-0.04408763721585274,
-0.019077634438872337,
-0.07551994174718857,
-0.049142442643642426,
0.048557162284851074,
0.04266344755887985,
0.01599191501736641,
0.040852878242731094,
0.0651143416762352,
0.03853603079915047,
0.021171115338802338,
0.008552969433367252,
-0.019108163192868233,
0.003440368454903364,
0.03194909170269966,
0.07222925126552582,
-0.015773432329297066,
0.0434143953025341,
-0.02666814997792244,
0.0029079015366733074,
-0.037955209612846375,
-0.0618099607527256,
0.010804083198308945,
0.006614765152335167,
0.11378563940525055,
-0.01821027137339115,
-0.018395697697997093,
0.03414742276072502,
0.04143815115094185,
0.07587818801403046,
0.06474674493074417,
-0.03933557868003845,
-0.04041079804301262,
0.01817222498357296,
-0.0204987283796072,
0.03740512207150459,
0.01889391615986824,
-0.07063999772071838,
2.179868682917513e-33,
-0.03599799796938896,
-0.044113270938396454,
0.017658516764640808,
0.010356432758271694,
0.009536786936223507,
0.018909335136413574,
-0.04536532610654831,
-0.013564187102019787,
0.029705990105867386,
0.06050696596503258,
-0.037728194147348404,
0.03074185736477375,
-0.044618070125579834,
0.0710466280579567,
-0.016329895704984665,
-0.038380153477191925,
-0.0544416718184948,
0.04054078459739685,
-0.01156011875718832,
-0.013968611136078835,
0.03929952159523964,
0.04759112372994423,
-0.041507288813591,
-0.01762131229043007,
-0.015586240217089653,
0.005283304490149021,
0.024985481053590775,
-0.08253233134746552,
-0.07037999480962753,
-0.008465667255222797,
-0.10831320285797119,
0.04257432371377945,
-0.016070326790213585,
0.01448865421116352,
0.012133629992604256,
-0.00767513969913125,
0.03820085898041725,
0.00014236860442906618,
0.022106237709522247,
-0.07382635027170181,
0.021098246797919273,
0.019651399925351143,
-0.04620801657438278,
-0.024689072743058205,
-0.03296234458684921,
-0.055840857326984406,
0.03869117051362991,
-0.04851308837532997,
-0.03785805031657219,
-0.0576552078127861,
0.019907750189304352,
-0.021075213328003883,
0.04093680530786514,
-0.18210798501968384,
-0.02128460258245468,
-0.03075939416885376,
-0.031176216900348663,
-0.014486907981336117,
0.08232565969228745,
-0.0072230068035423756,
-0.048898354172706604,
0.021489175036549568,
-0.02408212050795555,
0.06288857758045197,
0.056980304419994354,
0.0015212155412882566,
-0.043703846633434296,
0.03245748206973076,
0.0984724685549736,
0.05690746381878853,
-0.03957509621977806,
0.035561703145504,
-0.02630525454878807,
-0.09982193261384964,
0.04243982955813408,
-0.010090437717735767,
0.02220666967332363,
-0.05133213475346565,
-0.015155158005654812,
0.021621961146593094,
0.0402461439371109,
-0.07534769177436829,
-0.06286860257387161,
0.013956110924482346,
-0.0649135410785675,
-0.021972574293613434,
0.014809785410761833,
-0.09389257431030273,
-0.04888441041111946,
-0.0025074081495404243,
-0.06021180376410484,
-0.013982354663312435,
-0.008287547156214714,
-0.02708055078983307,
-0.06346521526575089,
-2.5404567178200345e-33,
-0.07055822014808655,
0.004070498049259186,
-0.037639979273080826,
-0.005118916742503643,
0.0013741031289100647,
-0.06025717407464981,
-0.027824658900499344,
-0.03501918911933899,
-0.007213999982923269,
-0.031014498323202133,
-0.03393405303359032,
-0.03765914961695671,
0.05020646005868912,
0.035645741969347,
-0.00791548378765583,
0.001983115915209055,
-0.009894166141748428,
0.07652001082897186,
0.049295131117105484,
-0.025719672441482544,
0.043457478284835815,
0.01008475013077259,
-0.08683530986309052,
-0.0015170141123235226,
-0.002773804124444723,
0.07295475900173187,
0.021713711321353912,
0.06298332661390305,
0.012305125594139099,
0.022411691024899483,
-0.06507739424705505,
-0.026666393503546715,
0.01468636468052864,
-0.029064130038022995,
-0.0784318819642067,
0.07054274529218674,
-0.04524768516421318,
-0.01224511954933405,
-0.07141939550638199,
-0.03271956369280815,
0.008129820227622986,
-0.048041727393865585,
-0.026111772283911705,
0.03353032469749451,
-0.00843073520809412,
-0.0514487586915493,
-0.03396428003907204,
0.0315871424973011,
-0.03232729062438011,
0.08189989626407623,
-0.0007098717032931745,
0.09331012517213821,
-0.09380219131708145,
0.015301181934773922,
-0.026261212304234505,
-0.029101107269525528,
0.00932145956903696,
-0.052894581109285355,
0.1138124093413353,
0.012047648429870605,
0.003064132295548916,
0.04694901779294014,
0.009308192878961563,
0.03289402648806572,
0.027863213792443275,
-0.010465790517628193,
0.023868920281529427,
0.03186452016234398,
0.1028718501329422,
0.009930823929607868,
0.05637522041797638,
0.044840600341558456,
-0.01563524454832077,
-0.045622147619724274,
0.034696970134973526,
0.032053858041763306,
-0.020137589424848557,
-0.04324319586157799,
-0.034660957753658295,
0.03811798244714737,
-0.03268647938966751,
-0.07338593900203705,
0.0534045584499836,
0.02298612706363201,
-0.025116464123129845,
-0.02206381782889366,
-0.006945732049643993,
0.05356839299201965,
0.028805207461118698,
-0.04905136674642563,
-0.018195947632193565,
-0.042495403438806534,
-0.036325085908174515,
0.018765216693282127,
-0.09711447358131409,
-5.0112898719589793e-8,
-0.08356492966413498,
-0.01449654996395111,
0.05345778539776802,
-0.02906939573585987,
0.00045500777196139097,
-0.04091213271021843,
0.02333255112171173,
0.06427790969610214,
-0.024543508887290955,
-0.0014172849478200078,
0.17416468262672424,
-0.08284317702054977,
-0.03772549703717232,
-0.02378278598189354,
0.09634224325418472,
0.06954699754714966,
0.061888474971055984,
0.010776463896036148,
0.028942590579390526,
0.04535239189863205,
0.0950162410736084,
0.03987136855721474,
0.023051723837852478,
-0.06650727242231369,
0.03902673348784447,
-0.032365832477808,
-0.08146987110376358,
0.15157285332679749,
0.0078037213534116745,
0.01858874224126339,
0.05833915248513222,
0.07174737751483917,
0.03109259530901909,
-0.04588828980922699,
-0.07439025491476059,
0.012946944683790207,
0.015018822625279427,
-0.01869877055287361,
-0.04002295434474945,
-0.019071189686655998,
-0.03909604996442795,
0.07255882769823074,
-0.02831576205790043,
0.02211177535355091,
0.014059764333069324,
-0.008482136763632298,
-0.0022121286019682884,
-0.08991716057062149,
0.09443484991788864,
0.022155549377202988,
0.0035259639844298363,
0.04119705408811569,
0.0019412210676819086,
0.1016552522778511,
0.064722441136837,
0.005182797554880381,
-0.006328264717012644,
-0.062606580555439,
-0.024093229323625565,
0.004474756307899952,
-0.03222671523690224,
0.08734136819839478,
-0.07410294562578201,
0.010639995336532593
] | 0 | 0 | 0 | 0 |
|
Training Free Guided Flow-Matching with Optimal Control
|
https://openreview.net/forum?id=61ss5RA1MM
|
[
"Luran Wang",
"Chaoran Cheng",
"Yizhen Liao",
"Yanru Qu",
"Ge Liu"
] |
Poster
|
Controlled generation with pre-trained Diffusion and Flow Matching models has vast applications. One strategy for guiding ODE-based generative models is through optimizing a target loss $R(x_1)$ while staying close to the prior distribution. Along this line, some recent work showed the effectiveness of guiding flow model by differentiating through its ODE sampling process. Despite the superior performance, the theoretical understanding of this line of methods is still preliminary, leaving space for algorithm improvement. Moreover, existing methods predominately focus on Euclidean data manifold, and there is a compelling need for guided flow methods on complex geometries such as SO(3), which prevails in high-stake scientific applications like protein design. We present OC-Flow, a general and theoretically grounded training-free framework for guided flow matching using optimal control. Building upon advances in optimal control theory, we develop effective and practical algorithms for solving optimal control in guided ODE-based generation and provide a systematic theoretical analysis of the convergence guarantee in both Euclidean and SO(3). We show that existing backprop-through-ODE methods can be interpreted as special cases of Euclidean OC-Flow. OC-Flow achieved superior performance in extensive experiments on text-guided image manipulation, conditional molecule generation, and all-atom peptide design.
|
flow matching, controlled generation, inverse problem
|
We develope a general and theoretically grounded framework for guided flow matching on both Euclidean and SO3 geometries taking inpiration from optimal control.
| 11,803 |
2410.18070
|
[
-0.088930644094944,
-0.05446809157729149,
0.048129476606845856,
-0.03648674488067627,
0.0367155484855175,
-0.0169049184769392,
-0.03552506864070892,
0.009333423338830471,
0.07339021563529968,
-0.034049540758132935,
-0.04825018718838692,
0.033195119351148605,
0.01868996024131775,
0.01337083987891674,
-0.011455187574028969,
0.00015871158393565565,
0.016591913998126984,
0.12198001146316528,
-0.059260908514261246,
-0.022278154268860817,
-0.03654586151242256,
0.0778905600309372,
0.031212443485856056,
0.05576653033494949,
-0.014037046581506729,
0.020906666293740273,
0.03500838205218315,
-0.08149975538253784,
0.003564252285286784,
-0.041881315410137177,
0.12086834758520126,
0.0008395079639740288,
-0.11858727037906647,
-0.06990117579698563,
-0.06462709605693817,
0.07576911151409149,
-0.06872259825468063,
-0.00807149987667799,
-0.06954338401556015,
0.004179921001195908,
0.02151939645409584,
0.10000181943178177,
0.08075909316539764,
0.034559495747089386,
0.03840653598308563,
0.0588061586022377,
-0.03639817610383034,
-0.05190012976527214,
0.033863842487335205,
0.009221778251230717,
-0.07347387075424194,
-0.03874891996383667,
-0.042364295572042465,
-0.003523370949551463,
0.05548376590013504,
0.015378494746983051,
0.03891279175877571,
-0.049137793481349945,
0.006923843175172806,
0.01084685605019331,
0.02266962267458439,
-0.11665274202823639,
-0.04719288647174835,
-0.022532843053340912,
-0.04404702037572861,
-0.09061123430728912,
0.03358680009841919,
0.07365727424621582,
0.026656879112124443,
0.058795228600502014,
-0.03819088637828827,
0.06682880222797394,
-0.03542543575167656,
0.03858232870697975,
0.034825585782527924,
0.07025523483753204,
-0.0016316837863996625,
0.0462242029607296,
-0.009329313412308693,
-0.015019548125565052,
0.05400140956044197,
0.04119163006544113,
0.044920414686203,
0.015238781459629536,
-0.027527665719389915,
0.05166003480553627,
-0.04688553512096405,
0.04635895788669586,
0.13331668078899384,
0.028148440644145012,
0.035941414535045624,
0.04514135420322418,
-0.02453797683119774,
-0.059872400015592575,
0.0008842246024869382,
-0.031173458322882652,
0.06046604365110397,
-0.014940117485821247,
0.04082946479320526,
0.0691441148519516,
0.023531891405582428,
-0.01637968048453331,
0.025670725852251053,
0.05650318041443825,
0.08509895950555801,
-0.08253496885299683,
0.01011266652494669,
0.09220287948846817,
0.04929868504405022,
-0.010060001164674759,
0.057548217475414276,
-0.06952563673257828,
0.04984385892748833,
0.061524707823991776,
-0.015517469495534897,
-0.049375951290130615,
0.015508272685110569,
-0.0709395706653595,
-0.006585204042494297,
0.08325383812189102,
-0.08916381001472473,
-0.015550249256193638,
-0.0298808254301548,
0.010836565867066383,
-0.008880119770765305,
0.0048761554062366486,
-0.06789927929639816,
3.132929013179499e-33,
0.046538230031728745,
0.020782431587576866,
0.05464131757616997,
-0.038160186260938644,
-0.0159385297447443,
0.0026408610865473747,
0.04015939310193062,
-0.0058457921259105206,
-0.028873788192868233,
0.02786845713853836,
-0.05519287660717964,
-0.10988929122686386,
-0.0830487385392189,
0.07912396639585495,
-0.011626895517110825,
-0.07995522022247314,
0.020206015557050705,
0.09715084731578827,
-0.061559997498989105,
0.043517839163541794,
0.03327285498380661,
-0.020356440916657448,
-0.037210918962955475,
-0.057621635496616364,
-0.0058614397421479225,
0.01290094293653965,
-0.05823788419365883,
-0.004476258996874094,
-0.003942856099456549,
-0.006440146826207638,
-0.06409024447202682,
-0.049047891050577164,
-0.03757971525192261,
0.0626489445567131,
0.06332807242870331,
-0.007577107287943363,
-0.0007169181481003761,
-0.049479566514492035,
-0.039465729147195816,
-0.05048161372542381,
0.003905148711055517,
0.01432350929826498,
-0.013126297853887081,
0.03156575933098793,
-0.1018715500831604,
-0.04291507974267006,
0.005793066695332527,
0.09170664846897125,
-0.01363244280219078,
-0.04950674623250961,
0.031097203493118286,
0.03390100970864296,
-0.024857809767127037,
-0.04663064703345299,
-0.00555040268227458,
-0.04124820604920387,
-0.04513347148895264,
0.047994788736104965,
0.01211861427873373,
0.10682076215744019,
-0.01773347519338131,
0.05050377547740936,
-0.057133838534355164,
0.03786388784646988,
0.041065607219934464,
0.02563459239900112,
-0.0361974760890007,
-0.03172953426837921,
0.09743357449769974,
0.012504477985203266,
-0.036660823971033096,
-0.02171328291296959,
-0.08493807911872864,
-0.033823076635599136,
0.09085182100534439,
-0.015163491480052471,
-0.01926385797560215,
-0.013863424770534039,
0.05053297430276871,
0.00377785530872643,
-0.11973930150270462,
0.08577919006347656,
-0.08105743676424026,
0.019243741407990456,
-0.020521825179457664,
-0.04281714931130409,
0.05938168242573738,
-0.03470807150006294,
-0.07139767706394196,
-0.06679272651672363,
-0.02920450270175934,
-0.09615496546030045,
-0.03448878228664398,
0.04082169756293297,
0.027659058570861816,
-3.211979906246344e-33,
0.012627189047634602,
-0.011523483321070671,
0.1345752477645874,
0.026306426152586937,
0.00028201533132232726,
-0.02411501668393612,
0.029182037338614464,
0.0038105081766843796,
0.09827285259962082,
-0.04631310701370239,
-0.02405511774122715,
-0.006267940625548363,
0.011977965012192726,
-0.046460796147584915,
0.02025708183646202,
-0.016368677839636803,
-0.06862853467464447,
-0.10899277776479721,
0.019193582236766815,
0.006236425135284662,
-0.030154690146446228,
0.0764937698841095,
-0.01798197440803051,
0.004906758666038513,
-0.029381824657320976,
0.01164157222956419,
-0.025113694369792938,
0.048331186175346375,
-0.008299957029521465,
0.019891083240509033,
-0.08628872036933899,
-0.023740043863654137,
-0.08558019995689392,
0.04215117543935776,
-0.13118809461593628,
0.10488153249025345,
0.06716163456439972,
0.07784675061702728,
-0.011610491201281548,
0.06556762009859085,
-0.02777843549847603,
-0.005280910059809685,
-0.05427122861146927,
0.1273600459098816,
0.026786182075738907,
0.021941035985946655,
-0.036567773669958115,
0.019751757383346558,
0.02418866567313671,
0.027460018172860146,
-0.08420699834823608,
0.03989195078611374,
0.014726024121046066,
-0.03117130883038044,
-0.05265769734978676,
-0.06760307401418686,
-0.005264218896627426,
0.016415443271398544,
0.028017671778798103,
-0.05904724448919296,
-0.06490641087293625,
-0.051481179893016815,
0.014296835288405418,
0.047205809503793716,
-0.020104674622416496,
-0.0056077963672578335,
0.03225264325737953,
0.014470744878053665,
0.02452176809310913,
0.03930385783314705,
-0.02058572508394718,
0.03170384466648102,
0.05991743132472038,
0.025292078033089638,
0.11749345809221268,
-0.05937109887599945,
0.003246485022827983,
-0.043966032564640045,
0.0038915143813937902,
-0.0007906497339718044,
0.0327644906938076,
0.07467199862003326,
0.06970326602458954,
0.03366773948073387,
0.15909910202026367,
0.029131632298231125,
0.006356331519782543,
-0.028275998309254646,
-0.015158128924667835,
0.0009829525370150805,
-0.019916066899895668,
-0.026905566453933716,
0.05879984050989151,
0.04609033837914467,
-0.055613573640584946,
-4.719651869322661e-8,
-0.04322045296430588,
-0.02247810550034046,
0.06256614625453949,
-0.010945041663944721,
0.029981309548020363,
0.05530243739485741,
-0.02369624935090542,
0.04605372995138168,
0.0028238813392817974,
-0.03759360685944557,
0.04443351924419403,
0.024518093094229698,
-0.032626673579216,
-0.07556324452161789,
-0.039121877402067184,
0.0007774767000228167,
0.01033798698335886,
-0.003664948744699359,
-0.0009759601671248674,
-0.028647219762206078,
-0.02083185687661171,
0.0027493522502481937,
-0.014311383478343487,
-0.017327554523944855,
0.09690669178962708,
-0.10797331482172012,
0.00189911387860775,
0.020463399589061737,
0.0448891706764698,
-0.010792690329253674,
-0.006106886547058821,
-0.043997541069984436,
0.06414791196584702,
0.08745281398296356,
0.06518587470054626,
-0.016249025240540504,
-0.02724330499768257,
-0.01119258627295494,
-0.04493386298418045,
0.0024771324824541807,
-0.05108488351106644,
0.1537858247756958,
-0.0799717828631401,
-0.05505499616265297,
0.014703628607094288,
0.10589544475078583,
0.04765603318810463,
-0.0017706904327496886,
-0.011796844191849232,
0.008642788045108318,
0.056237537413835526,
-0.02931186556816101,
-0.00949382409453392,
-0.01382380910217762,
0.051816876977682114,
-0.021911704912781715,
-0.021451283246278763,
-0.06515812128782272,
0.009116173721849918,
-0.010217280127108097,
-0.11412420868873596,
0.03848573938012123,
0.03268356993794441,
-0.04550963267683983
] | 0 | 0 | 0 | 0 |
|
Efficient Sparse PCA via Block-Diagonalization
|
https://openreview.net/forum?id=FAYIlGDBa1
|
[
"Alberto Del Pia",
"Dekun Zhou",
"Yinglun Zhu"
] |
Poster
|
Sparse Principal Component Analysis (Sparse PCA) is a pivotal tool in data analysis and dimensionality reduction. However, Sparse PCA is a challenging problem in both theory and practice: it is known to be NP-hard and current exact methods generally require exponential runtime. In this paper, we propose a novel framework to efficiently approximate Sparse PCA by (i) approximating the general input covariance matrix with a re-sorted block-diagonal matrix, (ii) solving the Sparse PCA sub-problem in each block, and (iii) reconstructing the solution to the original problem. Our framework is simple and powerful: it can leverage any off-the-shelf Sparse PCA algorithm and achieve significant computational speedups, with a minor additive error that is linear in the approximation error of the block-diagonal matrix. Suppose $g(k, d)$ is the runtime of an algorithm (approximately) solving Sparse PCA in dimension $d$ and with sparsity constant $k$. Our framework, when integrated with this algorithm, reduces the runtime to $\mathcal{O}\left(\frac{d}{d^\star} \cdot g(k, d^\star) + d^2\right)$, where $d^\star \leq d$ is the largest block size of the block-diagonal matrix. For instance, integrating our framework with the Branch-and-Bound algorithm reduces the complexity from $g(k, d) = \mathcal{O}(k^3\cdot d^k)$ to $\mathcal{O}(k^3\cdot d \cdot (d^\star)^{k-1})$, demonstrating exponential speedups if $d^\star$ is small. We perform large-scale evaluations on many real-world datasets: for exact Sparse PCA algorithm, our method achieves an average speedup factor of 100.50, while maintaining an average approximation error of 0.61%; for approximate Sparse PCA algorithm, our method achieves an average speedup factor of 6.00 and an average approximation error of -0.91%, meaning that our method oftentimes finds better solutions.
|
Sparse PCA, Block Diagonalization, Compurational Efficiency, Approximation Algorithms
| null | 11,799 |
2410.14092
|
[
-0.07622351497411728,
0.01618083193898201,
0.0092727430164814,
-0.037265509366989136,
0.031055329367518425,
-0.05669155716896057,
-0.045494385063648224,
-0.0401124469935894,
-0.003562597557902336,
0.01760411448776722,
-0.006663210690021515,
0.07341643422842026,
-0.05675886571407318,
0.033305492252111435,
-0.030851436778903008,
0.03151644021272659,
0.07737324386835098,
0.07834520190954208,
0.012768316082656384,
-0.013314077630639076,
-0.04880901426076889,
0.0026499514933675528,
-0.051530152559280396,
0.028757046908140182,
0.019285134971141815,
-0.015185631811618805,
0.07909347862005234,
-0.023984471336007118,
-0.04510650038719177,
0.06069067120552063,
0.013278020545840263,
0.020551446825265884,
-0.011873090639710426,
-0.06613359600305557,
0.011742711067199707,
0.09391384571790695,
0.06803782284259796,
0.03557465970516205,
-0.01294395700097084,
0.053969915956258774,
0.0850081518292427,
0.08643574267625809,
-0.061298783868551254,
0.06411348283290863,
0.04887481778860092,
-0.03260859474539757,
0.014662457630038261,
-0.08604061603546143,
0.05814950913190842,
-0.05717245489358902,
0.0015273650642484426,
0.01278185099363327,
-0.06959077715873718,
-0.01712728478014469,
-0.02959960140287876,
-0.09325553476810455,
0.009496096521615982,
-0.03874830901622772,
0.058432918041944504,
0.021311162039637566,
-0.014207622967660427,
-0.06340894848108292,
0.01996401511132717,
0.019757837057113647,
0.07032143324613571,
-0.015538649633526802,
0.06236817687749863,
0.017125876620411873,
-0.036976903676986694,
0.04266951605677605,
-0.029520384967327118,
0.05164210498332977,
-0.05530616268515587,
0.056008074432611465,
-0.0003832033253274858,
-0.04142938554286957,
-0.06243974715471268,
-0.008178839460015297,
0.06070568785071373,
0.009248957969248295,
0.026130802929401398,
0.06358189135789871,
-0.05966377630829811,
-0.04874158278107643,
0.04492995887994766,
-0.03553679957985878,
-0.04273725673556328,
0.01873181201517582,
0.06875944882631302,
-0.05529385060071945,
0.028997069224715233,
-0.018163563683629036,
0.011416219174861908,
-0.02762805111706257,
0.016130859032273293,
-0.08827327936887741,
0.1107696071267128,
-0.034008827060461044,
-0.004678420256823301,
0.06663072109222412,
-0.01473554689437151,
-0.03181718662381172,
-0.039745476096868515,
0.07095709443092346,
-0.03748001903295517,
-0.0252679530531168,
-0.0013566487468779087,
0.09618458151817322,
0.008596324361860752,
-0.05664995685219765,
0.007383827585726976,
-0.006512535270303488,
-0.03536079078912735,
0.043160248547792435,
0.04244305565953255,
-0.058065373450517654,
0.0728255957365036,
0.0006579106557182968,
0.08069924265146255,
0.04516928642988205,
-0.015747787430882454,
-0.06511552631855011,
0.09169737249612808,
0.0046967072412371635,
0.07869761437177658,
0.03645547479391098,
-0.05880282074213028,
2.602247129383985e-33,
0.10101289302110672,
0.1461867392063141,
0.06262106448411942,
-0.030231062322854996,
-0.0371529646217823,
0.002595883095636964,
0.07025182247161865,
-0.03125984966754913,
0.03845449537038803,
0.11026069521903992,
-0.07201647013425827,
0.010934739373624325,
-0.0029052088502794504,
0.01557628158479929,
-0.023391952738165855,
0.04476612061262131,
0.000315748417051509,
0.05924668535590172,
-0.005351139698177576,
-0.061720110476017,
0.10556560754776001,
0.007714227773249149,
0.06243489682674408,
-0.03094731643795967,
-0.04384654387831688,
-0.11275114119052887,
0.023933690041303635,
-0.12249398231506348,
0.017664043232798576,
-0.0043564289808273315,
-0.05022519826889038,
-0.05390993133187294,
0.0054954588413238525,
0.10814615339040756,
0.0043797921389341354,
-0.03471384197473526,
0.045742664486169815,
-0.10118525475263596,
0.055643677711486816,
-0.038162145763635635,
-0.044325873255729675,
-0.0073557510040700436,
0.04767907038331032,
-0.14855822920799255,
-0.08872411400079727,
-0.010397737845778465,
0.0024114882107824087,
0.006285460665822029,
0.028220729902386665,
-0.013795947656035423,
0.027392612770199776,
0.023252176120877266,
-0.023370344191789627,
0.04528519883751869,
-0.09234818816184998,
-0.07552651315927505,
-0.0038419668562710285,
-0.07146469503641129,
0.09504036605358124,
0.048502739518880844,
0.02352215349674225,
0.007532155141234398,
-0.033783938735723495,
-0.017002418637275696,
-0.006442337296903133,
-0.009726828895509243,
-0.020376130938529968,
0.0066031054593622684,
0.030177175998687744,
-0.04247036948800087,
-0.015966225415468216,
-0.013131871819496155,
-0.017505526542663574,
-0.0336143784224987,
0.0710885152220726,
-0.05295075476169586,
0.08628568798303604,
-0.028333820402622223,
-0.06670619547367096,
-0.02660759538412094,
-0.04022602736949921,
0.027542008087038994,
-0.0003693486796692014,
-0.07462204247713089,
-0.04266257584095001,
0.0003265195991843939,
0.042075205594301224,
-0.0674300342798233,
-0.05000797286629677,
-0.14793746173381805,
-0.06532225757837296,
0.01253930851817131,
0.021292349323630333,
-0.027769440785050392,
0.03274528682231903,
-1.441627741216392e-33,
-0.00021375819051172584,
-0.0057338327169418335,
0.11647573858499527,
0.0016443568747490644,
0.05378173291683197,
-0.13782602548599243,
-0.03967941179871559,
-0.04181018844246864,
0.02989131398499012,
-0.08462449163198471,
-0.04094414785504341,
-0.047588467597961426,
0.015831636264920235,
0.046402882784605026,
0.06380051374435425,
0.10467566549777985,
-0.095369853079319,
0.045279715210199356,
-0.00733009772375226,
0.01028881873935461,
-0.0522625558078289,
-0.03730863332748413,
0.014353211037814617,
-0.02028271183371544,
0.006337964441627264,
0.007894680835306644,
0.02796691283583641,
0.014497918076813221,
0.02024027705192566,
-0.0579044334590435,
-0.06506644189357758,
0.03879791498184204,
-0.12808004021644592,
0.044233519583940506,
-0.06545940786600113,
0.07613960653543472,
-0.05454633757472038,
0.03685873746871948,
-0.027053186669945717,
0.04836627468466759,
0.004284045659005642,
0.0174842681735754,
-0.03309360146522522,
-0.013456197455525398,
0.03613629937171936,
-0.014263848774135113,
-0.004672233480960131,
0.05383142828941345,
-0.0001167882583104074,
0.06616975367069244,
0.030500123277306557,
0.046275071799755096,
0.02532714605331421,
0.044993869960308075,
0.016712792217731476,
0.011029570363461971,
0.035642143338918686,
0.06694362312555313,
0.06315071135759354,
-0.04196799173951149,
0.014812009409070015,
-0.038037221878767014,
-0.052518147975206375,
0.010616098530590534,
0.039888396859169006,
0.03181497007608414,
0.07344266027212143,
-0.015336791053414345,
-0.054409805685281754,
0.042637333273887634,
-0.03106268122792244,
0.030357973650097847,
-0.02439049445092678,
0.007185343187302351,
-0.022822540253400803,
0.005740685388445854,
0.02567196637392044,
-0.014194395393133163,
-0.01255928073078394,
0.05744512751698494,
0.03310302644968033,
0.01470152847468853,
0.04433102533221245,
-0.08367662876844406,
0.034230101853609085,
0.11242203414440155,
0.03352270647883415,
-0.05100442096590996,
0.038374632596969604,
-0.055343419313430786,
-0.05223875120282173,
-0.019314059987664223,
0.12094097584486008,
0.032948266714811325,
0.02142239548265934,
-3.6217873145005797e-8,
-0.029606038704514503,
0.018729545176029205,
-0.04192933067679405,
-0.061321157962083817,
0.08736547827720642,
-0.04831846430897713,
0.013440381735563278,
-0.011351034045219421,
0.022509217262268066,
-0.08590157330036163,
0.09460458904504776,
-0.0773150622844696,
-0.020397625863552094,
-0.00900556892156601,
-0.02335249073803425,
0.08563600480556488,
-0.025256551802158356,
-0.026868512853980064,
-0.030749458819627762,
-0.017021333798766136,
-0.0071565862745046616,
0.026801859959959984,
0.011035566218197346,
0.04280975088477135,
0.011885502375662327,
-0.011640562675893307,
0.028718216344714165,
0.07752959430217743,
0.13056683540344238,
0.025876054540276527,
-0.07118653506040573,
0.019161265343427658,
0.09358610957860947,
-0.03776005282998085,
0.020414113998413086,
0.06238911673426628,
0.0630059465765953,
0.07557518780231476,
-0.07878375053405762,
-0.08034674823284149,
-0.08613564819097519,
-0.029255324974656105,
0.004615096375346184,
-0.045154523104429245,
0.06178855895996094,
0.019207414239645004,
-0.05543006584048271,
-0.007711476180702448,
0.03286394104361534,
0.02854296751320362,
0.005148343276232481,
0.004725961945950985,
0.0022777155973017216,
0.06566266715526581,
0.05217450112104416,
-0.009663905948400497,
-0.019176946952939034,
-0.017703620716929436,
0.09553143382072449,
-0.06291680783033371,
0.0031560719944536686,
-0.028440315276384354,
-0.04995688796043396,
-0.0069204275496304035
] | 0 | 0 | 0 | 0 |
|
The Last Iterate Advantage: Empirical Auditing and Principled Heuristic Analysis of Differentially Private SGD
|
https://openreview.net/forum?id=DwqoBkj2Mw
|
[
"Milad Nasr",
"Thomas Steinke",
"Borja Balle",
"Christopher A. Choquette-Choo",
"Arun Ganesh",
"Matthew Jagielski",
"Jamie Hayes",
"Abhradeep Guha Thakurta",
"Adam Smith",
"Andreas Terzis"
] |
Poster
|
We propose a simple heuristic privacy analysis of noisy clipped stochastic gradient descent (DP-SGD) in the setting where only the last iterate is released and the intermediate iterates remain hidden. Namely, our heuristic assumes a linear structure for the model.
We show experimentally that our heuristic is predictive of the outcome of privacy auditing applied to various training procedures. Thus it can be used prior to training as a rough estimate of the final privacy leakage. We also probe the limitations of our heuristic by providing some artificial counterexamples where it underestimates the privacy leakage.
The standard composition-based privacy analysis of DP-SGD effectively assumes that the adversary has access to all intermediate iterates, which is often unrealistic. However, this analysis remains the state of the art in practice. While our heuristic does not replace a rigorous privacy analysis, it illustrates the large gap between the best theoretical upper bounds and the privacy auditing lower bounds and sets a target for further work to improve the theoretical privacy analyses.
|
differential privacy, heuristics, privacy auditing
|
We propose a heuristic privacy analysis of DP-SGD based on linearization. We compare it to privacy auditing and theoretical upper bounds.
| 11,788 |
2410.06186
|
[
-0.1301041692495346,
-0.021387701854109764,
-0.013127937912940979,
0.0040881740860641,
0.07408184558153152,
0.014571413397789001,
0.04087340086698532,
-0.05683896318078041,
0.02361745946109295,
-0.026645105332136154,
0.03238683566451073,
0.006578863598406315,
0.022684454917907715,
-0.031729117035865784,
-0.016282793134450912,
-0.04834692180156708,
0.017181692644953728,
0.08523886650800705,
-0.050555307418107986,
-0.048437636345624924,
-0.037267446517944336,
-0.002068598521873355,
0.017770804464817047,
0.0018369448371231556,
0.0011520007392391562,
-0.02629137970507145,
0.04366779699921608,
-0.08636689931154251,
0.006419200450181961,
-0.0069105676375329494,
0.09019750356674194,
-0.08491674810647964,
0.00024665877572260797,
0.00882013887166977,
0.03417612239718437,
-0.03545505553483963,
-0.016568243503570557,
-0.011781621724367142,
0.0009874008828774095,
0.021751387044787407,
-0.03166468068957329,
-0.023252641782164574,
-0.028333036229014397,
0.01423337496817112,
-0.04553369805216789,
-0.02242170460522175,
0.05801675468683243,
-0.0025441301986575127,
-0.07345040142536163,
-0.05861969292163849,
-0.035787660628557205,
-0.010083275847136974,
-0.032632213085889816,
-0.04092855006456375,
-0.03862902894616127,
-0.0652843788266182,
0.009648763574659824,
-0.036516256630420685,
-0.023431070148944855,
0.04775319620966911,
-0.031999971717596054,
-0.09167902916669846,
-0.05444839969277382,
0.01971461996436119,
0.056676752865314484,
0.010017596185207367,
0.07803674042224884,
0.0531759075820446,
0.012508581392467022,
0.0316428504884243,
-0.022405680269002914,
0.0574980191886425,
-0.09838060289621353,
0.046464551240205765,
0.07652495056390762,
0.05187787115573883,
0.009994841180741787,
0.0032997687812894583,
-0.03323753550648689,
-0.10786871612071991,
0.0006944319466128945,
0.023806877434253693,
0.07465427368879318,
0.0032601102720946074,
-0.0018819168908521533,
-0.017682697623968124,
-0.0017327115638181567,
-0.011515544727444649,
0.054192956537008286,
0.0335976779460907,
-0.01572204753756523,
-0.0011200145818293095,
-0.10022290050983429,
0.0035742621403187513,
0.04942299425601959,
0.011682958342134953,
-0.025119664147496223,
-0.0065644835121929646,
0.021667130291461945,
0.05442702770233154,
-0.017262287437915802,
0.11892665177583694,
0.0060171992518007755,
-0.0353730134665966,
0.05816550925374031,
0.1046568751335144,
0.04196033999323845,
-0.022307565435767174,
0.044657837599515915,
0.0182096678763628,
0.05153388902544975,
0.04641437903046608,
-0.0079210065305233,
0.001472717267461121,
0.04662908986210823,
-0.04192576929926872,
-0.05643756687641144,
0.013587447814643383,
0.012778658419847488,
0.09660159051418304,
0.04058298468589783,
0.023039693012833595,
0.08478905260562897,
0.034919507801532745,
-0.04078714922070503,
-0.003690674202516675,
-0.0782913863658905,
2.081848066093281e-33,
0.012491538189351559,
0.01736677624285221,
-0.014365245588123798,
-0.0635581985116005,
-0.041186004877090454,
0.026444880291819572,
-0.005240534897893667,
0.04057169333100319,
-0.06755254417657852,
0.05911637097597122,
0.03757840394973755,
0.023684807121753693,
0.010094063356518745,
-0.01102465484291315,
0.056576453149318695,
0.06497850269079208,
0.027418581768870354,
0.07376296073198318,
0.0695817619562149,
-0.05397585779428482,
0.16975831985473633,
-0.0646739974617958,
0.006057750899344683,
-0.01977694220840931,
0.09087826311588287,
-0.07280092686414719,
0.06859210133552551,
-0.005114429630339146,
0.025387786328792572,
0.07058623433113098,
-0.06600866466760635,
0.0610145702958107,
0.022569065913558006,
0.05088319256901741,
0.10082580894231796,
0.03676879405975342,
-0.014293177984654903,
-0.026197848841547966,
0.049088023602962494,
0.005037167575210333,
0.00561484694480896,
-0.033825840801000595,
0.011817632243037224,
-0.017774952575564384,
-0.13646715879440308,
0.0020286752842366695,
-0.04998541623353958,
0.03403402119874954,
-0.03007339872419834,
-0.07854091376066208,
-0.02691655047237873,
0.02212080918252468,
-0.09087489545345306,
-0.006904955487698317,
-0.04160453751683235,
0.033966489136219025,
0.041212309151887894,
0.007610231172293425,
0.01762448623776436,
0.08558908849954605,
0.014387023635208607,
0.11483453959226608,
-0.011642780154943466,
-0.0024754267651587725,
-0.050669826567173004,
-0.007580664940178394,
-0.042647819966077805,
-0.03270372748374939,
0.0715484619140625,
-0.053216204047203064,
0.0037522963248193264,
0.007077708840370178,
-0.04431714117527008,
-0.04619108885526657,
-0.008342365734279156,
-0.06019964814186096,
0.07972351461648941,
0.018214484676718712,
0.04394717887043953,
-0.05556493625044823,
-0.005898358765989542,
0.039819248020648956,
0.0002499037072993815,
0.02318493463099003,
-0.12572669982910156,
-0.04053431376814842,
-0.05088811367750168,
0.021371526643633842,
-0.07541538029909134,
0.06438576430082321,
-0.08141899108886719,
-0.033273275941610336,
0.0011231359094381332,
0.008609820157289505,
-0.05839456617832184,
-4.19941609943019e-33,
-0.1511286497116089,
0.004187768325209618,
-0.043115731328725815,
0.026087846606969833,
0.041879937052726746,
-0.06397487968206406,
-0.03802895545959473,
0.029214706271886826,
0.026780348271131516,
0.022665848955512047,
-0.08058873564004898,
0.07367784529924393,
0.03627622872591019,
-0.026539338752627373,
-0.0018446787726134062,
0.04821110516786575,
0.018068406730890274,
0.04729423671960831,
-0.028603224083781242,
0.05065509304404259,
0.002947087399661541,
0.07321149855852127,
-0.019588444381952286,
0.05191069096326828,
0.06851062923669815,
-0.012691277079284191,
-0.09093979746103287,
0.07051204144954681,
0.058598898351192474,
-0.005622187163680792,
-0.01665649376809597,
-0.026558876037597656,
-0.10369888693094254,
-0.06574858725070953,
-0.061394933611154556,
-0.034540701657533646,
-0.04147389903664589,
0.038282256573438644,
0.028730865567922592,
0.05609835684299469,
0.035891249775886536,
-0.0033773158211261034,
-0.045409396290779114,
-0.055060870945453644,
0.06855795532464981,
-0.008314709179103374,
-0.06188935041427612,
0.05456675961613655,
0.048109568655490875,
0.001139541040174663,
0.0263315811753273,
0.025758489966392517,
0.04242327809333801,
0.06111525371670723,
-0.09576427191495895,
0.03147396072745323,
0.058848679065704346,
-0.009754148311913013,
0.07302264124155045,
0.050156593322753906,
0.010248101316392422,
0.01719934307038784,
-0.043257132172584534,
0.0454801507294178,
-0.00062334886752069,
-0.10986623913049698,
-0.03604162111878395,
-0.03396875038743019,
-0.025329770520329475,
0.011693793348968029,
0.02882741391658783,
-0.07566778361797333,
-0.07646151632070541,
0.03328939899802208,
-0.005075883585959673,
0.07552129030227661,
-0.0005844995030201972,
-0.0848735123872757,
-0.03212997317314148,
0.06513553857803345,
0.056939128786325455,
0.00481413584202528,
0.027605226263403893,
-0.09577666968107224,
0.055403199046850204,
-0.05398615077137947,
0.1280021369457245,
-0.08390099555253983,
-0.0007462401408702135,
0.0025717434473335743,
-0.017833296209573746,
-0.03851023688912392,
-0.08118483424186707,
-0.028184663504362106,
-0.03839031234383583,
-4.247129581358422e-8,
0.03921608626842499,
-0.05089769512414932,
0.07668246328830719,
0.07005178183317184,
0.09418920427560806,
-0.024079345166683197,
-0.013975018635392189,
0.023662399500608444,
-0.06782494485378265,
-0.040981192141771317,
0.07881469279527664,
-0.016864966601133347,
-0.03266074135899544,
-0.04107922688126564,
-0.0448613315820694,
0.03334163874387741,
-0.08031629770994186,
-0.03747786208987236,
-0.049198076128959656,
0.0067995390854775906,
-0.061379190534353256,
0.004527247045189142,
-0.02893868274986744,
-0.05275176465511322,
-0.009998603723943233,
-0.0296329315751791,
0.04078150913119316,
0.013594219461083412,
0.10060208290815353,
0.09135495126247406,
-0.02474306710064411,
-0.02659759670495987,
-0.009015452116727829,
0.0030078620184212923,
-0.08798863738775253,
0.13550278544425964,
-0.036726512014865875,
-0.037151649594306946,
0.039729807525873184,
0.015506532974541187,
0.011256049387156963,
-0.002847724361345172,
0.07954292744398117,
0.05749380588531494,
-0.06064441800117493,
-0.05620991438627243,
0.012796297669410706,
-0.09459077566862106,
0.007293453440070152,
-0.057387709617614746,
-0.008640432730317116,
-0.07433759421110153,
-0.04197443649172783,
0.08736059069633484,
0.1091604009270668,
-0.0016293305670842528,
0.008702114224433899,
-0.06332852691411972,
-0.021867554634809494,
0.08537040650844574,
0.009502009488642216,
0.009557045996189117,
-0.07777921110391617,
0.030420519411563873
] | 0 | 0 | 0 | 0 |
|
Trivialized Momentum Facilitates Diffusion Generative Modeling on Lie Groups
|
https://openreview.net/forum?id=DTatjJTDl1
|
[
"Yuchen Zhu",
"Tianrong Chen",
"Lingkai Kong",
"Evangelos Theodorou",
"Molei Tao"
] |
Poster
|
The generative modeling of data on manifolds is an important task, for which diffusion models in flat spaces typically need nontrivial adaptations. This article demonstrates how a technique called `trivialization' can transfer the effectiveness of diffusion models in Euclidean spaces to Lie groups. In particular, an auxiliary momentum variable was algorithmically introduced to help transport the position variable between data distribution and a fixed, easy-to-sample distribution. Normally, this would incur further difficulty for manifold data because momentum lives in a space that changes with the position. However, our trivialization technique creates a new momentum variable that stays in a simple fixed vector space. This design, together with a manifold preserving integrator, simplifies implementation and avoids inaccuracies created by approximations such as projections to tangent space and manifold, which were typically used in prior work, hence facilitating generation with high-fidelity and efficiency. The resulting method achieves state-of-the-art performance on protein and RNA torsion angle generation and sophisticated torus datasets. We also, arguably for the first time, tackle the generation of data on high-dimensional Special Orthogonal and Unitary groups, the latter essential for quantum problems. Code is available at https://github.com/yuchen-zhu-zyc/TDM.
|
non-Euclidean generative modeling, denoising diffusion, Lie group
| null | 11,776 |
2405.16381
|
[
-0.11636043339967728,
-0.02571621537208557,
0.030413398519158363,
-0.02844221517443657,
-0.04361926391720772,
0.0069134910590946674,
-0.06330211460590363,
-0.039675842970609665,
-0.026586541905999184,
-0.0474654883146286,
0.11410054564476013,
0.003575590904802084,
-0.03696950897574425,
0.030777985230088234,
0.005869080312550068,
-0.05643165484070778,
-0.009178119711577892,
0.15394316613674164,
-0.113543801009655,
-0.054242298007011414,
-0.07092448323965073,
0.10511808842420578,
0.004299748223274946,
0.03352881595492363,
0.056623898446559906,
0.022576965391635895,
0.09286084771156311,
-0.06251301616430283,
0.06598899513483047,
-0.07054857909679413,
0.12606477737426758,
-0.04212382435798645,
-0.07841455191373825,
-0.03536144644021988,
-0.031285181641578674,
0.04777159541845322,
-0.0046492028050124645,
0.012813504785299301,
-0.0230002049356699,
-0.0031084343791007996,
0.06524011492729187,
0.06464049965143204,
0.01917736604809761,
0.06312942504882812,
0.039377957582473755,
0.024645093828439713,
0.025699930265545845,
-0.022473864257335663,
0.025020519271492958,
-0.01747352071106434,
-0.016419505700469017,
-0.0036449909675866365,
-0.011893422342836857,
0.02342059090733528,
-0.01879296638071537,
-0.014488746412098408,
0.007871228270232677,
0.002249383367598057,
0.0023034117184579372,
-0.08893334865570068,
0.009863513521850109,
-0.04037201404571533,
0.0009226828115060925,
-0.008034802041947842,
-0.04178757220506668,
-0.02687818743288517,
0.06090177968144417,
0.08929018676280975,
0.021116573363542557,
0.004965829197317362,
-0.03544007986783981,
0.01997224986553192,
-0.15353642404079437,
0.07106669992208481,
0.021824322640895844,
-0.03932647779583931,
-0.03101464733481407,
0.015042156912386417,
-0.0048985895700752735,
-0.0059634000062942505,
0.04978552833199501,
-0.02466282807290554,
0.02342216670513153,
0.028057562187314034,
0.010989847593009472,
0.050010617822408676,
-0.033535633236169815,
0.11517611145973206,
-0.011082309298217297,
-0.05632143095135689,
-0.0077940174378454685,
-0.00013122324889991432,
-0.0443904809653759,
-0.04184691607952118,
0.04450578987598419,
0.024990549311041832,
0.04588960483670235,
0.016379909589886665,
0.009101342409849167,
0.06233521178364754,
0.0617896169424057,
-0.09533224999904633,
-0.0011492955964058638,
0.03397023677825928,
0.05122283473610878,
-0.09410198032855988,
0.0285047460347414,
0.0013602437684312463,
-0.011321443133056164,
0.02191028743982315,
0.008164756931364536,
0.031401243060827255,
-0.009161748923361301,
-0.03502823784947395,
0.0021130782552063465,
-0.1674794852733612,
0.0038011723663657904,
0.05859510228037834,
0.02272789180278778,
0.021440906450152397,
0.012438029982149601,
-0.0066635641269385815,
-0.02670116350054741,
0.09477370977401733,
-0.019190939143300056,
0.0208202563226223,
-0.10860337316989899,
5.559816905498273e-34,
0.01928958110511303,
0.05136067420244217,
0.015917805954813957,
0.09018360823392868,
-0.023125585168600082,
-0.00475231371819973,
0.03200724720954895,
0.00777716189622879,
0.02367144636809826,
-0.009085923433303833,
-0.13158346712589264,
-0.024498699232935905,
-0.060523007065057755,
0.015817634761333466,
-0.00735125970095396,
0.002546285977587104,
0.0008816625340841711,
0.08873886615037918,
0.014685438945889473,
-0.06402575969696045,
-0.005742084700614214,
0.06590716540813446,
-0.0002550280769355595,
-0.025467649102211,
-0.023629628121852875,
0.05157386511564255,
0.0333881638944149,
-0.0896439254283905,
-0.0992787703871727,
0.002431311411783099,
-0.069058358669281,
-0.03270011395215988,
-0.09124163538217545,
0.06764715909957886,
0.06369081139564514,
0.024946657940745354,
-0.0023187792394310236,
-0.08088009804487228,
-0.025465406477451324,
0.002365780295804143,
-0.05218890309333801,
0.04709919914603233,
-0.00910585094243288,
-0.02769339457154274,
-0.12210658192634583,
0.03951004892587662,
0.05728485435247421,
0.04510537534952164,
0.054071925580501556,
-0.06670208275318146,
0.04689269885420799,
-0.007645329460501671,
0.0064318059012293816,
-0.0008222095784731209,
-0.02165229059755802,
-0.055239152163267136,
-0.018046364188194275,
-0.04430495575070381,
-0.03807470574975014,
0.017644096165895462,
0.02703452669084072,
0.08578120172023773,
0.018658066168427467,
-0.027710698544979095,
-0.016077537089586258,
-0.0589420385658741,
-0.03070615977048874,
0.036318741738796234,
0.03126358985900879,
0.030030302703380585,
-0.05911416932940483,
0.028597502037882805,
-0.0682220309972763,
-0.04964828118681908,
0.05845613032579422,
0.03499772399663925,
0.04796631634235382,
-0.04465639591217041,
-0.03258538991212845,
0.011700154282152653,
-0.03398815914988518,
-0.025260750204324722,
-0.044028140604496,
-0.003928465768694878,
-0.0038343630731105804,
-0.062026940286159515,
0.020185206085443497,
-0.05312801152467728,
-0.046086475253105164,
-0.04380692541599274,
-0.010217411443591118,
-0.04268667474389076,
-0.01136455312371254,
0.03475396707653999,
-0.03465227410197258,
-1.3742503335726777e-33,
-0.06568276137113571,
0.034204576164484024,
0.03552693501114845,
0.04349369928240776,
-0.020118718966841698,
-0.017252514138817787,
0.014885324984788895,
-0.011941753327846527,
-0.010416802018880844,
-0.059407953172922134,
-0.08571500331163406,
0.04311387613415718,
0.08204877376556396,
0.004625086672604084,
0.03945481777191162,
0.03678591921925545,
-0.040131572633981705,
-0.09159622341394424,
-0.03515911474823952,
0.06683225929737091,
-0.03221924230456352,
0.03353486955165863,
-0.050470005720853806,
-0.003615342313423753,
-0.02945544943213463,
0.06781934201717377,
-0.014264288358390331,
0.004292815458029509,
0.03605201095342636,
0.015630312263965607,
0.04699641838669777,
-0.03845297917723656,
-0.10204024612903595,
-0.04469132795929909,
-0.1385839283466339,
0.1132606491446495,
-0.005272111855447292,
0.11436113715171814,
-0.029010850936174393,
0.003621028270572424,
-0.044321458786726,
0.001403476344421506,
-0.03592367097735405,
0.10699795186519623,
0.02658187970519066,
0.011002113111317158,
0.061541847884655,
-0.04269374534487724,
-0.051014501601457596,
0.09518234431743622,
0.01380287203937769,
0.12166295945644379,
0.026652801781892776,
0.04710964113473892,
-0.045593295246362686,
-0.008755426853895187,
-0.08401960134506226,
-0.05211573839187622,
0.05850735306739807,
-0.019497042521834373,
-0.05533165484666824,
-0.11731449514627457,
0.029111620038747787,
0.0011763391084969044,
-0.08357731997966766,
-0.00539127504453063,
-0.016610877588391304,
-0.05454788729548454,
-0.07556357979774475,
0.08832299709320068,
0.0007913947338238358,
-0.03864849731326103,
0.050347574055194855,
-0.018636047840118408,
0.0392906479537487,
-0.018520822748541832,
0.08742464333772659,
-0.029782235622406006,
0.008847621269524097,
-0.024072445929050446,
0.059087228029966354,
0.010142456740140915,
0.015888549387454987,
-0.021573064848780632,
0.09787960350513458,
0.03522692248225212,
0.0491892471909523,
0.01134943962097168,
0.020815765485167503,
-0.003842021105811,
0.02587888203561306,
0.021929681301116943,
0.022329282015562057,
0.15479029715061188,
-0.021490013226866722,
-4.4877950244881504e-8,
-0.030387740582227707,
-0.032044656574726105,
-0.0026658435817807913,
0.003912640735507011,
-0.017441630363464355,
0.020911982282996178,
0.005874700844287872,
0.06560740619897842,
-0.012767935171723366,
-0.02771560475230217,
0.025741096585989,
-0.07589732855558395,
-0.04068085923790932,
-0.038400668650865555,
0.07096075266599655,
0.10003940761089325,
-0.00930436234921217,
0.014725394546985626,
-0.0083153136074543,
-0.03092694655060768,
0.030440596863627434,
0.08269725739955902,
-0.04304715245962143,
0.0031284992583096027,
0.06526444107294083,
-0.07457070797681808,
0.010110592469573021,
-0.007519703824073076,
0.1052950769662857,
-0.03222011402249336,
0.019456051290035248,
0.011126364581286907,
0.011666434817016125,
0.07510414719581604,
-0.023879975080490112,
0.03775889053940773,
-0.0038557439111173153,
0.039943236857652664,
-0.01042612362653017,
-0.013252165168523788,
-0.06117250770330429,
0.08737440407276154,
-0.02823549695312977,
0.0587909072637558,
0.028223751112818718,
0.1004236564040184,
0.0875229686498642,
0.03528920188546181,
0.013038268312811852,
0.07391702383756638,
0.024473562836647034,
0.0035045123659074306,
0.004976062104105949,
0.022015955299139023,
0.009291322901844978,
0.09448743611574173,
-0.027124648913741112,
0.007204524707049131,
0.024349341168999672,
-0.06542561948299408,
-0.08059456199407578,
-0.00895325280725956,
-0.08680989593267441,
-0.02270595170557499
] | 0 | 0 | 0 | 0 |
|
GeoLoRA: Geometric integration for parameter efficient fine-tuning
|
https://openreview.net/forum?id=bsFWJ0Kget
|
[
"Steffen Schotthöfer",
"Emanuele Zangrando",
"Gianluca Ceruti",
"Francesco Tudisco",
"Jonas Kusch"
] |
Poster
|
Low-Rank Adaptation (LoRA) has become a widely used method for parameter-efficient fine-tuning of large-scale, pre-trained neural networks. However, LoRA and its extensions face several challenges, including the need for rank adaptivity, robustness, and computational efficiency during the fine-tuning process. We introduce GeoLoRA, a novel approach that addresses these limitations by leveraging dynamical low-rank approximation theory. GeoLoRA requires only a single backpropagation pass over the small-rank adapters, significantly reducing computational cost as compared to similar dynamical low-rank training methods and making it faster than popular baselines such as AdaLoRA. This allows GeoLoRA to efficiently adapt the allocated parameter budget across the model, achieving smaller low-rank adapters compared to heuristic methods like AdaLoRA and LoRA, while maintaining critical convergence, descent, and error-bound theoretical guarantees. The resulting method is not only more efficient but also more robust to varying hyperparameter settings. We demonstrate the effectiveness of GeoLoRA on several state-of-the-art benchmarks, showing that it outperforms existing methods in both
accuracy and computational efficiency
|
Low Rank, Finetuning, Robustness, Rank Adaptive, LoRA, Low Rank Adapter
| null | 11,771 |
2410.18720
|
[
-0.06632184237241745,
-0.09302546828985214,
0.02264726720750332,
0.06614759564399719,
0.008499525487422943,
0.0312294103205204,
-0.009504984132945538,
0.007304779719561338,
-0.0602303184568882,
-0.108281709253788,
-0.08860072493553162,
-0.032940369099378586,
0.0501575730741024,
0.023574352264404297,
-0.06637893617153168,
0.026756947860121727,
0.062396347522735596,
0.14402197301387787,
-0.10388913005590439,
-0.02933228574693203,
-0.03513466939330101,
-0.01080192532390356,
0.03915608301758766,
0.042658016085624695,
0.06373331695795059,
-0.027854835614562035,
-0.013341374695301056,
-0.022191662341356277,
-0.01032975409179926,
-0.05133341625332832,
0.020113229751586914,
0.011006208136677742,
-0.020516229793429375,
-0.013589190319180489,
-0.12167774885892868,
0.008113497868180275,
-0.061877038329839706,
-0.009242000989615917,
0.026109574362635612,
-0.00022949233243707567,
0.03185190632939339,
0.042540132999420166,
-0.005688714329153299,
-0.0015453299274668097,
0.07178756594657898,
0.001982190879061818,
0.04871132969856262,
-0.06468554586172104,
0.006837529595941305,
-0.011296454817056656,
-0.02275315672159195,
0.0000603126609348692,
-0.08857045322656631,
0.016219673678278923,
-0.03729426488280296,
0.02411026693880558,
0.05481592193245888,
-0.030564580112695694,
0.0528634749352932,
-0.028677474707365036,
-0.0013716933317482471,
-0.07289977371692657,
0.016285695135593414,
-0.05582873150706291,
0.03436082601547241,
-0.08791600912809372,
0.09036444127559662,
0.008971299976110458,
0.0021854517981410027,
-0.013161266222596169,
-0.014376315288245678,
0.09976744651794434,
-0.038632359355688095,
0.03530968725681305,
0.02577458694577217,
0.042130183428525925,
0.08242064714431763,
0.04491560161113739,
0.050021592527627945,
0.01677737943828106,
0.047304898500442505,
0.05188971757888794,
0.013693364337086678,
0.0013536160113289952,
0.14045172929763794,
0.036819834262132645,
-0.020332137122750282,
-0.022348560392856598,
0.06638944894075394,
-0.04789837449789047,
0.06759617477655411,
0.02380715310573578,
-0.025503646582365036,
-0.024551240727305412,
-0.025731438770890236,
0.009783868677914143,
0.0011522665154188871,
-0.08804100006818771,
-0.07334066182374954,
0.12369806319475174,
0.0017629704670980573,
-0.006072214338928461,
-0.013268708251416683,
-0.0062832944095134735,
0.03759453818202019,
-0.043625570833683014,
0.08478926122188568,
0.06713823229074478,
0.04365244507789612,
-0.09901940077543259,
-0.017063410952687263,
0.001697832252830267,
0.010144127532839775,
0.04915604740381241,
-0.07781007885932922,
0.0644039660692215,
0.006398530211299658,
-0.068876251578331,
0.0377189964056015,
0.04918283969163895,
-0.08769025653600693,
-0.007222317159175873,
-0.013115991838276386,
0.07353302091360092,
-0.00028772582300007343,
0.04126148298382759,
-0.07660705596208572,
4.278144466234528e-33,
0.037204254418611526,
0.06478523463010788,
0.08426769822835922,
-0.05282171443104744,
-0.023910580202937126,
-0.013060751371085644,
0.040726449340581894,
-0.0396280400454998,
-0.03278180584311485,
-0.02363310381770134,
-0.12641160190105438,
0.0006083005573600531,
-0.0512818917632103,
0.07354717701673508,
0.054281797260046005,
-0.014931818470358849,
0.013637552037835121,
0.008121744729578495,
-0.05220364034175873,
-0.04548434168100357,
0.045040931552648544,
-0.020909372717142105,
0.011504641734063625,
-0.04660801216959953,
0.012371699325740337,
0.06528299301862717,
0.02123979665338993,
0.028867557644844055,
-0.05115151032805443,
0.022253843024373055,
-0.007169532589614391,
-0.06797746568918228,
-0.00647188164293766,
0.040914397686719894,
0.04442689195275307,
-0.0695885643362999,
-0.051467496901750565,
0.0019278050167486072,
0.021639324724674225,
-0.05338595435023308,
-0.07060100883245468,
0.04888878017663956,
-0.004436181392520666,
-0.000588455586694181,
-0.02942700870335102,
-0.07648587226867676,
0.0027823192067444324,
0.10176576673984528,
-0.042022205889225006,
0.005192860029637814,
-0.041408903896808624,
0.029750533401966095,
-0.06878030300140381,
-0.015117446891963482,
-0.06776060163974762,
-0.05525563284754753,
0.10077719390392303,
0.03392399474978447,
0.01153695397078991,
0.05485834181308746,
-0.0032652567606419325,
0.03271118178963661,
-0.08189424127340317,
0.03868034854531288,
0.04478303715586662,
-0.04166274890303612,
0.010142548941075802,
0.058365218341350555,
0.05673062056303024,
-0.007343478966504335,
0.014362160116434097,
0.01682421751320362,
-0.02257491461932659,
0.02392919361591339,
0.049791667610406876,
-0.06148757413029671,
0.08246355503797531,
-0.09936806559562683,
0.02590782940387726,
0.03972713276743889,
-0.010197706520557404,
0.08630630373954773,
0.058560971170663834,
-0.04357438534498215,
-0.05304638668894768,
-0.08901086449623108,
0.052036549896001816,
-0.020957302302122116,
-0.016847580671310425,
-0.017368705943226814,
-0.07794403284788132,
-0.03006620891392231,
-0.010345177724957466,
-0.08089623600244522,
0.003786686575040221,
-2.1612084447821784e-33,
-0.0010453762952238321,
0.0016275985399261117,
0.0207455363124609,
0.021324818953871727,
-0.011487539857625961,
-0.02732473611831665,
-0.04474186897277832,
0.029950303956866264,
-0.05678447708487511,
-0.0035799844190478325,
-0.03327488899230957,
-0.012562183663249016,
0.04678775742650032,
-0.002001254353672266,
0.12529130280017853,
0.03141477704048157,
-0.13546207547187805,
-0.020887644961476326,
-0.013788901269435883,
0.007776414975523949,
0.02746894396841526,
0.10455957800149918,
-0.026979411020874977,
0.04541220888495445,
-0.09674623608589172,
-0.030742531642317772,
-0.013313867151737213,
0.07925192266702652,
-0.031307466328144073,
-0.009235808625817299,
-0.04659447818994522,
0.022026576101779938,
-0.04206211119890213,
0.035334669053554535,
-0.0319046787917614,
0.02928559109568596,
0.026879126206040382,
-0.06730614602565765,
-0.017613044008612633,
0.1179490014910698,
0.0027329556178301573,
-0.011884750798344612,
0.012525658123195171,
-0.01229097880423069,
0.009826451539993286,
0.061852917075157166,
-0.07100718468427658,
0.004041311796754599,
-0.03895166516304016,
0.0752280056476593,
-0.004535660147666931,
-0.007908041588962078,
-0.02085377648472786,
0.036763519048690796,
0.0003816571261268109,
-0.07410721480846405,
-0.019951922819018364,
0.027144961059093475,
0.06664741039276123,
-0.023322321474552155,
0.02897496707737446,
-0.03687308728694916,
0.01919623091816902,
-0.050801269710063934,
-0.06751283258199692,
0.045669715851545334,
0.012037741020321846,
0.004228983074426651,
0.04116783291101456,
0.08138648420572281,
-0.024035127833485603,
-0.04151785373687744,
0.013651796616613865,
-0.04198112711310387,
-0.1277962028980255,
-0.04042786359786987,
-0.029232678934931755,
0.03362207114696503,
-0.014792350120842457,
-0.10168427228927612,
0.04276435449719429,
0.033115845173597336,
0.042657285928726196,
0.029006198048591614,
0.12107530981302261,
0.08464638888835907,
0.08044946193695068,
0.07554902136325836,
0.02914741262793541,
0.01501784473657608,
-0.005043708253651857,
-0.009187222458422184,
0.032106585800647736,
0.055235572159290314,
-0.08131594955921173,
-4.2667352317948826e-8,
-0.08510778099298477,
0.05611446499824524,
-0.009011001326143742,
0.107609324157238,
0.003846517065539956,
0.010919219814240932,
0.028577502816915512,
0.058573901653289795,
-0.0036548590287566185,
0.0010027630487456918,
0.02892056107521057,
-0.027097364887595177,
0.004919150844216347,
-0.041264377534389496,
0.04272666200995445,
0.04110296443104744,
-0.014060208573937416,
0.017474403604865074,
-0.02384183555841446,
0.019429808482527733,
-0.02774147316813469,
0.12103482335805893,
0.016965748742222786,
-0.04276565462350845,
0.1111379861831665,
-0.09678568691015244,
-0.013775411993265152,
0.1147705689072609,
0.13167421519756317,
-0.047662004828453064,
0.054881684482097626,
-0.006047306582331657,
0.05168052017688751,
-0.01750246435403824,
-0.039467230439186096,
0.09677939862012863,
-0.0036651804111897945,
0.013974142260849476,
-0.05623409524559975,
-0.012068048119544983,
-0.00019712392531801015,
0.10771708935499191,
-0.043544966727495193,
-0.0011524074943736196,
0.020161757245659828,
0.019476421177387238,
0.02201159857213497,
-0.09289499372243881,
0.05936167389154434,
-0.0037560181226581335,
0.02495962753891945,
-0.05339952930808067,
-0.032744910567998886,
0.04117178916931152,
0.015747765079140663,
0.012883157469332218,
-0.02632278949022293,
-0.050052087754011154,
0.01329178549349308,
0.036579571664333344,
-0.054356567561626434,
-0.07760128378868103,
-0.07643837481737137,
0.02606401965022087
] | 0 | 0 | 0 | 0 |
|
Support is All You Need for Certified VAE Training
|
https://openreview.net/forum?id=oZkqkkvdND
|
[
"Changming Xu",
"Debangshu Banerjee",
"Deepak Vasisht",
"Gagandeep Singh"
] |
Poster
|
Variational Autoencoders (VAEs) have become increasingly popular and deployed in safety-critical applications. In such applications, we want to give certified probabilistic guarantees on performance under adversarial attacks. We propose a novel method, CIVET, for certified training of VAEs. CIVET depends on the key insight that we can bound worst-case VAE error by bounding the error on carefully chosen support sets at the latent layer. We show this point mathematically and present a novel training algorithm utilizing this insight. We show in an extensive evaluation across different datasets (in both the wireless and vision application areas), architectures, and perturbation magnitudes that our method outperforms SOTA methods achieving good standard performance with strong robustness guarantees.
|
Certified Training, Trustworthy Machine Learning, Variational Autoencoder, Wireless
|
Certified Training of VAEs by support set bounding of the latent layer.
| 11,759 | null |
[
-0.08149644732475281,
-0.02297731675207615,
0.035720642656087875,
0.017995551228523254,
0.0646127387881279,
0.05250277742743492,
0.03541598469018936,
-0.04061257839202881,
-0.05391845852136612,
-0.05987432971596718,
-0.006028345320373774,
-0.061496540904045105,
0.01758226566016674,
-0.017416641116142273,
-0.06868544220924377,
-0.025435034185647964,
0.09875870496034622,
0.07639718800783157,
-0.09768996387720108,
0.015361307188868523,
-0.018233295530080795,
0.0882696881890297,
-0.0020585579331964254,
0.02987498976290226,
0.07371218502521515,
-0.06006459891796112,
0.030482852831482887,
-0.016429055482149124,
0.041400667279958725,
-0.05166943371295929,
0.042988620698451996,
-0.04989510029554367,
-0.10604044049978256,
0.04690922424197197,
-0.09694525599479675,
-0.0016237973468378186,
-0.10780450701713562,
0.004926975350826979,
0.028464751318097115,
0.014417722821235657,
-0.04977338761091232,
0.031921133399009705,
-0.052152905613183975,
0.0492069274187088,
-0.014180527999997139,
-0.008838706649839878,
0.015356289222836494,
-0.053938981145620346,
-0.0009046313352882862,
-0.03591678664088249,
0.025973442941904068,
0.021580994129180908,
0.006758349481970072,
0.017886698246002197,
-0.09360923618078232,
-0.07888805866241455,
0.026425404474139214,
0.01995820552110672,
0.00806907843798399,
0.07862083613872528,
0.06691017001867294,
-0.07011973857879639,
-0.019757872447371483,
-0.0355403795838356,
0.004452916327863932,
-0.05866469442844391,
0.07889813184738159,
0.000601809355430305,
0.024783015251159668,
-0.04125427454710007,
-0.011796959675848484,
0.0684640035033226,
-0.05908812955021858,
0.08795590698719025,
0.05505827069282532,
0.11296461522579193,
0.025202708318829536,
-0.06187344342470169,
0.07622729241847992,
-0.06962186098098755,
0.018451422452926636,
0.004282115958631039,
0.011348119005560875,
0.04207846522331238,
0.1488834023475647,
0.04048876836895943,
-0.047134071588516235,
0.041976556181907654,
0.08844602853059769,
-0.010685590095818043,
-0.030531033873558044,
-0.01314876601099968,
-0.0070726596750319,
0.061228688806295395,
0.04113766551017761,
-0.020495165139436722,
0.022289199754595757,
-0.061742059886455536,
0.01760454662144184,
0.10477651655673981,
-0.01565985195338726,
-0.02799919992685318,
0.017077108845114708,
0.0005091872881166637,
0.11075622588396072,
-0.041817355901002884,
0.12742534279823303,
0.018352704122662544,
0.05218218266963959,
-0.0660521537065506,
0.04880357161164284,
0.04044795408844948,
-0.015465437434613705,
-0.07071937620639801,
0.0202548298984766,
0.036330536007881165,
-0.04372061416506767,
0.024758007377386093,
-0.012239490635693073,
0.012167070992290974,
-0.08645845204591751,
-0.019939644262194633,
0.00646055955439806,
-0.02104925364255905,
0.05330072343349457,
-0.07922682911157608,
0.006197371985763311,
1.9610123896165347e-33,
-0.00801046658307314,
-0.001714805024676025,
0.046440836042165756,
-0.04272245988249779,
0.02395370788872242,
-0.021162860095500946,
0.023375511169433594,
-0.003612176515161991,
-0.048395875841379166,
0.0015918478602543473,
-0.0493924617767334,
-0.049048714339733124,
-0.11023633182048798,
0.053653497248888016,
0.11157672107219696,
-0.002230512909591198,
-0.046907179057598114,
-0.03960805758833885,
-0.018635837361216545,
-0.012464205734431744,
0.11059446632862091,
-0.11645816266536713,
-0.03298896923661232,
-0.016387511044740677,
-0.06100234016776085,
-0.02772727981209755,
0.0651361346244812,
0.028993815183639526,
-0.010194646194577217,
0.0235256589949131,
-0.10750018805265427,
0.032219335436820984,
0.07484953850507736,
-0.016505610197782516,
0.03882026672363281,
-0.0058370959013700485,
-0.0246297437697649,
0.02489541471004486,
-0.030187757685780525,
-0.021459007635712624,
0.0414075031876564,
0.07254312187433243,
0.014056622982025146,
-0.09287838637828827,
-0.007182334084063768,
-0.06092538312077522,
-0.0042658415623009205,
0.08258908987045288,
-0.03822571039199829,
-0.07715490460395813,
0.00922349188476801,
0.002615934470668435,
-0.04844958707690239,
0.009856934659183025,
-0.018364649266004562,
0.0037226639688014984,
0.03284769877791405,
0.07652208209037781,
0.08070722967386246,
0.03850762918591499,
0.012090880423784256,
-0.019489020109176636,
0.0032099708914756775,
0.0017932986374944448,
-0.03168412670493126,
-0.05053504928946495,
-0.003720030188560486,
-0.038848407566547394,
-0.04578150436282158,
0.05157246068120003,
-0.06886988133192062,
0.018315505236387253,
-0.02519218809902668,
-0.07616492360830307,
0.0011643745237961411,
-0.022489115595817566,
0.0060622780583798885,
0.010516021400690079,
0.012084156274795532,
-0.0029637522529810667,
0.00562122231349349,
0.06493251025676727,
-0.003262573154643178,
0.04083642736077309,
-0.07698895782232285,
-0.04521259292960167,
-0.016980871558189392,
-0.04724622145295143,
-0.05631288141012192,
0.04831715673208237,
-0.04828730598092079,
-0.033593714237213135,
0.03337498754262924,
-0.007837247103452682,
-0.022211354225873947,
-1.8677207498314624e-33,
-0.0960157960653305,
0.07930953800678253,
-0.011019336991012096,
0.062330275774002075,
0.010304438881576061,
-0.05639158561825752,
-0.0047801402397453785,
0.10363895446062088,
-0.045407988131046295,
-0.004042489919811487,
-0.03357718139886856,
-0.006812583189457655,
0.03623458743095398,
-0.016389068216085434,
-0.0042746951803565025,
0.0019101541256532073,
-0.07732199877500534,
0.06171293184161186,
0.05700905993580818,
0.02798192948102951,
0.09501390904188156,
0.06271448731422424,
-0.005931421648710966,
0.04866769537329674,
0.06899100542068481,
-0.014330165460705757,
-0.0327640026807785,
0.10035021603107452,
0.014633896760642529,
-0.0019045808585360646,
-0.0016253860667347908,
-0.0403897650539875,
-0.029155245050787926,
-0.04318707436323166,
-0.016937972977757454,
-0.02999313734471798,
0.10273826122283936,
0.054841600358486176,
0.04410064220428467,
0.0663386881351471,
0.03491433709859848,
-0.017385048791766167,
-0.03940200433135033,
0.024007324129343033,
0.02226073667407036,
-0.03777875751256943,
0.01695750467479229,
-0.03312433883547783,
0.00403404189273715,
-0.0032574485521763563,
0.0261535607278347,
0.0029934069607406855,
-0.0701105147600174,
0.07562768459320068,
-0.001691692741587758,
-0.027526067569851875,
0.05124979466199875,
-0.03070160001516342,
0.07491575926542282,
0.05616828799247742,
-0.09283849596977234,
-0.01936044543981552,
0.02718360163271427,
-0.05691555142402649,
-0.015041718259453773,
-0.09377118945121765,
-0.0964968129992485,
0.02812102809548378,
0.033585742115974426,
0.031065642833709717,
0.042271289974451065,
-0.009716005995869637,
-0.045641813427209854,
0.033305082470178604,
0.009530291892588139,
-0.07551944255828857,
0.0697721540927887,
-0.03965670242905617,
0.04152316227555275,
-0.02714422345161438,
0.0031648543663322926,
-0.04242847114801407,
0.07573113590478897,
0.0034744655713438988,
0.07648497819900513,
0.11611492186784744,
0.056254394352436066,
-0.08921286463737488,
-0.009278004057705402,
0.02823706530034542,
-0.07643375545740128,
0.03128359094262123,
0.029802590608596802,
-0.006091712508350611,
-0.09730839729309082,
-4.07029610016707e-8,
-0.07933498919010162,
0.06734611839056015,
0.03852621838450432,
-0.021389955654740334,
0.05086442083120346,
-0.08299850672483444,
-0.020564788952469826,
0.060504425317049026,
0.019257696345448494,
-0.011530607007443905,
0.10741451382637024,
-0.02320818230509758,
-0.005782443098723888,
-0.04294903203845024,
-0.0028993706218898296,
0.060980621725320816,
-0.05938728153705597,
0.0994967371225357,
-0.021195638924837112,
0.019441381096839905,
-0.00513528473675251,
0.03945505619049072,
0.022181302309036255,
-0.052125539630651474,
0.06175864115357399,
-0.042798757553100586,
-0.028752468526363373,
0.09643319994211197,
-0.013990778475999832,
0.018276799470186234,
-0.0782463550567627,
-0.02731890045106411,
0.08214671909809113,
-0.005641549360007048,
-0.039892714470624924,
0.1547362059354782,
-0.04520825669169426,
-0.04124495014548302,
-0.055210940539836884,
-0.03618748113512993,
-0.02237686887383461,
-0.054119765758514404,
0.03333716467022896,
-0.007028010208159685,
-0.006207561586052179,
0.0435573048889637,
0.08888939023017883,
-0.07419977337121964,
-0.0040750447660684586,
-0.009049194864928722,
0.035614654421806335,
-0.05143076926469803,
0.03067113272845745,
0.07394006103277206,
0.044515930116176605,
-0.02159864269196987,
0.04153062030673027,
-0.12192816287279129,
0.04116155579686165,
0.0006155489245429635,
-0.013069982640445232,
0.05087387189269066,
-0.012919845059514046,
0.012215526774525642
] | 0 | 0 | 0 | 0 |
|
CONGO: Compressive Online Gradient Optimization
|
https://openreview.net/forum?id=4BFzTrIjPN
|
[
"Jeremy Carleton",
"Prathik Vijaykumar",
"Divyanshu Saxena",
"Dheeraj Narasimha",
"Srinivas Shakkottai",
"Aditya Akella"
] |
Poster
|
We address the challenge of zeroth-order online convex optimization where the objective function's gradient exhibits sparsity, indicating that only a small number of dimensions possess non-zero gradients. Our aim is to leverage this sparsity to obtain useful estimates of the objective function's gradient even when the only information available is a limited number of function samples. Our motivation stems from the optimization of large-scale queueing networks that process time-sensitive jobs. Here, a job must be processed by potentially many queues in sequence to produce an output, and the service time at any queue is a function of the resources allocated to that queue. Since resources are costly, the end-to-end latency for jobs must be balanced with the overall cost of the resources used. While the number of queues is substantial, the latency function primarily reacts to resource changes in only a few, rendering the gradient sparse. We tackle this problem by introducing the Compressive Online Gradient Optimization framework which allows compressive sensing methods previously applied to stochastic optimization to achieve regret bounds with an optimal dependence on the time horizon without the full problem dimension appearing in the bound. For specific algorithms, we reduce the samples required per gradient estimate to scale with the gradient's sparsity factor rather than its full dimensionality. Numerical simulations and real-world microservices benchmarks demonstrate CONGO's superiority over gradient descent approaches that do not account for sparsity.
|
online convex optimization, compressive sensing, regret analysis
|
A compressive sensing based approach to online convex optimization and its application to the optimization of queueing networks
| 11,756 |
2407.06325
|
[
-0.10026039928197861,
-0.043988924473524094,
-0.058917298913002014,
0.07093743234872818,
0.04282531514763832,
-0.01594058983027935,
0.03372501954436302,
-0.05534839630126953,
0.07530120760202408,
0.012411767616868019,
-0.07708793878555298,
-0.016410665586590767,
-0.024269048124551773,
0.009748085401952267,
-0.013994403183460236,
-0.0246410071849823,
0.14373871684074402,
0.024643586948513985,
-0.027967635542154312,
-0.048032525926828384,
0.009076561778783798,
-0.003856280818581581,
-0.013878279365599155,
0.00874040462076664,
0.0017955454532057047,
-0.09067514538764954,
-0.022031402215361595,
-0.0261475071310997,
-0.02690436691045761,
-0.0034818481653928757,
-0.003651224309578538,
-0.03189004212617874,
0.01882508397102356,
-0.013601125217974186,
0.03166680783033371,
0.07288436591625214,
-0.0627380833029747,
-0.0022872004192322493,
0.02965896762907505,
0.006484958343207836,
0.06671054661273956,
0.030120588839054108,
-0.018474457785487175,
0.028780419379472733,
0.047074928879737854,
0.0035750889219343662,
0.06338395178318024,
-0.005384746473282576,
0.034010957926511765,
-0.02072456292808056,
0.036760952323675156,
-0.035975996404886246,
-0.07738842070102692,
-0.005707159638404846,
0.0336473174393177,
0.022428663447499275,
0.032796911895275116,
-0.037690721452236176,
0.015523627400398254,
0.04092366248369217,
-0.031330257654190063,
-0.16538885235786438,
0.04578874260187149,
0.0037985763046890497,
-0.016632407903671265,
-0.012467626482248306,
0.06284873932600021,
0.023625783622264862,
-0.0017058354569599032,
0.09920547902584076,
-0.028422238305211067,
0.02643858827650547,
-0.0847143605351448,
0.03328462690114975,
0.005750440526753664,
0.030548978596925735,
0.09576384723186493,
0.0023135037627071142,
-0.012206066399812698,
-0.03182942792773247,
0.004573915619403124,
0.04149165004491806,
0.022458454594016075,
0.010763492435216904,
0.016360238194465637,
-0.04973422735929489,
0.029522687196731567,
-0.023853925988078117,
0.13961246609687805,
-0.10032491385936737,
-0.022619297727942467,
0.03465166687965393,
-0.09025110304355621,
0.021410664543509483,
-0.009316273033618927,
-0.06755506992340088,
0.05283650755882263,
0.0023151016794145107,
-0.027645176276564598,
0.06154477596282959,
0.002692620037123561,
-0.03234826773405075,
0.06210416555404663,
0.02121513895690441,
0.029822880402207375,
-0.01708022505044937,
0.021846162155270576,
0.08220705389976501,
-0.01171404030174017,
-0.033598314970731735,
0.00624228036031127,
0.025793643668293953,
0.05787361413240433,
-0.006548739969730377,
0.04457899183034897,
-0.015008504502475262,
0.016461635008454323,
-0.07756165415048599,
-0.029042726382613182,
0.14925116300582886,
-0.011469635181128979,
-0.04567902907729149,
0.006728508044034243,
-0.011280927807092667,
0.02326618880033493,
0.05076112598180771,
0.011831540614366531,
9.689545750885383e-34,
0.04606260731816292,
0.04694059118628502,
0.04768338426947594,
-0.09944424778223038,
0.09100808948278427,
-0.016743598505854607,
0.08087058365345001,
0.03230549395084381,
-0.032031163573265076,
0.042304627597332,
-0.10251263529062271,
-0.035529229789972305,
-0.06784902513027191,
0.06658563762903214,
0.007289882283657789,
-0.010049272328615189,
0.04103401303291321,
0.04242616891860962,
0.10638587921857834,
-0.08232773840427399,
0.03372501581907272,
-0.07192804664373398,
-0.08561895042657852,
-0.006003573536872864,
0.06612726300954819,
-0.08088667690753937,
0.05566006526350975,
-0.11330582946538925,
0.019215470179915428,
0.005703616421669722,
-0.06893717497587204,
0.024082345888018608,
0.007811280898749828,
0.017561592161655426,
0.06309571117162704,
0.05135367810726166,
-0.04270513355731964,
0.055335305631160736,
-0.0342598631978035,
-0.01815788261592388,
0.0022596726194024086,
0.025193531066179276,
-0.048242777585983276,
-0.049823492765426636,
-0.10537398606538773,
-0.10947713255882263,
0.03274918720126152,
-0.0035725338384509087,
-0.025520985946059227,
-0.019563725218176842,
0.0030147514771670103,
0.021246498450636864,
-0.00929375272244215,
0.021909059956669807,
-0.06525880843400955,
-0.020402511581778526,
0.035290855914354324,
0.0009727645083330572,
0.05512240156531334,
0.04866722971200943,
0.0014747927198186517,
-0.07312460243701935,
-0.04519474133849144,
-0.007460408844053745,
0.05736610293388367,
0.030273813754320145,
0.003925475291907787,
0.01602995954453945,
0.06938926130533218,
-0.05678655952215195,
-0.014942117035388947,
-0.006682273466140032,
0.10733672231435776,
-0.07728448510169983,
0.05858438089489937,
0.05986788123846054,
0.06300088763237,
-0.06351607292890549,
-0.025408580899238586,
0.005740986671298742,
-0.013652208261191845,
-0.029492409899830818,
-0.03698602318763733,
-0.10583657026290894,
-0.043509501963853836,
-0.030341094359755516,
0.05643046274781227,
-0.021655162796378136,
0.016156427562236786,
-0.06183309853076935,
-0.17354297637939453,
-0.01250727940350771,
-0.006094456650316715,
-0.07627031207084656,
-0.01792355813086033,
-1.590578671170915e-33,
-0.02060781978070736,
-0.05531510338187218,
-0.019318632781505585,
0.04765424504876137,
0.04042450338602066,
-0.03967822715640068,
-0.044132109731435776,
-0.02282569371163845,
-0.058396629989147186,
0.018508056178689003,
-0.07465236634016037,
-0.04869447648525238,
0.03032028302550316,
0.046859487891197205,
-0.021786294877529144,
0.034353189170360565,
-0.09399702399969101,
0.03735796734690666,
0.00694908807054162,
0.04539251700043678,
-0.013261890970170498,
0.01742204837501049,
-0.09975883364677429,
0.013937619514763355,
0.033486392349004745,
0.06174623221158981,
-0.011276807636022568,
0.03260974958539009,
-0.03271661326289177,
-0.025148332118988037,
-0.1119595617055893,
-0.03034723550081253,
-0.042720403522253036,
-0.003285503713414073,
0.0039560855366289616,
0.053224850445985794,
0.00454922579228878,
0.12207502871751785,
-0.02900935895740986,
0.0881001427769661,
0.12774421274662018,
-0.020321639254689217,
0.02166900411248207,
-0.03433661535382271,
0.032249048352241516,
-0.05827724188566208,
-0.0320172943174839,
-0.025989582762122154,
-0.0242023766040802,
0.048105087131261826,
-0.002111451467499137,
0.06438080221414566,
-0.07387017458677292,
0.12155387550592422,
-0.03489755839109421,
-0.04924621805548668,
-0.03686608746647835,
-0.03716983646154404,
0.0772085040807724,
0.03687719628214836,
-0.005037514492869377,
-0.00857762061059475,
-0.07354234904050827,
0.016098786145448685,
0.04409879818558693,
-0.02733997069299221,
0.031116558238863945,
-0.03829687833786011,
0.019067946821451187,
0.05588172748684883,
0.020886162295937538,
0.04577118903398514,
0.040359701961278915,
0.06496788561344147,
-0.07598817348480225,
-0.002832846948876977,
0.029935114085674286,
0.004752598237246275,
-0.027477405965328217,
0.06124909222126007,
0.07615445554256439,
0.029495904222130775,
0.10341048240661621,
0.01495202537626028,
0.007995297200977802,
0.07675424963235855,
0.1203509196639061,
0.044796280562877655,
0.04118653014302254,
-0.04362662136554718,
-0.0414409302175045,
-0.059865426272153854,
-0.00071759830461815,
-0.0482160858809948,
-0.028563763946294785,
-4.784115859024496e-8,
-0.060394495725631714,
0.03448760136961937,
0.0202200748026371,
-0.005696485750377178,
0.0748186782002449,
-0.050631314516067505,
0.03768342733383179,
-0.004247154574841261,
-0.04987039789557457,
-0.01728794537484646,
0.0810120478272438,
-0.06526421010494232,
-0.06783206015825272,
-0.024893196299672127,
0.010567017830908298,
0.05441321060061455,
-0.05401547625660896,
-0.0669960230588913,
0.015798117965459824,
-0.023152679204940796,
-0.03610701858997345,
0.0683969110250473,
-0.006961063947528601,
-0.020623110234737396,
0.09740451723337173,
-0.037523217499256134,
0.05771712586283684,
0.04039531201124191,
0.12106229364871979,
-0.03613247349858284,
-0.08873309940099716,
0.004591275472193956,
0.05762692168354988,
-0.04959343001246452,
-0.01311892457306385,
0.02185293845832348,
0.013483730144798756,
-0.04426976293325424,
-0.035916768014431,
-0.024830346927046776,
-0.03831065073609352,
0.01941375993192196,
-0.03022252768278122,
-0.0550110898911953,
0.0834004282951355,
-0.017573123797774315,
-0.031647518277168274,
-0.0001418461324647069,
-0.026282567530870438,
-0.03083755448460579,
0.002879346953704953,
-0.020087022334337234,
0.020097816362977028,
0.11898990720510483,
0.051339492201805115,
-0.029062343761324883,
-0.022470738738775253,
-0.03800622746348381,
0.008752918802201748,
0.09644592553377151,
0.018784448504447937,
0.006563997827470303,
-0.07604668289422989,
-0.039900295436382294
] | 0 | 0 | 0 | 0 |
|
Scalable Extraction of Training Data from Aligned, Production Language Models
|
https://openreview.net/forum?id=vjel3nWP2a
|
[
"Milad Nasr",
"Javier Rando",
"Nicholas Carlini",
"Jonathan Hayase",
"Matthew Jagielski",
"A. Feder Cooper",
"Daphne Ippolito",
"Christopher A. Choquette-Choo",
"Florian Tramèr",
"Katherine Lee"
] |
Poster
|
Large language models are prone to *memorizing* some of their training data. Memorized (and possibly sensitive) samples can then be extracted at generation time by adversarial or benign users. There is hope that *model alignment*---a standard training process that tunes a model to harmlessly follow user instructions---would mitigate the risk of extraction. However, we develop two novel attacks that undo a language model's alignment and recover thousands of training examples from popular proprietary aligned models such as OpenAI's ChatGPT. Our work highlights the limitations of existing safeguards to prevent training data leakage in production language models.
|
privacy, language models, data extraction, security
|
We show that aligned, production language models still memorize---and can be made to repeat---their training datasets through two different attacks.
| 11,754 | null |
[
-0.14005544781684875,
-0.03903406113386154,
0.03056294657289982,
0.04437337443232536,
0.08237943798303604,
-0.001476203789934516,
-0.028630713000893593,
-0.04252730309963226,
0.05959419161081314,
-0.027302736416459084,
0.01755828596651554,
-0.044146183878183365,
0.018564943224191666,
-0.04194837436079979,
-0.0068837422877550125,
-0.06401479244232178,
-0.02674204483628273,
0.09383486956357956,
-0.043592967092990875,
-0.06377178430557251,
-0.012295163236558437,
0.08632425963878632,
0.003686833195388317,
0.08947809785604477,
0.01697426848113537,
-0.011900086887180805,
0.02340320497751236,
-0.006930631119757891,
0.008921142667531967,
-0.0794079527258873,
0.023982390761375427,
0.029320981353521347,
0.008274827152490616,
0.06281710416078568,
0.04001776501536369,
0.01800001971423626,
-0.04656505212187767,
-0.025240182876586914,
0.05827563628554344,
-0.05062401667237282,
0.001295688678510487,
-0.02988252229988575,
-0.01654667966067791,
-0.011043736711144447,
0.046750351786613464,
-0.05854485183954239,
-0.01713225245475769,
-0.024818183854222298,
-0.01548309251666069,
0.009515239857137203,
-0.13140518963336945,
-0.02814052440226078,
0.010630263946950436,
0.048191506415605545,
-0.02598142996430397,
-0.026901204138994217,
0.043747711926698685,
0.029199419543147087,
-0.012912074103951454,
0.04379415884613991,
-0.02765062265098095,
-0.048561565577983856,
-0.11083822697401047,
0.0014376082690432668,
-0.08321810513734818,
-0.03550000116229057,
0.007239195052534342,
0.054865673184394836,
-0.002742877695709467,
0.05241968855261803,
-0.027620337903499603,
0.06482672691345215,
-0.03984611853957176,
-0.020656047388911247,
-0.023460863158106804,
0.1137002483010292,
-0.017812300473451614,
-0.04065027832984924,
0.0014540390111505985,
-0.08203525096178055,
0.017342928797006607,
-0.031832609325647354,
0.07212886214256287,
-0.006781779695302248,
0.05097575485706329,
0.03123316913843155,
0.04424876719713211,
0.020938988775014877,
0.13672995567321777,
0.006678576115518808,
-0.08028999716043472,
-0.055895887315273285,
0.10824383050203323,
0.05939052999019623,
0.06569089740514755,
0.04407227784395218,
0.05884368717670441,
-0.00933827180415392,
-0.05214236304163933,
0.1099584773182869,
-0.04036992788314819,
0.11185762286186218,
0.001271178130991757,
-0.03780718147754669,
-0.00035795968142338097,
-0.031135940924286842,
0.012038182467222214,
-0.03128236159682274,
0.06082821264863014,
-0.050647396594285965,
0.07588113099336624,
0.04668580740690231,
-0.02069004438817501,
0.025671176612377167,
0.022273996844887733,
-0.016092928126454353,
-0.023836152628064156,
-0.0590425468981266,
-0.0886681005358696,
0.0702848806977272,
-0.012039363384246826,
0.040664639323949814,
0.01032306719571352,
0.04499480500817299,
-0.00911033246666193,
-0.03218249976634979,
-0.012305328622460365,
1.6427008942469665e-34,
0.08724615722894669,
0.03969143331050873,
-0.006045705638825893,
0.02521812543272972,
0.027213837951421738,
0.022107252851128578,
-0.012735421769320965,
0.021947549656033516,
0.014683433808386326,
0.000573530443944037,
-0.038251202553510666,
0.017183369025588036,
-0.046928130090236664,
0.00738782761618495,
0.02086384780704975,
0.03673100471496582,
-0.026050128042697906,
0.10133820027112961,
0.0021914674434810877,
0.05292551591992378,
0.08853015303611755,
0.008251774124801159,
-0.019449062645435333,
-0.01078797411173582,
-0.013777383603155613,
0.022996578365564346,
-0.01239214837551117,
-0.024502594023942947,
-0.027805233374238014,
0.04694439098238945,
-0.05120205879211426,
-0.030082501471042633,
0.03187604248523712,
0.05659416690468788,
0.06088884919881821,
-0.019882114604115486,
-0.03298170119524002,
-0.0850520133972168,
0.017884835600852966,
-0.01197000127285719,
0.006236843764781952,
0.025055019184947014,
0.03111352026462555,
-0.014274812303483486,
-0.03302372619509697,
-0.018556030467152596,
-0.03949873521924019,
-0.027869319543242455,
0.04421576112508774,
-0.06564056873321533,
-0.003244674764573574,
0.08602142333984375,
-0.011724402196705341,
-0.021736156195402145,
-0.05553853511810303,
-0.03645965829491615,
0.025347627699375153,
-0.0002051505580311641,
0.02447831630706787,
0.10197025537490845,
-0.09301190823316574,
0.02067120559513569,
0.004678443074226379,
0.01392387319356203,
-0.023557260632514954,
-0.0609368234872818,
-0.029171280562877655,
-0.016380274668335915,
0.034584831446409225,
0.03203300014138222,
-0.034489214420318604,
-0.037834830582141876,
-0.10210765153169632,
-0.005154850427061319,
-0.033037152141332626,
-0.028975967317819595,
0.08201185613870621,
-0.02241358533501625,
0.06428384780883789,
-0.04835179075598717,
0.08417525142431259,
-0.024507861584424973,
0.06777507066726685,
-0.05832865834236145,
-0.10839913040399551,
-0.05968313664197922,
0.017416974529623985,
-0.02772630751132965,
0.0019196646753698587,
0.1132117286324501,
-0.0717451423406601,
-0.03621438145637512,
-0.031508829444646835,
-0.01227684784680605,
0.009146489202976227,
-8.542032757387474e-34,
-0.06674417108297348,
-0.013662965036928654,
-0.05165201798081398,
0.009964232333004475,
-0.06781020015478134,
-0.09081947803497314,
0.007936408743262291,
0.09556356072425842,
-0.06067299097776413,
-0.05696015805006027,
-0.04169292747974396,
-0.03160933032631874,
0.030428435653448105,
-0.03352845832705498,
0.027139168232679367,
-0.05598435178399086,
0.07417993992567062,
0.06606710702180862,
-0.008456889539957047,
0.07484715431928635,
0.02978038787841797,
0.029043395072221756,
-0.06799940764904022,
0.039135612547397614,
0.07242006808519363,
0.016143999993801117,
-0.009137998335063457,
0.0616990402340889,
-0.0023076809011399746,
0.033967554569244385,
-0.07696647197008133,
0.017776090651750565,
-0.08867210894823074,
-0.041974928230047226,
-0.07880350947380066,
0.0025936574675142765,
0.06712193042039871,
0.028662167489528656,
0.046525534242391586,
0.06580863893032074,
0.061721328645944595,
0.023091016337275505,
-0.15163977444171906,
0.022784434258937836,
-0.03540918603539467,
-0.04916137456893921,
-0.08043429255485535,
-0.0006521166651509702,
-0.03228333964943886,
-0.05263363569974899,
0.014472577720880508,
0.019193250685930252,
0.015634508803486824,
-0.07118308544158936,
-0.011108532547950745,
-0.0844040960073471,
0.0025841740425676107,
-0.04239405319094658,
-0.03364856168627739,
0.024500731378793716,
-0.04721694067120552,
0.002492103958502412,
0.013600973412394524,
-0.008231664076447487,
0.040191709995269775,
-0.029753686860203743,
0.011890188790857792,
0.049814216792583466,
-0.012032176367938519,
-0.02963869646191597,
0.0820368081331253,
-0.07161640375852585,
-0.07697807252407074,
0.11287473887205124,
-0.010139653459191322,
-0.05049895867705345,
0.0009056850685738027,
-0.07171095162630081,
0.02419354021549225,
-0.04573056101799011,
0.10841874033212662,
-0.019250450655817986,
0.07809024304151535,
0.025998180732131004,
0.08784687519073486,
0.011212211102247238,
0.10704220831394196,
0.034420620650053024,
-0.04219350963830948,
0.0782971978187561,
-0.0467098243534565,
-0.013817288912832737,
-0.029877591878175735,
0.09617108851671219,
0.002671536523848772,
-3.2924827308988824e-8,
-0.08124871551990509,
-0.05596848204731941,
0.061467185616493225,
0.1491037756204605,
-0.01103771012276411,
-0.0768507719039917,
-0.08217791467905045,
0.05904765799641609,
0.05698511749505997,
-0.01541856862604618,
0.0006498037255369127,
0.027088019996881485,
-0.0661424919962883,
-0.03448542207479477,
-0.013496136292815208,
0.11935830861330032,
-0.08627428114414215,
0.05692664161324501,
-0.0761067196726799,
0.004728302825242281,
-0.021800095215439796,
0.08937732875347137,
0.030797962099313736,
0.0049059148877859116,
0.04292580857872963,
-0.06641319394111633,
-0.024049371480941772,
0.10554438084363937,
0.05193803086876869,
0.010303249582648277,
-0.019255224615335464,
-0.07336050271987915,
0.01903882436454296,
-0.015234662219882011,
0.022809742018580437,
0.13782867789268494,
-0.04064517095685005,
-0.026902753859758377,
0.012978416867554188,
-0.0013722566654905677,
-0.010396198369562626,
0.05733814463019371,
-0.02662823535501957,
0.016012346372008324,
0.015483924187719822,
-0.037824779748916626,
-0.009194089099764824,
-0.09985733032226562,
0.040724657475948334,
-0.08034111559391022,
0.006257383618503809,
-0.00724210636690259,
0.02230685018002987,
0.03164096921682358,
-0.036084335297346115,
0.012986645102500916,
-0.021334633231163025,
-0.04600106552243233,
0.032859403640031815,
0.061437997967004776,
0.005905827973037958,
0.0007287004846148193,
0.05578652396798134,
-0.0006249287398532033
] | 0 | 0 | 0 | 0 |
|
Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
|
https://openreview.net/forum?id=590yfqz1LE
|
[
"Michael Aerni",
"Javier Rando",
"Edoardo Debenedetti",
"Nicholas Carlini",
"Daphne Ippolito",
"Florian Tramèr"
] |
Poster
|
Large language models memorize parts of their training data. Memorizing short snippets and facts is required to answer questions about the world and to be fluent in any language. But models have also been shown to reproduce long verbatim sequences of memorized text when prompted by a motivated adversary. In this work, we investigate an intermediate regime of memorization that we call non-adversarial reproduction, where we quantify the overlap between model responses and pretraining data when responding to natural and benign prompts. For a variety of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up to 15% of the text output by popular conversational language models overlaps with snippets from the Internet. In worst cases, we find generations where 100% of the content can be found exactly online. For the same tasks, we find that human-written text has far less overlap with Internet data. We further study whether prompting strategies can close this reproduction gap between models and humans. While appropriate prompting can reduce non-adversarial reproduction on average, we find that mitigating worst-case reproduction of training data requires stronger defenses—even for benign interactions.
|
large language models, memorization, data extraction, originality, privacy
|
We measure the frequency at which LLMs reproduce training data when not prompted to do so adversarially, and find that it can happen frequently even on accident.
| 11,752 |
2411.10242
|
[
-0.08847206085920334,
-0.04726767912507057,
-0.0022600353695452213,
0.03391198441386223,
0.06813859939575195,
0.006225406192243099,
-0.017548589035868645,
-0.015408149920403957,
0.0470438115298748,
-0.04098357632756233,
0.010251183994114399,
-0.04103991016745567,
0.07727465033531189,
-0.05652213096618652,
-0.021738793700933456,
-0.040801551192998886,
0.07774234563112259,
0.020172778517007828,
-0.05020242556929588,
-0.030293675139546394,
0.03127681463956833,
0.036030713468790054,
0.12508966028690338,
0.024249570444226265,
-0.010433193296194077,
-0.0499071329832077,
-0.030675077810883522,
-0.050187453627586365,
-0.005386038217693567,
-0.0004358204605523497,
0.007032208610326052,
0.02385948970913887,
-0.007995628751814365,
0.05251820385456085,
-0.09306612610816956,
-0.00018541509052738547,
-0.053833283483982086,
0.052372731268405914,
0.045336585491895676,
-0.08414336293935776,
-0.06632913649082184,
-0.04177344962954521,
-0.000926707056351006,
0.014247691258788109,
0.03968960791826248,
-0.037857141345739365,
-0.06766039878129959,
0.014907418750226498,
-0.058801110833883286,
0.0022176336497068405,
-0.08171959966421127,
-0.0012016120599582791,
0.0038344450294971466,
0.06970959901809692,
-0.05696217715740204,
-0.03795098885893822,
0.07389611005783081,
0.07231616228818893,
-0.0026272437535226345,
0.0182405486702919,
-0.058927737176418304,
-0.05887690186500549,
-0.10037826001644135,
0.014195602387189865,
-0.05462193116545677,
-0.02345488779246807,
0.04586421698331833,
0.10475806891918182,
-0.056053776293992996,
0.10097707808017731,
0.053952157497406006,
0.055128879845142365,
-0.06737709790468216,
0.037280064076185226,
0.07530893385410309,
0.03535158932209015,
-0.020820485427975655,
-0.04103576019406319,
0.008712696842849255,
-0.05698871985077858,
0.01986788958311081,
-0.0036268248222768307,
0.08360590040683746,
0.005849319510161877,
0.0713479071855545,
0.022882122546434402,
0.01538345217704773,
0.04687010496854782,
0.06346872448921204,
-0.012556327506899834,
-0.09270010143518448,
-0.07314436882734299,
0.055684179067611694,
0.03601553291082382,
0.017485572025179863,
0.0778031051158905,
-0.004721864592283964,
-0.05519770830869675,
-0.010218068026006222,
0.09215767681598663,
-0.01748834364116192,
0.07442044466733932,
-0.002930184593424201,
-0.04635412618517876,
0.026232603937387466,
0.04868343472480774,
0.02658352069556713,
0.023503081873059273,
0.03881744667887688,
-0.006335381418466568,
0.027270467951893806,
0.060068096965551376,
-0.02243570052087307,
-0.003575193928554654,
0.07955606281757355,
-0.0686395913362503,
0.06207762658596039,
-0.01480176206678152,
-0.023857982829213142,
0.06586828827857971,
0.05966731533408165,
0.05252707749605179,
0.04613161087036133,
0.046229247003793716,
0.039349038153886795,
-0.10972872376441956,
-0.01453323196619749,
1.267547956283654e-34,
0.13470284640789032,
-0.0014598654815927148,
0.0023677111603319645,
0.05699748918414116,
0.05507275462150574,
0.03734958916902542,
0.006469725631177425,
0.05496387556195259,
0.013281688094139099,
-0.028571605682373047,
-0.04385353624820709,
-0.00784214399755001,
-0.027369258925318718,
0.06491544097661972,
0.03820393979549408,
0.06805185973644257,
-0.05877980589866638,
0.05480959638953209,
0.002350157592445612,
0.011628547683358192,
0.06434988975524902,
-0.04512227326631546,
0.04564550891518593,
-0.00832762848585844,
0.008270307444036007,
-0.050622086971998215,
-0.0008447220898233354,
-0.016180232167243958,
-0.03066256269812584,
0.03913890942931175,
-0.09326352924108505,
-0.07431638240814209,
0.014246512204408646,
0.04145222157239914,
0.008496838621795177,
-0.03145773336291313,
-0.038272205740213394,
-0.056983113288879395,
0.01810389757156372,
-0.0037434487603604794,
0.02220398373901844,
0.037693578749895096,
0.05968654528260231,
0.02300727181136608,
-0.05449581891298294,
-0.07156506925821304,
0.0007488341070711613,
-0.03127449005842209,
-0.0808754563331604,
-0.05158662423491478,
-0.008036375977098942,
0.08891922235488892,
0.01839545927941799,
-0.0042485385201871395,
-0.06830860674381256,
0.003920521587133408,
-0.00044561989489011467,
-0.028495077043771744,
-0.028170667588710785,
0.028784222900867462,
0.02391054667532444,
0.03205491229891777,
0.017483072355389595,
0.06392142176628113,
0.02596922591328621,
-0.05470526963472366,
-0.06703830510377884,
0.036304544657468796,
0.016296548768877983,
-0.004801948554813862,
-0.023930933326482773,
-0.052271462976932526,
-0.07093015313148499,
-0.056096501648426056,
0.011873845010995865,
-0.00710570951923728,
0.06389078497886658,
-0.0419062003493309,
0.05662938207387924,
-0.004949064459651709,
0.09294429421424866,
-0.08986920118331909,
0.04059244692325592,
-0.0704653263092041,
-0.0904095396399498,
-0.0546518936753273,
0.022139431908726692,
-0.06733397394418716,
0.011017504148185253,
0.05141063779592514,
-0.05860587954521179,
-0.04036392271518707,
-0.0649125874042511,
-0.006133188493549824,
0.015019961632788181,
-1.9162328506769212e-33,
-0.03699009120464325,
-0.03160974383354187,
-0.05090215057134628,
0.05595478415489197,
-0.1165180429816246,
-0.005064025521278381,
0.0343807153403759,
0.06795420497655869,
-0.07245311886072159,
-0.06898001581430435,
-0.09645531326532364,
0.014741534367203712,
0.08354724198579788,
-0.022646130993962288,
-0.006008603610098362,
-0.049237072467803955,
0.0709259957075119,
0.07858327776193619,
0.006924903020262718,
0.005810387432575226,
0.036866579204797745,
0.04969002306461334,
-0.08040233701467514,
-0.03380309417843819,
-0.01186752412468195,
-0.037953950464725494,
-0.07932650297880173,
0.06427183002233505,
-0.0057050082832574844,
0.020656729117035866,
-0.04749438911676407,
0.026945563033223152,
-0.01823480613529682,
-0.03813309967517853,
-0.07857305556535721,
0.06827900558710098,
0.05351906642317772,
0.02197987213730812,
-0.02552899718284607,
0.09747479856014252,
0.0812804102897644,
-0.04204746335744858,
-0.09201180189847946,
-0.017640596255660057,
-0.060644812881946564,
-0.07522032409906387,
-0.08522696793079376,
0.013793403282761574,
0.032262980937957764,
-0.02026013284921646,
-0.004900433123111725,
0.04028742015361786,
-0.02075165882706642,
0.0014941811095923185,
-0.08100536465644836,
-0.11221743375062943,
-0.04212028533220291,
-0.06915751099586487,
0.07368644326925278,
-0.00786546804010868,
-0.06716277450323105,
-0.0058625307865440845,
-0.02673541195690632,
-0.055726196616888046,
0.045154862105846405,
-0.10743983089923859,
0.012630226090550423,
0.05394025892019272,
0.052242256700992584,
-0.002603149274364114,
0.12080492824316025,
-0.039171017706394196,
-0.027420518919825554,
-0.00006672480958513916,
0.022557277232408524,
-0.00014535074296873063,
0.016419151797890663,
-0.060249414294958115,
-0.04828355088829994,
-0.05261211842298508,
0.0641767829656601,
-0.010216545313596725,
0.003364047035574913,
0.047712769359350204,
0.0436817966401577,
0.018149547278881073,
0.03689367324113846,
0.0590878501534462,
-0.03210176154971123,
0.07934350520372391,
0.044604167342185974,
0.011831498704850674,
-0.05218442529439926,
0.033328842371702194,
0.004231886472553015,
-4.0171624249296656e-8,
-0.040729109197854996,
0.02139429561793804,
0.018909256905317307,
0.0946212112903595,
0.08122656494379044,
-0.032540641725063324,
-0.07449477165937424,
0.04569632560014725,
-0.013752960599958897,
-0.08465034514665604,
0.05328087881207466,
0.04938039183616638,
0.025079788640141487,
-0.06577541679143906,
-0.02349967509508133,
0.09079806506633759,
-0.05550191178917885,
0.0080519188195467,
-0.053083598613739014,
-0.013322562910616398,
0.06091896817088127,
0.011403479613363743,
-0.06906575709581375,
-0.04270371422171593,
0.0026278344448655844,
-0.02725045010447502,
-0.013081013225018978,
0.07738041132688522,
0.006267325021326542,
0.02219397947192192,
-0.04203851521015167,
-0.03842981159687042,
-0.09396866708993912,
-0.031453344970941544,
0.006731516215950251,
0.11260287463665009,
0.01842942275106907,
-0.06252697110176086,
0.005292197223752737,
-0.07099057734012604,
0.013729745522141457,
0.041288990527391434,
0.017909713089466095,
0.021709682419896126,
0.08473292738199234,
0.006630145013332367,
-0.013011055998504162,
-0.14516319334506989,
0.06940891593694687,
-0.0573577880859375,
0.008993471041321754,
-0.025459755212068558,
0.038753870874643326,
0.08685551583766937,
0.054555248469114304,
0.0026052731554955244,
-0.007301290985196829,
0.014058977365493774,
0.016478920355439186,
0.1086927130818367,
0.06777897477149963,
0.013244803063571453,
-0.024620305746793747,
-0.05720675364136696
] | 0 | 0 | 0 | 0 |
|
InsightBench: Evaluating Business Analytics Agents Through Multi-Step Insight Generation
|
https://openreview.net/forum?id=ZGqd0cbBvm
|
[
"Gaurav Sahu",
"Abhay Puri",
"Juan A. Rodriguez",
"Amirhossein Abaskohi",
"Mohammad Chegini",
"Alexandre Drouin",
"Perouz Taslakian",
"Valentina Zantedeschi",
"Alexandre Lacoste",
"David Vazquez",
"Nicolas Chapados",
"Christopher Pal",
"Sai Rajeswar",
"Issam H. Laradji"
] |
Poster
|
Data analytics is essential for extracting valuable insights from data that can assist organizations in making effective decisions. We introduce InsightBench, a benchmark dataset with three key features. First, it consists of 100 datasets representing diverse business use cases such as finance and incident management, each accompanied by a carefully curated set of insights planted in the datasets. Second, unlike existing benchmarks focusing on answering single queries, InsightBench evaluates agents based on their ability to perform end-to-end data analytics, including formulating questions, interpreting answers, and generating a summary of insights and actionable steps. Third, we conducted comprehensive quality assurance to ensure that each dataset in the benchmark had clear goals and included relevant and meaningful questions and analysis. Furthermore, we implement a two-way evaluation mechanism using LLaMA-3 as an effective, open-source evaluator to assess agents’ ability to extract insights. We also propose AgentPoirot, our baseline data analysis agent capable of performing end-to-end data analytics. Our evaluation on InsightBench shows that AgentPoirot outperforms existing approaches (such as Pandas Agent) that focus on resolving single queries. We also compare the performance of open- and closed-source LLMs and various evaluation strategies. Overall, this benchmark serves as a testbed to motivate further development in comprehensive automated data analytics and can be accessed here: https://github.com/ServiceNow/insight-bench.
|
Automated Data Analysis, Data Analytics Benchmark, LLM agents, Code Generation, LLM Evaluation
|
We propose a comprehensive benchmark to evaluate LLM-based agents on their ability to perform multi-step data analysis and discover interesting insights in data.
| 11,745 |
2407.06423
|
[
-0.073615662753582,
-0.012216516770422459,
-0.10595318675041199,
0.08236607909202576,
0.024078531190752983,
-0.009973082691431046,
0.03523993864655495,
-0.012713035568594933,
0.01358984224498272,
-0.027137240394949913,
-0.04137597233057022,
-0.04993097111582756,
0.016883542761206627,
-0.002894808305427432,
0.003315793816000223,
0.09484698623418808,
0.06902846693992615,
-0.02123422920703888,
-0.10258753597736359,
-0.04477435722947121,
-0.06417183578014374,
0.041162729263305664,
0.05236407369375229,
-0.02737792395055294,
-0.0005234075360931456,
0.017732826992869377,
0.008686171844601631,
-0.02555886283516884,
-0.027233848348259926,
0.011542430147528648,
0.04797191545367241,
0.052705660462379456,
-0.05283614620566368,
0.09233787655830383,
-0.011851336807012558,
0.031870804727077484,
-0.022491484880447388,
0.030272018164396286,
0.019293950870633125,
-0.0035112546756863594,
-0.0012245362158864737,
-0.011531636118888855,
-0.041467756032943726,
-0.030816201120615005,
0.06704386323690414,
-0.08504607528448105,
-0.031811583787202835,
-0.06039075553417206,
0.03571721166372299,
0.03948448598384857,
-0.1568152755498886,
-0.024955855682492256,
0.0001394887949572876,
0.04905581846833229,
-0.00009846594912232831,
0.05870046839118004,
-0.059066567569971085,
0.004380474332720041,
-0.010225456207990646,
-0.04136725142598152,
-0.025960274040699005,
-0.08228836953639984,
0.000729004736058414,
0.008727552369236946,
-0.04046778008341789,
-0.012762660160660744,
-0.045655228197574615,
0.015816466882824898,
-0.0727354884147644,
0.004012302495539188,
-0.05966993421316147,
0.012279022485017776,
-0.06032450124621391,
0.02867901884019375,
-0.07136156409978867,
0.06932762265205383,
0.02625826559960842,
0.02300259843468666,
0.051410019397735596,
-0.05301109328866005,
-0.03384196013212204,
0.00046598297194577754,
0.002932807197794318,
0.1620802879333496,
0.03521837294101715,
-0.008762221783399582,
0.05841400474309921,
0.05557585135102272,
-0.02150399051606655,
0.06381989270448685,
-0.009140652604401112,
0.0019462177297100425,
0.09822379052639008,
-0.03660191595554352,
0.0766509622335434,
0.03470175340771675,
-0.05749233812093735,
-0.08358816802501678,
0.02264743112027645,
0.04971763864159584,
0.03308700770139694,
0.08643423765897751,
-0.027371445670723915,
-0.06133929267525673,
-0.09384799003601074,
-0.01768902875483036,
0.0863947793841362,
0.0469529926776886,
-0.04150480777025223,
-0.025875506922602654,
0.026361636817455292,
-0.001624767086468637,
0.06490056216716766,
-0.05470910295844078,
0.09134887903928757,
-0.03283330798149109,
-0.07777328044176102,
0.08442437648773193,
-0.02279769815504551,
0.04473505914211273,
0.04721526801586151,
0.08854171633720398,
0.07093790173530579,
-0.015093033201992512,
0.04908530041575432,
0.0931117907166481,
-0.018692418932914734,
3.426089957434885e-33,
0.05900535732507706,
0.0005374910542741418,
0.06645049154758453,
0.020377781242132187,
-0.014349651522934437,
0.02971452660858631,
-0.06219513341784477,
0.04043123498558998,
-0.07385613769292831,
0.07911090552806854,
-0.008487438783049583,
0.06884420663118362,
-0.030633267015218735,
-0.0029521379619836807,
0.058507565408945084,
0.019692882895469666,
-0.07566797733306885,
0.10253342986106873,
-0.046112366020679474,
-0.015664061531424522,
0.1129116341471672,
-0.07164572179317474,
0.03794661536812782,
0.014382409863173962,
0.1047346219420433,
-0.031047603115439415,
0.04766297712922096,
-0.012372741475701332,
-0.028632497414946556,
0.012041599489748478,
-0.05049624294042587,
0.03631192073225975,
-0.00030300088110379875,
0.08910975605249405,
-0.04580959677696228,
-0.016493206843733788,
-0.09049097448587418,
-0.06595464050769806,
0.018990013748407364,
0.044277891516685486,
-0.08254779130220413,
0.06204759702086449,
0.017188530415296555,
0.011218906380236149,
-0.10141858458518982,
-0.020687786862254143,
-0.03856859728693962,
0.005408313125371933,
0.08033724874258041,
-0.039674222469329834,
-0.016064785420894623,
-0.027245042845606804,
0.10964258015155792,
0.026658961549401283,
-0.050076622515916824,
-0.00034014094853773713,
0.020116642117500305,
-0.03040858544409275,
0.018688075244426727,
0.018852541223168373,
-0.0455315038561821,
-0.0002641004684846848,
-0.07721573859453201,
0.02585182897746563,
-0.06923192739486694,
0.021427951753139496,
0.006107316352427006,
0.1577272117137909,
0.010785689577460289,
0.025578508153557777,
0.02838217094540596,
0.034003566950559616,
0.057295508682727814,
-0.03253097087144852,
0.080597423017025,
-0.027089809998869896,
0.013489997945725918,
0.025229325518012047,
0.03226986527442932,
-0.027063315734267235,
0.012655205093324184,
-0.011404936201870441,
0.027438635006546974,
-0.03808761015534401,
-0.02807527966797352,
0.00369857344776392,
0.036662086844444275,
-0.018866850063204765,
-0.06519430875778198,
0.02492562122642994,
-0.08939351886510849,
0.03105214796960354,
-0.10163764655590057,
-0.000677016912959516,
0.00014783577353227884,
-2.4951015539905027e-33,
-0.04919581487774849,
-0.06044568866491318,
-0.07695676386356354,
0.03296380490064621,
-0.011462721973657608,
-0.01625972054898739,
-0.050370801240205765,
-0.006688279565423727,
0.045685797929763794,
-0.04455915838479996,
-0.007173341233283281,
-0.03838784992694855,
0.02257119119167328,
-0.05352666974067688,
-0.003557385876774788,
-0.027527643367648125,
-0.0231662318110466,
-0.1127442792057991,
-0.023810315877199173,
0.005837696138769388,
-0.02828168123960495,
0.04271189123392105,
-0.040537700057029724,
-0.040962934494018555,
-0.015244750306010246,
0.06797846406698227,
0.012983771041035652,
0.0013076587347313762,
0.0245404951274395,
-0.0032019754871726036,
-0.025304503738880157,
-0.028592461720108986,
-0.0262557752430439,
-0.02635558508336544,
-0.05060146749019623,
0.040650635957717896,
0.07942631095647812,
-0.11556544899940491,
-0.05681031569838524,
0.047349706292152405,
0.029553931206464767,
0.003229787340387702,
-0.024403760209679604,
0.06760337203741074,
-0.008418800309300423,
0.0716693103313446,
-0.042934615164995193,
0.019477473571896553,
-0.017051760107278824,
-0.04886883124709129,
-0.05630103126168251,
-0.021328572183847427,
-0.018303921446204185,
0.03613308072090149,
-0.015602600760757923,
-0.07072145491838455,
0.03798684850335121,
-0.07693611085414886,
-0.049903955310583115,
0.02663804404437542,
-0.06540871411561966,
0.10527374595403671,
0.06637527048587799,
0.03566165640950203,
0.026961488649249077,
0.0046072425320744514,
0.003555773524567485,
-0.035574719309806824,
-0.03827008977532387,
-0.07292809337377548,
0.026999305933713913,
-0.09031037241220474,
0.016890566796064377,
-0.022930791601538658,
0.046303775161504745,
0.050056491047143936,
-0.013047056272625923,
-0.07163265347480774,
-0.030476605519652367,
0.007036007009446621,
0.05697130039334297,
-0.0022454301360994577,
0.03327322006225586,
0.03267205134034157,
-0.01634911634027958,
0.01838277094066143,
0.006301118992269039,
0.07391954213380814,
0.000428451516199857,
0.0668586790561676,
-0.10064896941184998,
-0.0468430370092392,
-0.09219885617494583,
0.05675504356622696,
-0.018101714551448822,
-3.9774445070861475e-8,
-0.06905833631753922,
0.06613579392433167,
0.03733622282743454,
0.02160418964922428,
-0.023987529799342155,
-0.0815909206867218,
-0.012581546790897846,
0.15219427645206451,
0.005324190016835928,
-0.00789306778460741,
0.0914144515991211,
-0.04178689792752266,
-0.06619148701429367,
0.01902054063975811,
0.059195395559072495,
0.012742582708597183,
0.030327335000038147,
-0.015112928114831448,
-0.004487677477300167,
-0.024368571117520332,
0.05800936743617058,
0.03496258333325386,
0.023966709151864052,
-0.03804624080657959,
0.06560494750738144,
-0.06180882453918457,
-0.0901339128613472,
0.0780130922794342,
0.026231519877910614,
-0.025452783331274986,
-0.008815806359052658,
-0.04045752435922623,
0.05531512573361397,
-0.02971980720758438,
0.027473485097289085,
-0.01784726232290268,
-0.0030568609945476055,
-0.034343358129262924,
0.04187385365366936,
-0.01600739359855652,
-0.029254009947180748,
0.09574178606271744,
-0.013298533856868744,
0.04723120108246803,
-0.018713481724262238,
-0.06489124894142151,
-0.09407104551792145,
0.02974604442715645,
0.0899554118514061,
-0.07416953146457672,
-0.010814300738275051,
-0.11641176044940948,
0.06561717391014099,
0.08883566409349442,
0.06621167808771133,
0.0074720317497849464,
0.00808313861489296,
-0.0019476222805678844,
0.057077568024396896,
0.01691780425608158,
0.027772488072514534,
-0.05005712807178497,
-0.00782476644963026,
0.042128924280405045
] |
https://github.com/servicenow/insight-bench
| 39 | 0 | 0 | 0 |
Optimizing Posterior Samples for Bayesian Optimization via Rootfinding
|
https://openreview.net/forum?id=I6UbnkUveF
|
[
"Taiwo Adebiyi",
"Bach Do",
"Ruda Zhang"
] |
Poster
|
Bayesian optimization devolves the global optimization of a costly objective function to the global optimization of a sequence of acquisition functions. This inner-loop optimization can be catastrophically difficult if it involves posterior sample paths, especially in higher dimensions. We introduce an efficient global optimization strategy for posterior samples based on global rootfinding. It provides gradient-based optimizers with two sets of judiciously selected starting points, designed to combine exploration and exploitation. The number of starting points can be kept small without sacrificing optimization quality. Remarkably, even with just one point from each set, the global optimum is discovered most of the time. The algorithm scales practically linearly to high dimensions, breaking the curse of dimensionality. For Gaussian process Thompson sampling (GP-TS), we demonstrate remarkable improvement in both inner- and outer-loop optimization, surprisingly outperforming alternatives like EI and GP-UCB in most cases. Our approach also improves the performance of other posterior sample-based acquisition functions, such as variants of entropy search. Furthermore, we propose a sample-average formulation of GP-TS, which has a parameter to explicitly control exploitation and can be computed at the cost of one posterior sample.
|
Bayesian optimization, global optimization, acquisition function, Thompson sampling
| null | 11,744 |
2410.22322
|
[
-0.06726890057325363,
-0.06575638800859451,
0.01560780219733715,
-0.020085683092474937,
0.002045817207545042,
-0.07893159240484238,
-0.00672817463055253,
0.028551867231726646,
0.012648535892367363,
-0.04015078768134117,
-0.057649560272693634,
0.01124334242194891,
0.0185574758797884,
0.011912495829164982,
-0.05472761392593384,
0.017701270058751106,
0.026698270812630653,
0.10441334545612335,
-0.04900901019573212,
-0.0549740344285965,
-0.01161939837038517,
0.039696745574474335,
0.060652513056993484,
-0.07127693295478821,
-0.05466274172067642,
-0.002809795318171382,
0.061733879148960114,
-0.046502139419317245,
0.04805004596710205,
0.026249857619404793,
0.07179701328277588,
-0.005724950693547726,
0.019029170274734497,
-0.0503004789352417,
0.03336188942193985,
0.019565200433135033,
-0.1451476365327835,
-0.009995214641094208,
0.021929221227765083,
0.018349338322877884,
0.009146270342171192,
0.03824225068092346,
0.0289267860352993,
0.06597266346216202,
0.050955988466739655,
-0.030637871474027634,
-0.021214663982391357,
-0.027622820809483528,
0.07573577761650085,
0.008326933719217777,
-0.033417489379644394,
-0.02697553113102913,
-0.07834798097610474,
-0.05384375527501106,
0.04253522306680679,
0.02672668732702732,
0.07773551344871521,
-0.04753375053405762,
0.08585016429424286,
-0.0029010281432420015,
-0.007653286214917898,
-0.137455552816391,
0.024351423606276512,
-0.054299935698509216,
0.027464279904961586,
0.029715221375226974,
0.05317263677716255,
0.05459913983941078,
0.0350760817527771,
0.09927482903003693,
-0.04274589940905571,
0.11467713117599487,
-0.06542297452688217,
0.030376987531781197,
-0.009653910994529724,
0.03336350619792938,
-0.020994778722524643,
-0.010279910638928413,
-0.015806466341018677,
-0.020899882540106773,
0.018884319812059402,
0.017907550558447838,
0.002469906583428383,
-0.028710318729281425,
0.065175861120224,
0.010462092235684395,
0.00243734335526824,
0.015167375095188618,
0.17189230024814606,
0.05493834614753723,
-0.02924155443906784,
-0.0144726587459445,
-0.08397644758224487,
-0.049537740647792816,
-0.016978908330202103,
-0.006083148531615734,
-0.02625304087996483,
-0.06391493231058121,
0.029105670750141144,
0.021649476140737534,
0.04221973568201065,
0.07570932805538177,
0.07336925715208054,
-0.02866498939692974,
-0.03746764361858368,
-0.060828644782304764,
-0.00509791262447834,
0.06304603815078735,
0.05584065988659859,
-0.08080533891916275,
0.0010979981161653996,
0.0530497245490551,
0.08100540935993195,
-0.022421173751354218,
-0.031055063009262085,
0.050212644040584564,
0.026600627228617668,
-0.013527504168450832,
-0.008657576516270638,
0.12476102262735367,
0.0038549245800822973,
-0.026425013318657875,
-0.039976101368665695,
0.050929322838783264,
0.007130342535674572,
-0.01294904388487339,
-0.07098419964313507,
3.707353429492011e-33,
0.02170722931623459,
-0.01600317284464836,
0.06123749166727066,
0.0000875121186254546,
-0.019014419987797737,
0.060309674590826035,
0.03433078154921532,
-0.015878165140748024,
-0.05834183469414711,
0.04525553435087204,
-0.05112038180232048,
0.02100995182991028,
-0.1186886876821518,
-0.010168344713747501,
0.11880476027727127,
0.008759018033742905,
-0.015748465433716774,
0.08742235600948334,
-0.05562340095639229,
0.019131526350975037,
0.05932998284697533,
-0.05240137130022049,
-0.018796809017658234,
-0.047626785933971405,
0.03522515669465065,
-0.008657089434564114,
0.014441768638789654,
0.0008883446571417153,
-0.04633941501379013,
0.007662229705601931,
-0.10375893115997314,
-0.048575203865766525,
-0.07532872259616852,
0.08755040913820267,
0.07484342902898788,
0.00030122094904072583,
0.038105543702840805,
-0.019567856565117836,
-0.0494864247739315,
0.02553585171699524,
-0.01751779206097126,
0.014827506616711617,
0.10623655468225479,
0.013286151923239231,
-0.1202293187379837,
-0.09727337211370468,
-0.019048545509576797,
-0.03588137775659561,
0.03273798152804375,
-0.06946688145399094,
-0.013359365053474903,
-0.003096821252256632,
0.03889647498726845,
-0.050636738538742065,
0.003353628097102046,
0.004829524550586939,
0.011708793230354786,
-0.003909253980964422,
-0.021082447841763496,
0.08065526932477951,
0.0012106526410207152,
0.004734110087156296,
-0.038720570504665375,
0.03138469159603119,
-0.030057573691010475,
0.031950440257787704,
-0.03637443855404854,
0.01074919942766428,
0.06318001449108124,
0.02543373592197895,
0.001830337685532868,
0.012268428690731525,
0.050543420016765594,
-0.0768546536564827,
0.07562121003866196,
0.005408991128206253,
0.0648413598537445,
0.038768090307712555,
-0.0273466594517231,
-0.04807646572589874,
-0.041961751878261566,
0.02920016646385193,
-0.05807914584875107,
-0.0006400701240636408,
0.043138254433870316,
-0.04258852079510689,
-0.006521071307361126,
-0.06383175402879715,
-0.07094883918762207,
-0.050795942544937134,
-0.13493843376636505,
0.03089388832449913,
-0.05699075385928154,
0.00096934475004673,
-0.05426870286464691,
-2.9775158030313307e-33,
-0.01704617589712143,
-0.08305976539850235,
0.09600461274385452,
0.018682818859815598,
-0.01234173122793436,
-0.07556276023387909,
-0.01681782491505146,
-0.10183949023485184,
-0.0023262554313987494,
-0.06751495599746704,
-0.09333649277687073,
-0.006233451422303915,
0.04626551643013954,
-0.016032252460718155,
0.016514863818883896,
0.07986953854560852,
-0.01602982170879841,
0.03708998113870621,
0.03690847381949425,
-0.033371858298778534,
-0.004223874304443598,
0.05584730580449104,
-0.02275959774851799,
-0.004166141152381897,
-0.04056485369801521,
0.04702204838395119,
0.03470517694950104,
0.0636221170425415,
-0.010804231278598309,
0.012151921167969704,
-0.07323887944221497,
0.08176308870315552,
-0.0861811414361,
0.06934204697608948,
-0.12580443918704987,
0.08516830950975418,
0.07946889847517014,
0.026102455332875252,
0.014735613949596882,
0.10616283863782883,
0.014391756616532803,
0.025092363357543945,
0.01926850900053978,
-0.012793694622814655,
0.014801559038460255,
0.006481373682618141,
0.024428771808743477,
0.00416216254234314,
-0.026532119140028954,
-0.007893494330346584,
-0.055398937314748764,
0.07463276386260986,
-0.06001954525709152,
0.049133043736219406,
-0.09918118268251419,
-0.05785376951098442,
-0.05085338279604912,
-0.05280667915940285,
0.11207310855388641,
-0.01640378311276436,
0.014099903404712677,
0.02397632598876953,
-0.01463368721306324,
0.011193529702723026,
-0.037918929010629654,
0.041651543229818344,
0.07105139642953873,
0.059416305273771286,
-0.014270619489252567,
0.026432381942868233,
-0.054892804473638535,
-0.023830413818359375,
0.03360828012228012,
0.06478939950466156,
-0.018576892092823982,
0.03551548346877098,
-0.007104134187102318,
-0.07543012499809265,
0.013051481917500496,
-0.0009243912063539028,
0.018654419109225273,
0.05496416613459587,
-0.01723027601838112,
0.08896472305059433,
0.020051391795277596,
0.03696645423769951,
0.030501950532197952,
0.07527963072061539,
0.04084629938006401,
-0.09054970741271973,
-0.04295981675386429,
-0.01697430945932865,
0.0546928346157074,
0.0376269705593586,
-0.11163167655467987,
-4.2398092148232536e-8,
-0.12850163877010345,
0.020192185416817665,
0.03660143166780472,
0.049243997782468796,
0.08564914017915726,
-0.002239477587863803,
-0.044297151267528534,
0.08336696773767471,
0.01156611181795597,
-0.10396623611450195,
0.055776409804821014,
-0.024007564410567284,
-0.026100076735019684,
-0.031642720103263855,
0.00504256971180439,
0.023759422823786736,
-0.0111079765483737,
-0.02178928256034851,
-0.001387939671985805,
-0.029716845601797104,
-0.015982212498784065,
0.06421667337417603,
0.03269420191645622,
-0.01679239235818386,
0.03344840183854103,
-0.0978739783167839,
0.06175224483013153,
0.026959311217069626,
0.12285929173231125,
-0.01647290028631687,
-0.03898271918296814,
-0.0037540351040661335,
-0.008507243357598782,
0.08606058359146118,
-0.004316972568631172,
0.01951002888381481,
-0.05715404078364372,
0.004083889536559582,
-0.03701096773147583,
-0.04143081605434418,
-0.025653107091784477,
0.03415072709321976,
-0.04477714002132416,
-0.0002982075675390661,
-0.009069533087313175,
0.03131198137998581,
0.0175647996366024,
-0.0011378657072782516,
0.08695957809686661,
-0.0017880203668028116,
0.04243571683764458,
-0.024292852729558945,
-0.06020209938287735,
0.06120520830154419,
0.06305456161499023,
0.03907915577292442,
0.008218523114919662,
-0.05168140307068825,
0.009059550240635872,
0.030348408967256546,
-0.07876163721084595,
0.01862400770187378,
-0.09826766699552536,
-0.003919495735317469
] |
https://github.com/uquh/tsroots
| 7 | 0 | 0 | 0 |
Model-Agnostic Knowledge Guided Correction for Improved Neural Surrogate Rollout
|
https://openreview.net/forum?id=3ep9ZYMZS3
|
[
"Bharat Srikishan",
"Daniel O'Malley",
"Mohamed Mehana",
"Nicholas Lubbers",
"Nikhil Muralidhar"
] |
Poster
|
Modeling the evolution of physical systems is critical to many applications in science and engineering. As the evolution of these systems is governed by partial differential equations (PDEs), there are a number of computational simulations which resolve these systems with high accuracy. However, as these simulations incur high computational costs, they are infeasible to be employed for large-scale analysis. A popular alternative to simulators are neural network surrogates which are trained in a data-driven manner and are much more computationally efficient. However, these surrogate models suffer from high rollout error when used autoregressively, especially when confronted with training data paucity. Existing work proposes to improve surrogate rollout error by either including physical loss terms directly in the optimization of the model or incorporating computational simulators as `differentiable layers' in the neural network. Both of these approaches have their challenges, with physical loss functions suffering from slow convergence for stiff PDEs and simulator layers requiring gradients which are not always available, especially in legacy simulators. We propose the Hybrid PDE Predictor with Reinforcement Learning (HyPER) model: a model-agnostic, RL based, cost-aware model which combines a neural surrogate, RL decision model, and a physics simulator (with or without gradients) to reduce surrogate rollout error significantly. In addition to reducing in-distribution rollout error by **47%-78%**, HyPER learns an intelligent policy that is adaptable to changing physical conditions and resistant to noise corruption. Code available at https://github.com/scailab/HyPER.
|
deep learning, knowledge guided machine learning, scientific machine learning, computational fluid dynamics, reinforcement learning
|
We propose a novel model-agnostic, cost-aware method which combines a neural surrogate, decision model, and simulator to significantly reduce rollout error when performing time-series PDE prediction.
| 11,738 |
2503.10048
|
[
-0.07035049051046371,
-0.05309358984231949,
0.03733091428875923,
0.09354516863822937,
-0.02686418779194355,
-0.016480475664138794,
-0.10807108879089355,
0.0198359414935112,
-0.018324242904782295,
-0.013823235407471657,
0.0030920235440135,
-0.009936547838151455,
-0.016621017828583717,
0.034595832228660583,
-0.11821337789297104,
-0.0024786931462585926,
-0.013359223492443562,
0.1332883983850479,
-0.07905784994363785,
0.0076697519980371,
-0.00846668891608715,
0.02772011235356331,
-0.040072251111269,
-0.04158072918653488,
-0.006339587736874819,
0.007194382604211569,
-0.019088666886091232,
-0.023380210623145103,
-0.037653084844350815,
-0.0009377324022352695,
0.07539018988609314,
-0.04308278486132622,
-0.0690985769033432,
-0.028493860736489296,
-0.03132273256778717,
0.021460851654410362,
-0.07039723545312881,
0.0334576778113842,
0.03031376749277115,
-0.012383499182760715,
0.005151841789484024,
0.012755006551742554,
-0.009839755482971668,
0.07778998464345932,
0.18357376754283905,
-0.0399886853992939,
0.03704865649342537,
-0.07233373820781708,
0.03898812457919121,
0.032331470400094986,
-0.0016934546874836087,
-0.03531842306256294,
0.016872230917215347,
0.040203921496868134,
0.0465940497815609,
0.03914594277739525,
0.05537158250808716,
-0.018447741866111755,
-0.02331511117517948,
0.05547055974602699,
-0.013800674118101597,
-0.09624271094799042,
-0.020124606788158417,
-0.05384685844182968,
-0.05880296230316162,
0.002483162796124816,
0.06813609600067139,
0.05031650513410568,
0.012811348773539066,
0.06214359775185585,
-0.06138903275132179,
0.07018385082483292,
-0.04145623371005058,
-0.06486288458108902,
0.005148541182279587,
0.05719822272658348,
0.09198205918073654,
0.031373243778944016,
0.06148598715662956,
-0.0778723880648613,
0.01163286529481411,
0.0008563394076190889,
-0.03354880213737488,
-0.10289032012224197,
0.07240753620862961,
0.017175011336803436,
0.04444441571831703,
0.06947465986013412,
0.07967767119407654,
-0.050255946815013885,
0.03203319385647774,
-0.01279065478593111,
0.03080957569181919,
-0.03991435840725899,
0.010322610847651958,
-0.009633220732212067,
0.05300845950841904,
-0.09182531386613846,
-0.032779399305582047,
0.047267790883779526,
0.01717805303633213,
0.06282171607017517,
0.025116080418229103,
0.008191467262804508,
0.03225414827466011,
-0.03370494022965431,
0.04241403937339783,
0.06464450806379318,
0.001250006607733667,
-0.04581564664840698,
0.023224564269185066,
0.03377527371048927,
-0.022428393363952637,
0.044975195080041885,
0.06649697571992874,
-0.02597164735198021,
0.03295694291591644,
-0.09216376394033432,
0.0027095782570540905,
0.1005353331565857,
-0.11202731728553772,
-0.02419539913535118,
0.02484419010579586,
0.04238271340727806,
0.03614254295825958,
-0.045848771929740906,
-0.0442705862224102,
2.5842205561721406e-33,
-0.026658467948436737,
-0.0028546347748488188,
-0.0038022552616894245,
-0.006852563004940748,
0.031691718846559525,
-0.07331279665231705,
0.0658276230096817,
0.028502587229013443,
-0.008608208037912846,
-0.012890717014670372,
-0.05288074165582657,
0.024252204224467278,
-0.07468950003385544,
0.08453531563282013,
0.0798121839761734,
0.03631148859858513,
-0.035415709018707275,
0.023540254682302475,
-0.016187654808163643,
-0.08300388604402542,
0.023521699011325836,
-0.06857255101203918,
-0.011496836319565773,
-0.07091443985700607,
-0.0056001669727265835,
0.02201329544186592,
0.026828458532691002,
0.06122850626707077,
-0.02690180391073227,
0.0038819420151412487,
-0.07611711323261261,
0.052484411746263504,
-0.008804121054708958,
0.019053546711802483,
0.05731761083006859,
0.037401892244815826,
0.023225152865052223,
-0.0024733140598982573,
-0.059536512941122055,
0.0060995155945420265,
-0.03836344927549362,
0.017450572922825813,
0.04929373040795326,
0.05399837717413902,
-0.06763464212417603,
-0.09047240763902664,
0.06497908383607864,
0.005679765250533819,
-0.02406400814652443,
-0.09226534515619278,
-0.07601507008075714,
0.0366826057434082,
-0.026856528595089912,
-0.05207844078540802,
-0.04320646449923515,
-0.011436982080340385,
0.06252363324165344,
0.027332056313753128,
-0.011612685397267342,
0.04925530031323433,
-0.05235114321112633,
-0.019157929345965385,
-0.017375195398926735,
0.07252516597509384,
-0.016898056492209435,
0.04373988136649132,
-0.044792793691158295,
0.03689563274383545,
0.0052005210891366005,
-0.014816707000136375,
-0.00027646456146612763,
0.032944321632385254,
0.02663569338619709,
-0.05212872475385666,
0.06951641291379929,
-0.040026433765888214,
0.05272669717669487,
-0.018159044906497,
0.006188264116644859,
-0.051836881786584854,
0.01305778231471777,
0.07409325242042542,
-0.051207285374403,
-0.07477717101573944,
0.011497200466692448,
-0.06181919202208519,
-0.010771307162940502,
-0.007753767538815737,
-0.06652902811765671,
-0.0758921205997467,
-0.1326858401298523,
-0.003801519749686122,
0.007635386660695076,
-0.06316385418176651,
0.013509415090084076,
-4.275321810424616e-33,
-0.057539135217666626,
0.0453072115778923,
-0.020731931552290916,
0.06323288381099701,
-0.01335303857922554,
-0.01987694576382637,
-0.035435594618320465,
0.06899971514940262,
-0.04830008000135422,
-0.09020069986581802,
-0.04916222020983696,
-0.07176787406206131,
0.05508135259151459,
-0.010403059422969818,
0.020280299708247185,
0.022189732640981674,
-0.12960465252399445,
-0.10130520164966583,
0.07770441472530365,
0.013670196756720543,
0.005794374272227287,
0.03387093544006348,
-0.032093632966279984,
0.09023913741111755,
-0.07873411476612091,
-0.014915748499333858,
-0.04470402002334595,
0.12818102538585663,
0.033852823078632355,
-0.01429075375199318,
-0.009378737770020962,
-0.009604751132428646,
-0.02648571878671646,
-0.0021174743305891752,
-0.03156028687953949,
0.14573074877262115,
0.07934024184942245,
-0.025324786081910133,
0.012260610237717628,
0.027403615415096283,
0.04937295243144035,
-0.03465581685304642,
-0.05648452043533325,
-0.01871434971690178,
0.05310218036174774,
-0.024480734020471573,
-0.062914177775383,
-0.023172562941908836,
0.008955702185630798,
0.024520302191376686,
-0.0013413256965577602,
0.013950042426586151,
-0.05031811445951462,
0.011876476928591728,
-0.023472461849451065,
-0.019885187968611717,
-0.027247697114944458,
-0.006513675209134817,
0.02217177115380764,
-0.016382429748773575,
-0.030767470598220825,
-0.07029327750205994,
0.03234315663576126,
0.020197859033942223,
-0.026174409314990044,
0.0049192109145224094,
-0.025552986189723015,
0.07301359623670578,
0.006607551593333483,
0.016483258455991745,
-0.08178815990686417,
0.006551291793584824,
0.05882011726498604,
-0.0451248399913311,
0.042594656348228455,
-0.08637973666191101,
-0.03719291463494301,
-0.06031155586242676,
0.022097472101449966,
-0.007778934668749571,
0.044182296842336655,
-0.007906234823167324,
0.0788545086979866,
0.045919157564640045,
0.002464336110278964,
0.03624790534377098,
0.045668307691812515,
0.055116865783929825,
-0.00025897970772348344,
-0.03039986453950405,
0.008538058958947659,
-0.010346301831305027,
0.0024781012907624245,
0.0379764623939991,
-0.09371001273393631,
-4.5831445305566376e-8,
-0.00738678639754653,
0.11926651746034622,
0.16749387979507446,
0.023051496595144272,
0.05552845448255539,
-0.04408649727702141,
-0.0006001038127578795,
0.0010640990221872926,
-0.030179748311638832,
-0.022021010518074036,
0.05217110365629196,
-0.029102515429258347,
0.02323055826127529,
-0.03342942148447037,
-0.05164423957467079,
0.10443969815969467,
-0.02395336516201496,
0.008937464095652103,
0.021729569882154465,
0.00477944128215313,
-0.03480426222085953,
0.058489058166742325,
0.009588023647665977,
-0.04938831925392151,
0.10767348855733871,
-0.04229661077260971,
-0.07132326811552048,
0.06729079782962799,
0.04314836487174034,
-0.04418712109327316,
0.005719123408198357,
-0.03557963669300079,
0.0853051021695137,
0.031498029828071594,
0.022100253030657768,
0.10043605417013168,
0.1421886831521988,
-0.025551874190568924,
-0.011540420353412628,
0.05626017972826958,
-0.019312722608447075,
0.044365108013153076,
-0.06397928297519684,
0.017766989767551422,
0.004690417554229498,
-0.02672506496310234,
-0.06252793222665787,
-0.07474548369646072,
0.05914495512843132,
0.007360846269875765,
0.06785666942596436,
0.000982919940724969,
-0.06299877911806107,
0.00593214575201273,
0.07644907385110855,
-0.044913824647665024,
-0.027040516957640648,
-0.036769550293684006,
0.05127768963575363,
0.08871888369321823,
-0.03887178376317024,
-0.04325129836797714,
-0.0773419737815857,
0.0665505900979042
] |
https://github.com/scailab/hyper
| 1 | 0 | 0 | 0 |
Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians
|
https://openreview.net/forum?id=1durmugh3I
|
[
"Ishan Amin",
"Sanjeev Raja",
"Aditi S. Krishnapriyan"
] |
Poster
|
The foundation model (FM) paradigm is transforming Machine Learning Force Fields (MLFFs), leveraging general-purpose representations and scalable training to perform a variety of computational chemistry tasks. Although MLFF FMs have begun to close the accuracy gap relative to first-principles methods, there is still a strong need for faster inference speed. Additionally, while research is increasingly focused on general-purpose models which transfer across chemical space, practitioners typically only study a small subset of systems at a given time. At test time, MLFFs must also obey physical constraints unique to the downstream use case, such as energy conservation for molecular dynamics simulations. This underscores the need for fast, specialized MLFFs relevant to specific downstream applications, which preserve test-time physical soundness while maintaining train-time scalability. In this work, we introduce a method for transferring general-purpose representations from MLFF foundation models to smaller, faster MLFFs specialized to specific regions of chemical space. We formulate our approach as an architecture-agnostic knowledge distillation procedure, where the smaller "student" MLFF is trained to match the Hessians of the energy predictions of the "teacher" foundation model. We demonstrate our approach across multiple recent foundation models, large-scale datasets, chemical subsets, and downstream tasks. Our specialized MLFFs can be up to 20 times faster than the original foundation model, while retaining, and in some cases exceeding, its performance and that of undistilled models. We also show that distilling from a teacher model with a direct force parameterization into a student model trained with conservative forces (i.e., computed as derivatives of the potential energy) successfully leverages the representations from the large-scale teacher for improved accuracy, while maintaining energy conservation during test-time molecular dynamics simulations. More broadly, our work suggests a new paradigm for MLFF development, in which foundation models are released along with smaller, specialized simulation ``engines" for common chemical subsets. The implementation of our method is available at https://github.com/ASK-Berkeley/MLFF-distill.
|
machine learning force fields, graph neural networks, knowledge distillation
|
We distill large machine learning force field foundation models into small, specialized models using knowledge distillation from energy Hessians.
| 11,734 |
2501.09009
|
[
-0.06131855398416519,
-0.11567467451095581,
0.05728394165635109,
0.051384471356868744,
0.08280573785305023,
-0.05549335852265358,
-0.06456972658634186,
-0.02506416104733944,
-0.005946436431258917,
-0.03617340698838234,
-0.07123677432537079,
-0.07752875238656998,
-0.03683444857597351,
0.023344628512859344,
-0.04878498241305351,
-0.016572250053286552,
0.0365576334297657,
0.08391910791397095,
-0.12655839323997498,
-0.003591065062209964,
-0.010602131485939026,
0.01425965316593647,
0.005288306623697281,
0.044219501316547394,
0.04768545180559158,
0.021470356732606888,
0.01224078144878149,
0.0504433773458004,
-0.017603209242224693,
-0.03488725423812866,
0.0652959942817688,
-0.008763163350522518,
-0.06224416196346283,
0.001422464963980019,
0.0007939392817206681,
0.08944932371377945,
-0.023513175547122955,
0.05406292900443077,
-0.007614549249410629,
-0.010415326803922653,
-0.028259549289941788,
0.022544413805007935,
0.05118918418884277,
0.03775918111205101,
0.10822225362062454,
0.005479814484715462,
0.07731414586305618,
-0.08905623108148575,
0.05207596346735954,
0.0024044609162956476,
0.027886617928743362,
-0.04270738363265991,
-0.03291621059179306,
0.046154413372278214,
-0.0024512154050171375,
0.03191594406962395,
0.07838577777147293,
-0.043192408978939056,
-0.04043780639767647,
-0.01725492998957634,
-0.02049751952290535,
-0.05487578362226486,
-0.04910812899470329,
-0.020189858973026276,
-0.010938801802694798,
-0.011456333100795746,
0.04328665882349014,
0.06523755937814713,
0.03465455770492554,
-0.026141798123717308,
-0.09309758245944977,
0.09715358167886734,
-0.047695811837911606,
0.028460338711738586,
0.05973052233457565,
0.0652671679854393,
0.052756167948246,
0.07448463886976242,
0.07958140969276428,
-0.10943173617124557,
0.03644197806715965,
0.060557521879673004,
0.025724638253450394,
-0.039685867726802826,
0.03638974204659462,
0.013191338628530502,
0.03134315460920334,
0.07436951249837875,
0.11124786734580994,
-0.031057924032211304,
0.022257110103964806,
0.027708476409316063,
0.015029694885015488,
-0.05040089413523674,
-0.008659948594868183,
0.013490840792655945,
0.03963139280676842,
-0.056477949023246765,
-0.031043102964758873,
0.10216396301984787,
-0.03667883947491646,
0.10710598528385162,
0.00652761897072196,
-0.025447996333241463,
0.01340388786047697,
-0.03976157680153847,
0.026704855263233185,
0.08897913992404938,
0.028586555272340775,
-0.03176441043615341,
0.06075223535299301,
0.07920005172491074,
0.0032832373399287462,
0.006878947373479605,
0.04288209229707718,
-0.012925629504024982,
0.013253702782094479,
-0.12751077115535736,
0.004317339975386858,
0.03954870626330376,
-0.09672348946332932,
-0.004697727505117655,
-0.03173020854592323,
0.047629185020923615,
0.011722487397491932,
-0.019688475877046585,
-0.1341085433959961,
1.9186521649877074e-33,
-0.013780472800135612,
0.061342041939496994,
0.09492117911577225,
0.03424818813800812,
0.032838765531778336,
-0.08833695948123932,
0.0058710286393761635,
-0.006465490907430649,
-0.04317018389701843,
0.044239889830350876,
-0.03804662078619003,
0.041838712990283966,
-0.05267157405614853,
-0.038118790835142136,
-0.0028132922016084194,
0.01111247856169939,
-0.01593487337231636,
0.011231817305088043,
-0.02127464860677719,
-0.051545944064855576,
0.01808522455394268,
-0.026420148089528084,
-0.04105309024453163,
-0.03901213780045509,
0.014266003854572773,
0.05458522588014603,
0.013548108749091625,
0.03419797495007515,
-0.0419086217880249,
-0.015401902608573437,
-0.06749626994132996,
-0.037306223064661026,
-0.07876229286193848,
0.05281959846615791,
0.023239821195602417,
-0.052156224846839905,
0.030380981042981148,
-0.07880508899688721,
0.025944631546735764,
-0.05748683586716652,
-0.029217738658189774,
0.038967229425907135,
0.04355071112513542,
-0.020408038049936295,
-0.044453419744968414,
0.009567744098603725,
0.05352555215358734,
-0.006776438094675541,
-0.03364494442939758,
-0.0779915452003479,
0.01767677441239357,
0.008420548401772976,
0.03869487717747688,
-0.06900474429130554,
-0.0179839339107275,
0.02510368824005127,
0.03664989024400711,
-0.01924150437116623,
0.06371647864580154,
0.02794751524925232,
-0.027111437171697617,
0.04767114296555519,
-0.031769994646310806,
0.09827481210231781,
0.015298119746148586,
0.0644741803407669,
-0.10552005469799042,
0.06824558228254318,
0.051560722291469574,
0.033667344599962234,
-0.07306253165006638,
0.0041088322177529335,
0.02012801356613636,
-0.07114202529191971,
0.06674322485923767,
-0.02214176021516323,
0.040354594588279724,
-0.08053943514823914,
0.020057247951626778,
0.0428256019949913,
-0.0008866649004630744,
-0.0038588247261941433,
-0.062006741762161255,
-0.03106142394244671,
-0.06103285402059555,
-0.10000758618116379,
-0.005464455112814903,
-0.10946934670209885,
-0.015742221847176552,
-0.07314911484718323,
-0.10893087834119797,
-0.07271450012922287,
0.0595933236181736,
-0.02554943598806858,
-0.023535359650850296,
-2.940387446618624e-33,
-0.009362795390188694,
-0.001152563956566155,
0.0023150991182774305,
0.04756082966923714,
0.021664107218384743,
-0.012694149278104305,
-0.080863818526268,
-0.005455693695694208,
0.010786574333906174,
-0.004833750426769257,
0.03228232264518738,
-0.056690532714128494,
-0.010733547620475292,
-0.04675464332103729,
-0.017801329493522644,
0.009509521536529064,
-0.03542505204677582,
-0.013979307375848293,
0.0432715080678463,
0.029087983071804047,
-0.022357149049639702,
0.07508209347724915,
-0.035922445356845856,
0.009244764223694801,
-0.04577645659446716,
0.009332899004220963,
-0.004431925248354673,
0.05141996964812279,
-0.018571719527244568,
0.03409522771835327,
-0.0563722662627697,
0.00521648908033967,
-0.013472992926836014,
-0.01305573619902134,
-0.08954986184835434,
0.05922956392168999,
0.023219138383865356,
0.01139874942600727,
0.00693875877186656,
0.06632670760154724,
0.0009297847282141447,
-0.049463097006082535,
-0.06796625256538391,
-0.017663437873125076,
-0.010437186807394028,
0.05041832476854324,
-0.07194714993238449,
-0.03377176448702812,
0.0601012259721756,
-0.0034573657903820276,
0.03878890350461006,
0.013128362596035004,
-0.05733388662338257,
-0.01813722401857376,
-0.0171157568693161,
-0.001033420325256884,
-0.0022496823221445084,
0.0010847932426258922,
0.0560431182384491,
0.01541995070874691,
-0.05213529244065285,
-0.09091567248106003,
0.017801325768232346,
0.019150732085108757,
-0.08540726453065872,
-0.02419831044971943,
-0.06812991946935654,
0.06882926821708679,
0.04228130728006363,
0.01134099904447794,
-0.06797586381435394,
0.036593347787857056,
0.07855752855539322,
0.06954342126846313,
0.0030156278517097235,
-0.07473766803741455,
0.10303271561861038,
-0.11396295577287674,
-0.06719455868005753,
0.022442961111664772,
0.03734968602657318,
-0.016221433877944946,
0.021633554250001907,
0.014061286114156246,
0.057893525809049606,
0.026313437148928642,
0.07466712594032288,
0.035833798348903656,
0.03424034267663956,
-0.06668198853731155,
-0.014601055532693863,
0.020414795726537704,
0.04755958169698715,
0.0753178745508194,
-0.07543600350618362,
-4.2055287252651397e-8,
0.012531961314380169,
0.06769966334104538,
0.11072494089603424,
-0.04100468382239342,
-0.019622614607214928,
0.09181896597146988,
-0.05614152178168297,
-0.0035329116508364677,
-0.0066058156080543995,
0.010074633173644543,
0.08245347440242767,
0.014597767032682896,
0.022336363792419434,
-0.060660503804683685,
-0.0007240382255986333,
0.08423380553722382,
-0.004638068377971649,
-0.02005067467689514,
-0.04573560878634453,
0.0014545306330546737,
0.027575718238949776,
0.05284304544329643,
0.03601513430476189,
-0.02026802860200405,
0.14059045910835266,
-0.13438507914543152,
-0.0024412425700575113,
0.02212754637002945,
0.03769247233867645,
0.054246701300144196,
-0.05270577594637871,
0.030125446617603302,
0.0038749324157834053,
0.023204009979963303,
0.04031691700220108,
0.05492639169096947,
-0.0100681371986866,
-0.010103206150233746,
-0.023648902773857117,
0.057345062494277954,
-0.06883574277162552,
0.07959089428186417,
-0.1777539998292923,
0.01232083234935999,
-0.012624460272490978,
0.016578158363699913,
-0.0652453675866127,
0.003926437348127365,
0.058978646993637085,
0.10162903368473053,
0.07265499979257584,
0.024884972721338272,
-0.009693003259599209,
0.0007496464531868696,
0.03898428753018379,
0.047944966703653336,
-0.09659956395626068,
-0.03158031404018402,
0.020273935049772263,
-0.07860533893108368,
0.047956373542547226,
-0.04519420862197876,
0.003668606048449874,
-0.016404496505856514
] |
https://github.com/ASK-Berkeley/MLFF-distill
| 11 | 0 | 0 | 0 |
Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study
|
https://openreview.net/forum?id=r8J3DSD5kF
|
[
"Shawn Tan",
"Songlin Yang",
"Aaron Courville",
"Rameswar Panda",
"Yikang Shen"
] |
Poster
|
The self-attention mechanism traditionally relies on the softmax operator, necessitating positional embeddings like RoPE, or position biases to account for token order.
But current methods using still face length generalisation challenges.
We investigate an alternative attention mechanism based on the stick-breaking process in larger scale settings.
The method works as follows: For each token before the current, we determine a break point, which represents the proportion of the stick, the weight of the attention, to allocate to the current token.
We repeat this on the remaining stick, until all tokens are allocated a weight, resulting in a sequence of attention weights.
This process naturally incorporates recency bias, which has linguistic motivations for grammar parsing (Shen et al., 2017).
We study the implications of replacing the conventional softmax-based attention mechanism with stick-breaking attention.
We then discuss implementation of numerically stable stick-breaking attention and adapt Flash Attention to accommodate this mechanism.
When used as a drop-in replacement for current softmax+RoPE attention systems, we find that stick-breaking attention performs competitively with current methods on length generalisation and downstream tasks.
Stick-breaking also performs well at length generalisation, allowing a model trained with $2^{11}$ context window to perform well at $2^{14}$ with perplexity improvements.
|
transformer, attention, stick-breaking, softmax, length extrapolation
|
Using the stick-breaking process formulation as a replacement for softmax attention.
| 11,727 | null |
[
-0.013277452439069748,
-0.10775598138570786,
0.07401296496391296,
0.000313270662445575,
0.08153021335601807,
-0.02417069859802723,
0.011394912376999855,
0.04378461092710495,
0.045716434717178345,
-0.022773731499910355,
-0.03765709325671196,
-0.02446717582643032,
-0.009704113937914371,
0.034719374030828476,
-0.010091312229633331,
0.016476480290293694,
0.054224006831645966,
0.0678790882229805,
-0.09527090936899185,
-0.021255671977996826,
0.0721999928355217,
-0.025618160143494606,
0.07332508265972137,
0.0004947613924741745,
-0.009936466813087463,
0.011722825467586517,
-0.07195699959993362,
-0.054924122989177704,
0.10250701755285263,
-0.028585204854607582,
0.02202812395989895,
-0.04641474783420563,
0.038142651319503784,
0.08154462277889252,
-0.059574007987976074,
0.07060333341360092,
-0.10595572739839554,
0.0077267540618777275,
-0.002692886395379901,
-0.0636463314294815,
0.017271505668759346,
-0.0061953323893249035,
0.018252907320857048,
0.049951132386922836,
0.13269414007663727,
0.0034168942365795374,
0.03708444535732269,
-0.005512561649084091,
-0.040320172905921936,
-0.004032386001199484,
-0.018728824332356453,
0.005734718404710293,
-0.01811842806637287,
0.11690464615821838,
-0.030207742005586624,
-0.004047940019518137,
-0.01826128549873829,
0.08622445911169052,
0.04680244252085686,
0.008020663633942604,
0.004023767076432705,
-0.05800340324640274,
0.03851804509758949,
-0.03905177116394043,
0.014756479300558567,
-0.05459560826420784,
-0.0002580253058113158,
0.01895313896238804,
-0.11550168693065643,
0.1386898159980774,
-0.03150828555226326,
0.0485004223883152,
-0.034208234399557114,
0.08399394899606705,
0.06670485436916351,
0.07216749340295792,
-0.01656162738800049,
-0.04469645395874977,
0.013525261543691158,
-0.048962563276290894,
-0.000522984133567661,
-0.053971197456121445,
0.07160511612892151,
0.0074367825873196125,
0.0924200788140297,
-0.009904351085424423,
0.015615515410900116,
0.02154584415256977,
0.00984575878828764,
0.04633461683988571,
-0.04568751901388168,
-0.07490817457437515,
0.06350526958703995,
-0.008542997762560844,
0.03022555075585842,
-0.00467281648889184,
-0.07056111842393875,
-0.0094587542116642,
-0.07022163271903992,
0.06253748387098312,
0.008400327526032925,
0.07134688645601273,
0.03993924707174301,
-0.025016427040100098,
0.018564200028777122,
-0.055122070014476776,
-0.05469745025038719,
0.02827242575585842,
0.03393552079796791,
-0.11157763004302979,
0.003267587162554264,
0.0536084920167923,
-0.011390396393835545,
0.0037841058801859617,
0.047877177596092224,
-0.08095414191484451,
-0.00014786656538490206,
-0.019466890022158623,
0.03815591707825661,
0.049084100872278214,
0.055307455360889435,
-0.017528662458062172,
-0.07402773201465607,
0.02695588208734989,
0.025960268452763557,
0.018498405814170837,
0.02259467914700508,
1.9730874716643643e-34,
0.028390660881996155,
0.06971275061368942,
-0.034559592604637146,
-0.06520590931177139,
-0.0019274670630693436,
-0.015399031341075897,
-0.036755576729774475,
0.0632459968328476,
-0.057069770991802216,
0.04758095368742943,
0.042938824743032455,
-0.002286994829773903,
-0.017026033252477646,
0.07452359795570374,
0.07336550205945969,
-0.032840464264154434,
0.029607433825731277,
-0.0005985941388644278,
-0.06638418138027191,
-0.03512334078550339,
-0.0355927012860775,
0.028193805366754532,
-0.002615210600197315,
-0.05114280804991722,
0.006204948294907808,
-0.018644211813807487,
0.07262770086526871,
-0.05960780754685402,
-0.09290335327386856,
0.006888132076710463,
-0.14071272313594818,
-0.057822488248348236,
-0.02808346226811409,
0.002429817570373416,
0.02267448976635933,
-0.12453143298625946,
-0.0035228654742240906,
-0.04271905869245529,
0.008997207507491112,
-0.007217973005026579,
-0.0056219096295535564,
0.04019800201058388,
-0.043600790202617645,
0.0966063141822815,
-0.03920948877930641,
-0.0743229016661644,
-0.034904420375823975,
0.11291347444057465,
-0.015390009619295597,
-0.06843362003564835,
0.03855594992637634,
0.04151024669408798,
0.008034991100430489,
-0.08175577968358994,
0.008282179944217205,
-0.011857550591230392,
0.02138477936387062,
0.04141176864504814,
-0.010228713974356651,
-0.0008615738479420543,
0.06132688745856285,
0.02947762794792652,
-0.008216923102736473,
0.11261294782161713,
-0.028819970786571503,
0.05232546105980873,
-0.038066331297159195,
0.0760624036192894,
0.05349203944206238,
0.005245296750217676,
-0.008545494638383389,
0.07005245238542557,
-0.07645794004201889,
-0.050441790372133255,
0.029375523328781128,
0.0256726685911417,
0.12372652441263199,
-0.10526633262634277,
-0.0010452482383698225,
0.030883116647601128,
-0.0005506138550117612,
0.0013911389978602529,
0.00703230919316411,
-0.005236390978097916,
-0.04119591414928436,
-0.07430346310138702,
0.05015301704406738,
-0.1198774129152298,
-0.002583215245977044,
-0.037178054451942444,
0.0035041477531194687,
-0.08222617208957672,
-0.01900489814579487,
0.0013862644555047154,
-0.044673141092061996,
-2.441242331381245e-34,
-0.04752625897526741,
0.006137679796665907,
-0.04934515431523323,
0.01225904282182455,
-0.038943495601415634,
-0.028170371428132057,
-0.04939443618059158,
0.0018378084059804678,
-0.06660158932209015,
-0.04520803689956665,
-0.09037434309720993,
-0.02551771141588688,
-0.005800081416964531,
0.04570382833480835,
0.020605186000466347,
-0.015513289719820023,
-0.027164988219738007,
0.06846126914024353,
-0.021818585693836212,
-0.006987961009144783,
0.014495306648314,
0.10235284268856049,
-0.0297589972615242,
0.11626027524471283,
-0.012950031086802483,
0.07572135329246521,
0.017450464889407158,
-0.0022620519157499075,
-0.05638821795582771,
-0.0800899863243103,
-0.1056494489312172,
-0.019018149003386497,
0.040054403245449066,
0.02394293248653412,
-0.07484754174947739,
0.04164847359061241,
-0.020225077867507935,
0.02198003977537155,
-0.05863838270306587,
0.031121687963604927,
0.07034655660390854,
-0.06145963445305824,
-0.029902368783950806,
0.010437751188874245,
-0.03426915779709816,
-0.049383196979761124,
-0.08629585802555084,
0.01871355250477791,
-0.060636475682258606,
0.08468423038721085,
-0.040773868560791016,
0.03873361274600029,
-0.06040915474295616,
0.033700548112392426,
-0.04267395660281181,
-0.027052447199821472,
-0.08130725473165512,
-0.023496733978390694,
0.0052554006688296795,
-0.035529643297195435,
-0.04491881653666496,
-0.02209063246846199,
0.05025767162442207,
-0.015967248007655144,
0.0704454779624939,
-0.0405285619199276,
0.056027255952358246,
-0.022971097379922867,
-0.003380269045010209,
-0.060570135712623596,
0.11096886545419693,
0.005070224404335022,
-0.006543868687003851,
0.04424352943897247,
-0.04502951353788376,
0.05871133878827095,
0.02486003190279007,
-0.060727253556251526,
-0.06890328973531723,
-0.03739915043115616,
-0.05099349468946457,
-0.003060090821236372,
0.037485428154468536,
0.020171545445919037,
-0.020915068686008453,
0.06460709869861603,
0.07266299426555634,
0.08159859478473663,
0.02933124266564846,
0.00786751601845026,
-0.01929648220539093,
0.022960497066378593,
-0.02765871211886406,
0.054982464760541916,
-0.0971054956316948,
-4.1772196368583536e-8,
-0.06828370690345764,
0.03104141168296337,
-0.05557245388627052,
0.08620545268058777,
0.07368406653404236,
0.018984558060765266,
-0.03814081847667694,
0.049578167498111725,
-0.03686763346195221,
-0.06737914681434631,
0.049433860927820206,
0.037677109241485596,
-0.00029025101684965193,
-0.033936575055122375,
-0.028507256880402565,
0.06450853496789932,
0.0408669114112854,
-0.013722592033445835,
-0.05485913157463074,
-0.023795094341039658,
-0.014833268709480762,
0.0670459121465683,
0.02121412381529808,
-0.03774594888091087,
-0.04094819724559784,
-0.07446583360433578,
-0.07712570577859879,
0.14197096228599548,
0.05724473297595978,
-0.06300053745508194,
0.0447399877011776,
0.06735796481370926,
-0.003961141686886549,
-0.014291149564087391,
-0.022367317229509354,
0.052815116941928864,
0.025116391479969025,
0.04338658228516579,
0.039231203496456146,
0.06620780378580093,
0.07692428678274155,
0.034250643104314804,
0.012808873318135738,
0.0731535404920578,
-0.011632705107331276,
0.03206688165664673,
-0.055113133043050766,
-0.05525215342640877,
0.06765468418598175,
-0.0708683505654335,
0.057167183607816696,
-0.06356486678123474,
-0.020156409591436386,
0.04620879516005516,
0.012882398441433907,
-0.005786466877907515,
-0.008237314410507679,
0.03713858872652054,
-0.04138739034533501,
0.034915268421173096,
0.02950441651046276,
0.025282075628638268,
-0.02440115250647068,
0.04605115205049515
] | 0 | 0 | 0 | 0 |
|
FlowDec: A flow-based full-band general audio codec with high perceptual quality
|
https://openreview.net/forum?id=uxDFlPGRLX
|
[
"Simon Welker",
"Matthew Le",
"Ricky T. Q. Chen",
"Wei-Ning Hsu",
"Timo Gerkmann",
"Alexander Richard",
"YI-CHIAO WU"
] |
Poster
|
We propose FlowDec, a neural full-band audio codec for general audio sampled at 48 kHz that combines non-adversarial codec training with a stochastic postfilter based on a novel conditional flow matching method. Compared to the prior work ScoreDec which is based on score matching, we generalize from speech to general audio and move from 24 kbit/s to as low as 4 kbit/s, while improving output quality and reducing the required postfilter DNN evaluations from 60 to 6 without any fine-tuning or distillation techniques. We provide theoretical insights and geometric intuitions for our approach in comparison to ScoreDec as well as another recent work that uses flow matching, and conduct ablation studies on our proposed components. We show that FlowDec is a competitive alternative to the recent GAN-dominated stream of neural codecs, achieving FAD scores better than those of the established GAN-based codec DAC and listening test scores that are on par, and producing qualitatively more natural reconstructions for speech and harmonic structures in music.
|
audio, audio codec, generative models, flow matching, postfilter, signal enhancement
|
FlowDec is a flow-based postfilter codec for general audio without adversarial training, and a competitive alternative to current GAN-based SOTA codecs.
| 11,725 |
2503.01485
|
[
-0.11161477863788605,
-0.10240591317415237,
0.017791353166103363,
-0.02330111712217331,
-0.014189265668392181,
0.03362928703427315,
-0.005807351786643267,
-0.130416139960289,
-0.010607645846903324,
-0.07339399307966232,
-0.12078215926885605,
-0.07385032624006271,
-0.11021141707897186,
-0.06867928802967072,
-0.09265683591365814,
-0.01375956553965807,
0.05990177020430565,
0.11380720138549805,
-0.012679439038038254,
-0.02046893909573555,
0.01711104065179825,
0.03570454567670822,
-0.01188330352306366,
0.03594300523400307,
0.11251513659954071,
0.06792422384023666,
-0.03960159420967102,
-0.0035874363966286182,
0.030119528993964195,
-0.04470488801598549,
0.08182347565889359,
0.06576686352491379,
0.068374402821064,
0.03914536535739899,
-0.1270347684621811,
-0.006164723541587591,
0.0004914774326607585,
-0.015723858028650284,
-0.011390775442123413,
-0.05507393181324005,
0.04596848040819168,
0.1036878153681755,
0.014389421790838242,
-0.00852180179208517,
-0.03436441719532013,
-0.03866007179021835,
-0.04746045917272568,
-0.057142868638038635,
-0.09859026968479156,
0.03766634687781334,
0.04925625026226044,
-0.0017555684316903353,
-0.02527233399450779,
0.11531238257884979,
-0.05569610744714737,
0.024380512535572052,
0.06593514233827591,
-0.019094552844762802,
0.01783662475645542,
0.0568830668926239,
-0.03933393210172653,
-0.07532054930925369,
0.028500782325863838,
-0.04385773465037346,
0.018709560856223106,
-0.07949694246053696,
-0.048160310834646225,
0.07768892496824265,
0.005775254685431719,
0.007523057982325554,
-0.020568737760186195,
0.048302631825208664,
0.0321592353284359,
0.08374518901109695,
0.07907909154891968,
0.058197807520627975,
0.03480766341090202,
-0.01862778700888157,
0.020882511511445045,
-0.025068502873182297,
0.010843230411410332,
-0.013929682783782482,
0.1067887395620346,
-0.004701082594692707,
0.0625557228922844,
0.04896126314997673,
0.015611164271831512,
-0.022979505360126495,
-0.03983667492866516,
-0.013765159994363785,
-0.08742662519216537,
0.012885275296866894,
-0.023026304319500923,
-0.015375044196844101,
0.0919160321354866,
0.07558345049619675,
0.00673663429915905,
-0.05326800048351288,
0.03882201761007309,
0.0574631467461586,
0.024318192154169083,
0.05519478768110275,
-0.001210924587212503,
-0.042711373418569565,
0.009640711359679699,
-0.04481912776827812,
0.092623271048069,
0.08159895241260529,
0.04707452282309532,
-0.07129423320293427,
0.03276748210191727,
0.020581642165780067,
0.046250659972429276,
-0.04782198369503021,
0.04043398052453995,
0.014850967563688755,
-0.05282062664628029,
0.0071527003310620785,
0.047542374581098557,
0.03196253627538681,
-0.1310008466243744,
0.06417594105005264,
-0.11331149935722351,
-0.01259161438792944,
0.021857380867004395,
-0.007784638088196516,
-0.012179836630821228,
2.1501499816768177e-33,
0.0391252264380455,
0.00041345960926264524,
0.04134165123105049,
0.007051215507090092,
0.034306515008211136,
-0.07077665627002716,
-0.07388674467802048,
0.03106461465358734,
0.04983026161789894,
0.0295867919921875,
-0.03989415243268013,
-0.023143723607063293,
-0.08170534670352936,
0.013422160409390926,
-0.054827816784381866,
0.018265049904584885,
-0.08967725187540054,
-0.024700529873371124,
-0.004053698852658272,
0.013852433301508427,
0.05786984786391258,
-0.038482990115880966,
-0.014950016513466835,
-0.00010805289639392868,
0.07927852869033813,
0.007114412263035774,
0.01634978875517845,
0.03412710130214691,
-0.018697671592235565,
0.029328420758247375,
-0.12029270082712173,
-0.04096847027540207,
0.06434378772974014,
0.001635518274269998,
0.013439158909022808,
-0.020207932218909264,
-0.03186720982193947,
0.038159701973199844,
-0.020230628550052643,
-0.03349752351641655,
-0.0415109284222126,
0.017125751823186874,
-0.07321644574403763,
-0.013977812603116035,
-0.07174721360206604,
-0.1037498190999031,
-0.00962719414383173,
0.07822854816913605,
0.0008315467857755721,
0.017510099336504936,
0.03603145480155945,
0.06037004664540291,
0.0008305850788019598,
0.015666916966438293,
-0.03376777097582817,
-0.06621739268302917,
0.04904001206159592,
0.07826300710439682,
0.059484273195266724,
0.11617624014616013,
0.05849187821149826,
0.036134812980890274,
-0.0386878103017807,
0.009578979574143887,
0.04544986039400101,
0.06037565693259239,
-0.006614812649786472,
0.05015147104859352,
0.05215862765908241,
-0.09933319687843323,
-0.05486109107732773,
-0.02458806335926056,
-0.025419937446713448,
0.041908420622348785,
-0.008688244968652725,
-0.04013214260339737,
0.06480863690376282,
-0.0413748100399971,
0.03915150463581085,
0.0018292834283784032,
-0.04197586327791214,
0.056421756744384766,
-0.003680598456412554,
-0.04243446886539459,
-0.05292366445064545,
0.044689249247312546,
0.010754600167274475,
-0.139312744140625,
-0.04228239133954048,
-0.03092307224869728,
-0.08731669932603836,
-0.024888452142477036,
0.05725110322237015,
-0.04324398934841156,
0.056901365518569946,
-2.592460955242397e-33,
0.01125127263367176,
0.11493638902902603,
-0.03096192702651024,
0.047530628740787506,
-0.022779366001486778,
0.037192974239587784,
0.10645661503076553,
0.050307974219322205,
0.057030364871025085,
0.02994733490049839,
0.04587353765964508,
-0.04078439250588417,
0.03025110810995102,
-0.07856211811304092,
0.026604436337947845,
-0.06508901715278625,
-0.06432989239692688,
0.02981874719262123,
0.026202503591775894,
0.061133839190006256,
0.05053243041038513,
0.1025310531258583,
0.012686891481280327,
-0.002819971414282918,
-0.10731729120016098,
-0.012968155555427074,
-0.057073213160037994,
0.03328300267457962,
0.02846599742770195,
-0.07408611476421356,
0.06638246029615402,
-0.022209420800209045,
-0.07065851241350174,
-0.05437346175312996,
-0.03417971730232239,
-0.02362905815243721,
0.08864516019821167,
0.04855075478553772,
-0.0242757610976696,
0.04842856898903847,
0.0294711384922266,
-0.004760274197906256,
-0.009472286328673363,
0.015938423573970795,
0.014661077409982681,
-0.01976660266518593,
-0.06185166537761688,
0.05151902511715889,
0.01372342649847269,
-0.0027689768467098475,
0.04468180984258652,
0.03555619716644287,
0.026161909103393555,
0.024688703939318657,
-0.03990156203508377,
-0.08256355673074722,
-0.022088322788476944,
-0.01103602908551693,
0.008429214358329773,
0.07651834189891815,
-0.026121146976947784,
-0.025573763996362686,
-0.005756933707743883,
-0.06672821193933487,
0.01221491675823927,
0.0120151462033391,
-0.02073846012353897,
0.07225794345140457,
0.05756083503365517,
0.022134754806756973,
-0.03600471839308739,
-0.04605862870812416,
-0.05604027211666107,
0.04523811489343643,
-0.03323345631361008,
-0.021843545138835907,
-0.003287006402388215,
-0.020825190469622612,
-0.03366551548242569,
0.0046118819154798985,
-0.011608589440584183,
0.005602637771517038,
-0.0010726149193942547,
0.017950808629393578,
0.1706816405057907,
0.07794933021068573,
-0.0032825428061187267,
-0.10386960953474045,
-0.013570057228207588,
0.09400786459445953,
-0.04091641306877136,
0.06146236136555672,
0.015689950436353683,
-0.05747710168361664,
-0.044002924114465714,
-4.025331890034067e-8,
-0.04175330325961113,
0.06635816395282745,
-0.02757042832672596,
0.029982790350914,
-0.012656224891543388,
-0.052555229514837265,
-0.024095823988318443,
0.01828717812895775,
0.0046698846854269505,
-0.03296084329485893,
0.0756651982665062,
0.0068660760298371315,
0.011394888162612915,
-0.045095622539520264,
-0.04765952751040459,
-0.005880865268409252,
-0.04590172320604324,
0.09112508594989777,
-0.0012454907409846783,
-0.02452591247856617,
0.033521026372909546,
0.07821354269981384,
0.013516765087842941,
-0.019658558070659637,
0.04519817605614662,
-0.06754358112812042,
0.08184422552585602,
0.03550691157579422,
0.019572310149669647,
0.018518710508942604,
-0.002242231508716941,
0.016403231769800186,
-0.004928283393383026,
-0.08477538824081421,
0.03844232112169266,
0.019223617389798164,
0.004088814370334148,
-0.032903123646974564,
-0.07017254829406738,
0.017189670354127884,
-0.0167863592505455,
0.03663655370473862,
-0.025155195966362953,
-0.03700009360909462,
-0.06775005161762238,
-0.024559523910284042,
0.022006653249263763,
-0.028934050351381302,
-0.02140939049422741,
0.03342669829726219,
0.009236812591552734,
0.010135918855667114,
-0.02379755862057209,
-0.00702526792883873,
0.08116798847913742,
-0.013854517601430416,
-0.03638862818479538,
-0.02596212364733219,
-0.06561397016048431,
0.049884531646966934,
-0.03675011172890663,
0.03873521089553833,
0.004042649641633034,
0.004051698371767998
] |
https://github.com/facebookresearch/flowdec
| 141 | 0 | 0 | 0 |
End-to-end Learning of Gaussian Mixture Priors for Diffusion Sampler
|
https://openreview.net/forum?id=iXbUquaWbl
|
[
"Denis Blessing",
"Xiaogang Jia",
"Gerhard Neumann"
] |
Poster
|
Diffusion models optimized via variational inference (VI) have emerged as a promising tool for generating samples from unnormalized target densities. These models create samples by simulating a stochastic differential equation, starting from a simple, tractable prior, typically a Gaussian distribution. However, when the support of this prior differs greatly from that of the target distribution, diffusion models often struggle to explore effectively or suffer from large discretization errors. Moreover, learning the prior distribution can lead to mode-collapse, exacerbated by the mode-seeking nature of reverse Kullback-Leibler divergence commonly used in VI.
To address these challenges, we propose end-to-end learnable Gaussian mixture priors (GMPs). GMPs offer improved control over exploration, adaptability to target support, and increased expressiveness to counteract mode collapse. We further leverage the structure of mixture models by proposing a strategy to iteratively refine the model through the addition of mixture components during training. Our experimental results demonstrate significant performance improvements across a diverse range of real-world and synthetic benchmark problems when using GMPs without requiring additional target evaluations.
|
Variational Inference, Sampling, Diffusion Models, Mixture Models
|
We focus on improving existing diffusion-based sampling methods via end-to-end learning of Gaussian mixture priors
| 11,724 |
2503.00524
|
[
0.00628522178158164,
-0.13393691182136536,
0.11127926409244537,
0.08633960038423538,
0.04602433741092682,
0.007994681596755981,
-0.0027293353341519833,
-0.04711252078413963,
0.000048093828809214756,
-0.05550723895430565,
-0.00839628092944622,
-0.02330218069255352,
-0.03600307181477547,
0.027259595692157745,
0.014099967665970325,
-0.07731080055236816,
0.08073319494724274,
0.08731593191623688,
-0.05362193286418915,
0.03335576876997948,
-0.0062407683581113815,
0.07323227822780609,
0.01051999069750309,
-0.03490418940782547,
0.04469158872961998,
0.07515811175107956,
0.06741085648536682,
-0.04240117594599724,
0.047291144728660583,
-0.058857399970293045,
0.10957103967666626,
0.017354145646095276,
-0.04483529552817345,
-0.03322954475879669,
-0.042940396815538406,
0.03311451897025108,
-0.05677216872572899,
0.041750289499759674,
0.044104237109422684,
0.0353904664516449,
-0.015265279449522495,
0.012547212652862072,
-0.011521373875439167,
0.06710564345121384,
0.06511758267879486,
-0.037048447877168655,
0.0026938735973089933,
-0.08444534987211227,
0.07949405908584595,
0.02726295031607151,
0.09566283226013184,
-0.05274539068341255,
-0.05055483058094978,
0.017635740339756012,
0.01552833616733551,
0.0501263253390789,
0.03181713446974754,
-0.007081279531121254,
0.01775718852877617,
-0.009859520010650158,
-0.06099662929773331,
-0.1222110465168953,
-0.05186530947685242,
-0.07951653748750687,
-0.008317858912050724,
0.005845639389008284,
0.016523068770766258,
0.05731149762868881,
0.03850522264838219,
0.09472252428531647,
0.01825147680938244,
0.08254854381084442,
-0.09722229093313217,
0.01427985168993473,
-0.02984030172228813,
0.005830344744026661,
-0.022250020876526833,
-0.03561374545097351,
0.014124483801424503,
-0.03199281170964241,
0.04375240206718445,
-0.0020548920147120953,
0.011606195010244846,
0.0359165258705616,
0.05643988773226738,
-0.055467620491981506,
0.026473689824342728,
0.09772010147571564,
0.07498571276664734,
-0.007830125279724598,
-0.04757842421531677,
0.08286890387535095,
-0.030276454985141754,
-0.0508616603910923,
0.04584016650915146,
0.01171338651329279,
0.00657687010243535,
-0.0910448357462883,
0.09351431578397751,
0.061396945267915726,
0.0016982168890535831,
0.026456184685230255,
0.03923648223280907,
-0.03280532360076904,
-0.0036043787840753794,
-0.12877264618873596,
0.0007107057608664036,
0.051223043352365494,
0.009279967285692692,
0.006038280203938484,
0.003996309358626604,
0.027910156175494194,
0.06935356557369232,
-0.01996195875108242,
0.046513427048921585,
-0.03731995075941086,
0.038915570825338364,
-0.0038992734625935555,
0.023714276030659676,
0.05427587777376175,
0.02447531931102276,
-0.02481687068939209,
-0.05280541256070137,
0.0632421225309372,
0.006451557856053114,
0.009183917194604874,
-0.011726238764822483,
3.251656881348427e-33,
-0.020815830677747726,
-0.06962312012910843,
0.06923527270555496,
0.07261356711387634,
0.04311167448759079,
0.04045324772596359,
-0.029482796788215637,
-0.02522137202322483,
-0.03298109024763107,
0.014598801732063293,
-0.009917127899825573,
0.04063235595822334,
-0.16281716525554657,
0.10205981880426407,
-0.0005947930621914566,
-0.0026953488122671843,
-0.088243268430233,
0.08585777878761292,
-0.03925115987658501,
-0.046899329870939255,
0.03671395778656006,
-0.03747812286019325,
-0.05428040400147438,
-0.05290685594081879,
-0.04679781198501587,
-0.017805391922593117,
0.07922100275754929,
-0.03635722026228905,
0.015737447887659073,
-0.014186445623636246,
-0.09929601848125458,
-0.032766520977020264,
-0.06317035853862762,
0.04680183157324791,
0.033988870680332184,
0.061252299696207047,
-0.019695792347192764,
-0.004910406190901995,
-0.05999727547168732,
-0.03372248262166977,
-0.032489970326423645,
0.04504500329494476,
-0.017877567559480667,
-0.012617269530892372,
-0.15687532722949982,
-0.06861215084791183,
0.008697203360497952,
-0.06630116701126099,
0.006004693452268839,
-0.07776959985494614,
-0.03030260093510151,
-0.031128495931625366,
-0.011294153518974781,
-0.07749704271554947,
-0.04694003611803055,
0.017310643568634987,
0.03413988649845123,
0.026672976091504097,
-0.050137992948293686,
0.0044142757542431355,
0.04635519161820412,
0.0447496622800827,
-0.003145519644021988,
0.03625804930925369,
0.01566491462290287,
0.029771825298666954,
-0.04866298288106918,
0.08691070228815079,
-0.05483127757906914,
0.028636405244469643,
-0.011595534160733223,
0.0330340601503849,
-0.038145679980516434,
-0.11194375902414322,
0.06029006838798523,
0.004518262576311827,
0.029873069375753403,
0.0002965546736959368,
0.017338868230581284,
-0.014617426320910454,
-0.0000895484845386818,
-0.02870657667517662,
-0.06599597632884979,
-0.03234970569610596,
0.008463027887046337,
0.016149621456861496,
-0.05034564062952995,
-0.07918401062488556,
-0.0941934585571289,
-0.05924907326698303,
-0.08418437093496323,
-0.027803057804703712,
-0.017742900177836418,
0.058163419365882874,
-0.00031907574157230556,
-3.1391609699487805e-33,
-0.03626322001218796,
0.0346396267414093,
0.02892756648361683,
0.03150935098528862,
0.017727650701999664,
0.00634352071210742,
0.016185084357857704,
0.01044223178178072,
0.08305519819259644,
-0.07767495512962341,
-0.02592693641781807,
-0.0006212450098246336,
0.02580496296286583,
-0.013217099942266941,
-0.004076811950653791,
0.011985905468463898,
-0.02612881548702717,
0.08424855768680573,
0.03673172742128372,
-0.007984654977917671,
-0.05716697499155998,
0.0011842866661027074,
-0.0017609838396310806,
0.01830185577273369,
-0.05585513636469841,
0.031153451651334763,
-0.02224615402519703,
0.11305944621562958,
0.007860059849917889,
-0.02265303209424019,
0.008753936737775803,
-0.023798806592822075,
0.02136967144906521,
-0.0445382297039032,
-0.1338568925857544,
0.1309751719236374,
0.03572038933634758,
0.06417499482631683,
-0.051867954432964325,
0.03338978812098503,
0.022518884390592575,
0.0009358952520415187,
0.0033346964046359062,
-0.002573565347120166,
0.05555931106209755,
-0.034606918692588806,
0.059032440185546875,
-0.028028428554534912,
0.03885620832443237,
0.008364494889974594,
-0.023171961307525635,
0.0913994088768959,
-0.12912771105766296,
0.045696258544921875,
-0.0724409893155098,
-0.021131783723831177,
-0.03456081449985504,
-0.06304380297660828,
0.05622868984937668,
-0.028022635728120804,
-0.005473265890032053,
-0.028773339465260506,
0.058811575174331665,
-0.0771368220448494,
-0.06571806967258453,
-0.0733562707901001,
-0.02174651063978672,
0.01684771664440632,
-0.008806359954178333,
0.014381811954081059,
-0.016034414991736412,
-0.029541440308094025,
0.09827920794487,
0.013316280208528042,
0.0724034532904625,
-0.0812644213438034,
0.028001239523291588,
0.015509748831391335,
-0.013089846819639206,
-0.02569812908768654,
0.12107909470796585,
-0.008164106868207455,
-0.010742587968707085,
0.000405885191867128,
0.07847397774457932,
0.08021719753742218,
-0.03906476870179176,
0.055702924728393555,
0.0334261991083622,
0.02206948772072792,
-0.043330781161785126,
-0.036633431911468506,
0.0447181835770607,
0.127484530210495,
-0.09618809074163437,
-4.2649347165024665e-8,
-0.07317880541086197,
0.02091747336089611,
-0.002611567499116063,
-0.03875275328755379,
0.08080597221851349,
0.0010237630922347307,
-0.029848812147974968,
0.034908443689346313,
-0.015319060534238815,
-0.07507673650979996,
0.05734860524535179,
-0.0008489847532473505,
-0.030144434422254562,
-0.016395365819334984,
-0.029065502807497978,
0.09554736316204071,
0.026397638022899628,
-0.028391292318701744,
0.016166692599654198,
-0.015990911051630974,
0.0023461123928427696,
0.01360307913273573,
0.015096188522875309,
-0.032884061336517334,
0.11472810804843903,
0.0031872361432760954,
0.000377511401893571,
0.07139366865158081,
0.09234271198511124,
-0.04119235277175903,
-0.004265557508915663,
0.015477337874472141,
-0.06237582117319107,
0.039129361510276794,
-0.05469189211726189,
0.13351315259933472,
0.020461872220039368,
0.007610649336129427,
-0.01936241425573826,
-0.05599702149629593,
-0.026959843933582306,
-0.005416983272880316,
-0.006140392739325762,
0.024310633540153503,
-0.008561486378312111,
0.08644701540470123,
0.0533212274312973,
-0.012082592584192753,
0.04603751376271248,
0.018228868022561073,
0.05164922773838043,
-0.06520160287618637,
0.022136040031909943,
0.05924302712082863,
0.0635811910033226,
-0.011083238758146763,
-0.02557791955769062,
-0.005120815709233284,
0.020753230899572372,
0.01754978857934475,
-0.019382670521736145,
-0.03723623976111412,
-0.04724989831447601,
0.05739046633243561
] | 0 | 0 | 0 | 0 |
|
IterGen: Iterative Semantic-aware Structured LLM Generation with Backtracking
|
https://openreview.net/forum?id=ac93gRzxxV
|
[
"Shubham Ugare",
"Rohan Gumaste",
"Tarun Suresh",
"Gagandeep Singh",
"Sasa Misailovic"
] |
Poster
|
Large Language Models (LLMs) are widely used for tasks such as natural language and code generation, but their outputs often suffer from issues like hallucination, toxicity, and incorrect results. Current libraries for structured LLM generation rely on left-to-right decoding without support for backtracking, limiting the ability to correct or refine outputs mid-generation.
To address this, we introduce IterGen, a user-friendly library for iterative, grammar-guided LLM generation that enables users to move both forward and backward within the generated output based on grammar symbols.
By leveraging a symbol-to-position mapping and maintaining the key-value (KV) cache state, IterGen ensures efficient and structured generation while allowing for corrections during the process. We demonstrate IterGen's effectiveness in two important applications: reducing privacy leakage in LLM outputs, improving the accuracy of LLM-generated SQL and Vega-Lite queries.
Our code and additional resources are available at https://structuredllm.com.
|
LLM, Grammar, Formal Languages, Parser, Decoding
| null | 11,719 |
2410.07295
|
[
-0.015647776424884796,
-0.05066373571753502,
0.009636967442929745,
0.044212955981492996,
0.0007113610627129674,
-0.016492696478962898,
-0.04321026802062988,
0.012117003090679646,
-0.028970275074243546,
-0.002417094772681594,
-0.023699477314949036,
-0.07252189517021179,
0.08353777974843979,
-0.061672840267419815,
0.06326992809772491,
0.04177476465702057,
0.021038195118308067,
0.04631563276052475,
-0.027420954778790474,
-0.08801060169935226,
-0.0022257212549448013,
0.09042420238256454,
-0.003515996504575014,
0.034256067126989365,
0.039743561297655106,
0.04419030621647835,
0.006122559774667025,
0.013530570082366467,
0.06324754655361176,
-0.023972278460860252,
0.04194502905011177,
0.021774522960186005,
-0.046429142355918884,
0.07537318021059036,
0.006164349615573883,
0.08052494376897812,
-0.0206783227622509,
-0.04428815841674805,
-0.02121557481586933,
-0.09955769032239914,
-0.06078753620386124,
0.006042742170393467,
0.007208418101072311,
0.03394130989909172,
0.02324596606194973,
0.0074653965421020985,
0.010271587409079075,
-0.014871444553136826,
-0.0351053923368454,
-0.008916511200368404,
0.016397345811128616,
0.01874159835278988,
0.017539389431476593,
-0.034078579396009445,
-0.03697321191430092,
0.03218939155340195,
0.025604594498872757,
0.01757536455988884,
-0.08811726421117783,
-0.06249497830867767,
-0.02105659805238247,
-0.08430229872465134,
-0.03297698497772217,
0.0008280957699753344,
-0.026754993945360184,
-0.04112516716122627,
0.07071986794471741,
0.0330374501645565,
0.0030648461543023586,
0.015196922235190868,
-0.07822416722774506,
-0.017286118119955063,
-0.010723775252699852,
0.10258471220731735,
-0.04647141695022583,
0.10833641141653061,
0.003442835994064808,
-0.014565594494342804,
0.07865425944328308,
-0.07307981699705124,
0.023082289844751358,
-0.018554117530584335,
0.03286696597933769,
0.0737570971250534,
0.04800332337617874,
-0.024091655388474464,
0.08243439346551895,
0.03538508340716362,
0.07420768588781357,
0.1350775510072708,
-0.012884184718132019,
-0.08800875395536423,
0.11853278428316116,
0.03320235013961792,
-0.015208027325570583,
0.04269078001379967,
0.06486836820840836,
-0.06429177522659302,
-0.021210387349128723,
0.05249560996890068,
0.023495763540267944,
0.1095600426197052,
0.031313978135585785,
-0.10513748228549957,
-0.02711152844130993,
0.022656485438346863,
0.029155224561691284,
0.02671062760055065,
0.015257752500474453,
-0.022047480568289757,
0.055336691439151764,
0.11129780858755112,
0.017358744516968727,
-0.047345440834760666,
0.03202719986438751,
0.010450786910951138,
-0.02302071638405323,
-0.0757000669836998,
0.09862519055604935,
0.004306333605200052,
-0.02724483422935009,
-0.02816111221909523,
-0.05660663917660713,
0.11124423891305923,
-0.05116398632526398,
-0.11155453324317932,
-0.04004671424627304,
-1.0510443965587473e-33,
0.026870612055063248,
-0.017844336107373238,
-0.03340963274240494,
0.03069452941417694,
0.02166862227022648,
0.04182601347565651,
0.039513446390628815,
0.060853511095047,
-0.10500228404998779,
0.024037275463342667,
0.046096108853816986,
-0.06930211186408997,
-0.0404585637152195,
0.06447534263134003,
0.030395610257983208,
-0.014160603284835815,
-0.0218611229211092,
0.041025254875421524,
-0.09642627090215683,
-0.010755988769233227,
0.031149689108133316,
0.030163202434778214,
-0.01710888184607029,
-0.01920689083635807,
0.06699072569608688,
0.09958389401435852,
0.06255806237459183,
-0.054840732365846634,
-0.042559925466775894,
0.01532005239278078,
-0.0449230782687664,
-0.08377429097890854,
0.01463429257273674,
-0.0257076658308506,
0.021284019574522972,
-0.03724760562181473,
-0.00256145978346467,
-0.010411676950752735,
0.03282247856259346,
0.004727560561150312,
0.027432747185230255,
0.024419866502285004,
-0.025601612403988838,
0.015590175054967403,
-0.060137517750263214,
0.0007424759096466005,
-0.0305190272629261,
0.012873273342847824,
-0.04223315417766571,
-0.06401383131742477,
0.031169524416327477,
0.08441391587257385,
-0.058694351464509964,
0.02976291999220848,
-0.002735419664531946,
0.01514777634292841,
0.027282902970910072,
0.019831541925668716,
0.03366454318165779,
0.03029150329530239,
0.005869804881513119,
0.10732494294643402,
0.004275482147932053,
0.08594062179327011,
0.005741503089666367,
-0.00933169573545456,
0.01417546160519123,
0.02303042821586132,
0.02113412320613861,
0.005610767286270857,
-0.06442465633153915,
-0.039754562079906464,
-0.03285909444093704,
0.002988748950883746,
-0.033724360167980194,
0.005437356419861317,
0.010080153122544289,
-0.0784263014793396,
-0.007553284987807274,
-0.05910373106598854,
0.04995948448777199,
0.03323220834136009,
-0.03932766988873482,
-0.012032568454742432,
0.00963683519512415,
-0.11993985623121262,
0.008303312584757805,
-0.039866287261247635,
-0.04897942394018173,
0.014442358165979385,
-0.008081139996647835,
-0.09585315734148026,
0.09219858795404434,
0.012164081446826458,
-0.08704313635826111,
-1.0328718967435078e-33,
-0.07664942741394043,
-0.05242717266082764,
-0.02548280730843544,
0.08926135301589966,
-0.04504847154021263,
-0.11487633734941483,
-0.005778249818831682,
0.04943448677659035,
-0.022451624274253845,
-0.0406336672604084,
-0.08189386874437332,
0.010829634964466095,
0.013797684572637081,
0.008066846057772636,
0.10866496711969376,
0.032331231981515884,
0.039005547761917114,
-0.014680817723274231,
-0.03320594131946564,
0.10460947453975677,
-0.051847051829099655,
0.10371439158916473,
-0.05771152675151825,
0.04680636152625084,
-0.040391288697719574,
0.0476321317255497,
-0.029486529529094696,
0.06307587772607803,
0.00024052544904407114,
-0.008696606382727623,
-0.003915201406925917,
0.051553938537836075,
0.04109758138656616,
-0.1230684220790863,
-0.08091920614242554,
-0.032506294548511505,
-0.0053727226331830025,
-0.009368005208671093,
-0.004354186821728945,
0.0397317111492157,
0.07595881074666977,
-0.03293420001864433,
-0.04143858700990677,
0.052579376846551895,
-0.044197771698236465,
0.03889396786689758,
-0.15845082700252533,
0.026862889528274536,
0.028452666476368904,
-0.008794184774160385,
0.05865786597132683,
-0.009754610247910023,
-0.010481178760528564,
-0.03530097007751465,
0.015629766508936882,
-0.06666732579469681,
-0.04166345298290253,
-0.04280726611614227,
-0.053361956030130386,
0.006737406365573406,
-0.08109892904758453,
-0.05893531069159508,
0.08163394778966904,
-0.10753196477890015,
0.042709220200777054,
-0.051071953028440475,
-0.0915001779794693,
-0.04316491633653641,
-0.02410459704697132,
-0.028907405212521553,
0.09486129879951477,
-0.05826931446790695,
-0.08404959738254547,
0.06512327492237091,
0.050129931420087814,
-0.012138496153056622,
0.010256079956889153,
-0.05209876596927643,
0.03938223049044609,
-0.015515568666160107,
0.022238653153181076,
0.04560549929738045,
0.05991080403327942,
0.020083094015717506,
0.04920537397265434,
-0.05778022110462189,
0.03010019287467003,
0.06813140213489532,
-0.055825334042310715,
0.05390246585011482,
-0.022273289039731026,
0.04372897371649742,
-0.03728324919939041,
0.006589899770915508,
-0.03852766007184982,
-4.030290767786937e-8,
-0.060182228684425354,
-0.05861777067184448,
-0.029750360175967216,
0.07207992672920227,
0.014000066556036472,
-0.05809115245938301,
0.008421345613896847,
0.06433876603841782,
0.029782751575112343,
-0.08382123708724976,
0.026890086010098457,
-0.024551263079047203,
-0.0640411227941513,
-0.018182426691055298,
0.012627528049051762,
0.0811481848359108,
-0.028626270592212677,
0.014891324564814568,
-0.061736706644296646,
0.0006384000880643725,
-0.04491161182522774,
0.09380246698856354,
-0.04614099860191345,
-0.028643853962421417,
0.058210890740156174,
-0.010855816304683685,
0.036869026720523834,
0.06467071920633316,
0.13649562001228333,
-0.01690046675503254,
0.026228701695799828,
-0.021967411041259766,
-0.02824995294213295,
-0.010926119983196259,
-0.0052963909693062305,
0.061108049005270004,
-0.022874077782034874,
-0.034683696925640106,
0.043182723224163055,
-0.00036150027881376445,
0.08370574563741684,
0.014599323272705078,
-0.03675385192036629,
0.07916951924562454,
-0.08154387772083282,
-0.03246505558490753,
0.0013119811192154884,
-0.05367531627416611,
0.011426649987697601,
-0.11686331778764725,
-0.04139304161071777,
-0.05908066779375076,
-0.011705431155860424,
-0.030550064519047737,
-0.01659991964697838,
0.06507383286952972,
-0.06951814889907837,
0.03215419128537178,
0.07302276045084,
0.010584448464214802,
-0.0077789840288460255,
0.0058366646990180016,
-0.01638984866440296,
-0.004755963571369648
] |
https://github.com/uiuc-arc/itergen
| 12 | 0 | 0 | 0 |
OpenPRM: Building Open-domain Process-based Reward Models with Preference Trees
|
https://openreview.net/forum?id=fGIqGfmgkW
|
[
"Kaiyan Zhang",
"Jiayuan Zhang",
"Haoxin Li",
"Xuekai Zhu",
"Ermo Hua",
"Xingtai Lv",
"Ning Ding",
"Biqing Qi",
"Bowen Zhou"
] |
Poster
|
Scaling inference-time computation is increasingly seen as the next frontier in scaling laws for large language models. Previous work in mathematics and coding has demonstrated the remarkable potential for inference-time scaling. During such scaling, fine-grained supervision through process-based reward models (PRMs) is essential for enhancement. However, exploration of inference-time scaling and PRMs in open-domain problems remains limited, where lacking exact answers and obtaining process supervision prove challenging. In this paper, we explore the construction of PRMs for open-domain tasks, specifically for instruction-following tasks. Utilizing existing outcome-based reward models (ORMs), we develop sentence-level preference trees based on the prefix similarity of parallel sampled candidates from datasets like UltraFeedback. This setup allows us to derive weak supervision for processes via back-propagation from outcome-level rewards. Subsequently, we integrate ORMs and PRMs under the same pairwise ranking objectives, resulting in our newly developed reward models, named OpenPRM. This approach significantly enhances the scalability of process-level supervision in open domains at minimal cost. We assess the performance of OpenPRM across various reward benchmarks, demonstrating its competitive edge over traditional ORMs in open domains and PRMs in specialized domains. Additionally, we investigate the scalability of inference-time computation for open-domain instructions. Our results highlight the limitations of ORMs’ scalability, while OpenPRM shows superior performance in scaled settings. Despite these advances, achieving automatic fine-grained supervision for open-domain inference-time scaling remains a substantial challenge. We hope these findings will spur further development of process supervision reward models in open-domain scenarios.
|
large language models, reward models, open-domain instruction following
| null | 11,716 | null |
[
-0.0770791694521904,
-0.07159756869077682,
0.0046613821759819984,
0.01738620549440384,
0.04523435980081558,
0.03459257632493973,
-0.013478419743478298,
0.05007033050060272,
0.046193260699510574,
0.01572606712579727,
-0.017364900559186935,
-0.04297255352139473,
0.03220519423484802,
0.08244428038597107,
0.04633626714348793,
0.0037203016690909863,
0.046366915106773376,
0.024477194994688034,
-0.09234192967414856,
-0.08871173113584518,
0.03170904517173767,
0.05101696029305458,
0.050971128046512604,
-0.00972787756472826,
-0.029385754838585854,
-0.006912049371749163,
0.0191795714199543,
-0.01417402271181345,
0.050537411123514175,
0.019418824464082718,
0.03313925117254257,
-0.041151922196149826,
0.020683931186795235,
-0.06860543042421341,
-0.07874522358179092,
0.08617256581783295,
-0.10033556073904037,
-0.062276121228933334,
-0.09286370128393173,
-0.04979921132326126,
-0.0012518060393631458,
-0.006469463463872671,
0.003288557752966881,
0.06567800790071487,
0.044418442994356155,
-0.031623147428035736,
-0.009882448241114616,
-0.0474463514983654,
-0.041641466319561005,
-0.001424059271812439,
-0.0178166925907135,
-0.036446817219257355,
-0.0026108953170478344,
0.010928953066468239,
0.001464350032620132,
0.06902599334716797,
0.06721009314060211,
-0.026975857093930244,
0.009573790244758129,
-0.07078149169683456,
-0.1257941871881485,
-0.12144505977630615,
-0.0910005196928978,
-0.04369243606925011,
-0.007634325884282589,
-0.05557793751358986,
-0.0009836293756961823,
0.003464876674115658,
-0.06384459882974625,
0.002711256965994835,
-0.009213658981025219,
0.11978062987327576,
-0.008709645830094814,
0.07220549136400223,
0.02527540922164917,
0.05535275861620903,
-0.052155692130327225,
-0.026198824867606163,
-0.025740908458828926,
-0.011220071464776993,
-0.05461941286921501,
-0.040273379534482956,
0.02208404801785946,
0.004024879541248083,
0.1078772246837616,
0.002352816518396139,
0.06588699668645859,
0.06448093801736832,
0.05692067742347717,
0.04016854986548424,
-0.04020962119102478,
-0.03268389776349068,
-0.004044598434120417,
0.018432889133691788,
-0.032186929136514664,
0.012139826081693172,
-0.10260973125696182,
-0.057098232209682465,
-0.017275460064411163,
0.061498790979385376,
0.020485511049628258,
0.14188620448112488,
-0.03839029744267464,
-0.05037619173526764,
0.01806000806391239,
0.018608039245009422,
-0.01201640535145998,
0.0819922685623169,
0.03978285938501358,
-0.12975770235061646,
0.04811904951930046,
0.017352523282170296,
0.09896272420883179,
0.007393485400825739,
-0.008418690413236618,
-0.055777668952941895,
0.012989702634513378,
-0.07846765965223312,
0.0199478380382061,
0.03276796638965607,
0.0008858973160386086,
0.02194702811539173,
-0.006266097538173199,
-0.02553250640630722,
0.007382047828286886,
0.029532799497246742,
0.005870393011718988,
6.263242826934907e-34,
0.08189786970615387,
0.04351770877838135,
0.09927953034639359,
-0.026309790089726448,
0.06843210756778717,
0.028660520911216736,
0.017006555572152138,
0.03693854808807373,
-0.10943365097045898,
-0.02178431861102581,
0.035302478820085526,
0.03307436779141426,
-0.09968363493680954,
0.06227273866534233,
0.07604668289422989,
-0.0320570282638073,
0.028207851573824883,
0.04210368171334267,
-0.012003296986222267,
-0.00032296477002091706,
0.03893229737877846,
0.02191973850131035,
-0.006802309770137072,
-0.019331814721226692,
0.07518196105957031,
-0.022470884025096893,
0.016892507672309875,
-0.08209662139415741,
-0.018164673820137978,
-0.04202871397137642,
-0.011962474323809147,
-0.044798363000154495,
-0.02155163325369358,
0.005439008120447397,
0.009682507254183292,
0.00280407234095037,
0.02132066898047924,
-0.047660499811172485,
0.006406556814908981,
-0.05592828989028931,
0.0011275919387117028,
-0.0279399361461401,
0.05029812455177307,
0.03471776098012924,
-0.08694475889205933,
-0.1072223037481308,
0.018193433061242104,
0.006254938431084156,
-0.01272116880863905,
-0.041698478162288666,
0.08223561197519302,
0.028234397992491722,
0.037796054035425186,
-0.07189984619617462,
-0.033176857978105545,
0.020009610801935196,
-0.01224543433636427,
0.05060437321662903,
0.006606938783079386,
0.06556516140699387,
0.054090067744255066,
0.035940513014793396,
-0.03750905022025108,
0.03868681192398071,
-0.04444478079676628,
0.028392426669597626,
-0.046027105301618576,
0.015853723511099815,
0.11278335005044937,
0.0457736998796463,
-0.014085580594837666,
0.03258282318711281,
-0.02893272414803505,
-0.08880021423101425,
0.08121392875909805,
0.07989569753408432,
0.02644062601029873,
-0.13211439549922943,
0.03762221708893776,
-0.028546342626214027,
0.00882489699870348,
-0.0425112321972847,
-0.013841341249644756,
-0.02897118590772152,
0.006103968247771263,
-0.023463774472475052,
0.0053237066604197025,
-0.08636375516653061,
-0.005136172287166119,
-0.01118999533355236,
-0.04717700928449631,
0.007930787280201912,
-0.034052059054374695,
0.04181937500834465,
0.012585584074258804,
-8.52194741602276e-34,
0.0056070731952786446,
-0.04946855083107948,
0.011774297803640366,
0.03709765523672104,
0.021852098405361176,
-0.015439006499946117,
-0.07624182105064392,
-0.02493673376739025,
-0.030635733157396317,
-0.06558486819267273,
-0.04451753944158554,
-0.0007803920889273286,
0.057792339473962784,
-0.003907632082700729,
0.02879181317985058,
-0.04093945026397705,
-0.02087981067597866,
0.11764687299728394,
0.008446167223155499,
0.04343975707888603,
0.010329035110771656,
0.19014586508274078,
-0.07959296554327011,
0.059817735105752945,
0.006092533469200134,
-0.015520543791353703,
0.03982239216566086,
0.02470492571592331,
-0.03290729224681854,
0.012644588947296143,
-0.03840502351522446,
-0.06216855347156525,
-0.018713779747486115,
-0.037547383457422256,
-0.05680614709854126,
0.011116975918412209,
0.07086628675460815,
-0.04416552558541298,
-0.07682953029870987,
0.11326956003904343,
0.0013548046117648482,
-0.06134385988116264,
-0.03269631415605545,
-0.031063903123140335,
-0.03039270080626011,
0.021604269742965698,
-0.12270352244377136,
-0.015236129984259605,
-0.036515671759843826,
-0.011047090403735638,
-0.07575952261686325,
0.08957359194755554,
-0.016978003084659576,
0.08106029778718948,
-0.02569618634879589,
-0.11583910137414932,
-0.056506045162677765,
-0.10875425487756729,
0.11764809489250183,
-0.019778629764914513,
-0.009639986790716648,
0.04803141951560974,
0.05134408921003342,
-0.027596216648817062,
0.04358281195163727,
-0.05747772753238678,
0.050885263830423355,
0.02609209157526493,
-0.04080110788345337,
-0.05219929292798042,
-0.03330501168966293,
0.024608757346868515,
0.044111885130405426,
0.00700024701654911,
0.015494639053940773,
0.011836778372526169,
0.010162347927689552,
-0.026984700933098793,
-0.039952244609594345,
0.00028768149786628783,
0.05937553569674492,
-0.002502861898392439,
0.04343968629837036,
0.0024147722870111465,
0.035653192549943924,
0.05744240805506706,
0.14057473838329315,
0.10699685662984848,
0.031007064506411552,
0.03995035961270332,
-0.035262174904346466,
0.02209007740020752,
0.056020088493824005,
0.05459502339363098,
-0.03201950341463089,
-3.956416705364063e-8,
-0.12328187376260757,
0.008796541020274162,
-0.0013702362775802612,
0.09530708938837051,
-0.033560313284397125,
-0.033265501260757446,
-0.0063746836967766285,
0.03635008633136749,
-0.03563854470849037,
-0.01407195720821619,
0.06669149547815323,
0.03432459011673927,
0.002766094636172056,
-0.0909922644495964,
0.019285807386040688,
0.09473177045583725,
0.058445751667022705,
-0.0030809594318270683,
0.034325920045375824,
-0.025273486971855164,
0.013698511756956577,
0.05164323374629021,
0.04529991373419762,
0.03527265787124634,
-0.03179222717881203,
-0.08029024302959442,
0.053164999932050705,
0.10980341583490372,
0.09512536972761154,
0.00003613683657022193,
0.014677069149911404,
-0.026681330054998398,
0.021703166887164116,
-0.006520826369524002,
0.050310101360082626,
0.0003532604605425149,
0.005462474189698696,
-0.002632815856486559,
0.0038636671379208565,
-0.01743902638554573,
0.03449862822890282,
0.021316934376955032,
-0.0045656426809728146,
0.01074815820902586,
-0.006992327515035868,
0.02882678247988224,
-0.05159655585885048,
-0.08233615010976791,
0.009202050045132637,
-0.0420251190662384,
0.024644197896122932,
-0.04113021120429039,
-0.001542355865240097,
0.011472072452306747,
0.06292524188756943,
0.04947961866855621,
-0.0148135581985116,
-0.04295354336500168,
-0.034294016659259796,
0.04837680980563164,
0.08876913040876389,
-0.03399597108364105,
-0.034724026918411255,
0.050731636583805084
] | 0 | 0 | 0 | 0 |
|
SPAM: Spike-Aware Adam with Momentum Reset for Stable LLM Training
|
https://openreview.net/forum?id=L9eBxTCpQG
|
[
"Tianjin Huang",
"Ziquan Zhu",
"Gaojie Jin",
"Lu Liu",
"Zhangyang Wang",
"Shiwei Liu"
] |
Poster
|
Large Language Models (LLMs) have demonstrated exceptional performance across diverse tasks, yet their training remains highly resource intensive and susceptible to critical challenges such as training instability. A predominant source of this instability stems from gradient and loss spikes, which disrupt the learning process, often leading to costly interventions like checkpoint recovery and experiment restarts, further amplifying inefficiencies. This paper presents a comprehensive investigation into gradient spikes observed during LLM training, revealing their prevalence across multiple architectures and datasets. Our analysis shows that these spikes can be up to 1000× larger than typical gradients, substantially deteriorating model performance. To address this issue, we propose Spike-Aware Adam with Momentum Reset (SPAM), a novel optimizer designed to counteract gradient spikes through momentum reset and spike-aware gradient clipping. Extensive experiments, including both pre-training and fine-tuning, demonstrate that SPAM consistently surpasses Adam and its variants across a range of model scales. Additionally, SPAM facilitates memory-efficient training by enabling sparse momentum, where only a subset of momentum terms are maintained and updated. When operating under memory constraints, SPAM outperforms state-of-the-art memory-efficient optimizers such as GaLore and Adam-Mini. Our work underscores the importance
of mitigating gradient spikes in LLM training and introduces an effective optimization strategy that enhances both training stability and resource efficiency at scale. Code is submitted.
|
Gradient Spikes, Spike-Aware Adam, LLMs
|
We present SPAM , an optimizer that mitigates gradient spikes in LLM training, improving stability and efficiency.
| 11,699 |
2501.06842
|
[
-0.04953807592391968,
-0.13452591001987457,
0.04797261208295822,
0.06542449444532394,
0.03453405573964119,
0.030665237456560135,
0.008729823864996433,
-0.0517294816672802,
0.04428695887327194,
-0.07509098201990128,
-0.01853932812809944,
0.018099110573530197,
-0.01884332299232483,
0.012156828306615353,
-0.02305246703326702,
0.013265120796859264,
0.05130181834101677,
0.09463947266340256,
-0.06871973723173141,
-0.03330036625266075,
-0.0031760793644934893,
0.010524330660700798,
0.036478083580732346,
0.047036390751600266,
0.07115476578474045,
-0.01661076582968235,
-0.042022738605737686,
-0.04208020493388176,
0.022371530532836914,
-0.05526180565357208,
0.0943697988986969,
-0.03770449012517929,
-0.07148107886314392,
0.008723250590264797,
-0.025252627208828926,
0.02333253249526024,
-0.06633470207452774,
0.011354039423167706,
0.02118244767189026,
-0.04342004284262657,
0.0262366384267807,
-0.05254998803138733,
-0.016616547480225563,
0.006293123122304678,
0.08896476030349731,
0.012123004533350468,
0.02845004014670849,
-0.019116835668683052,
0.026375705376267433,
-0.02939019165933132,
-0.03206943720579147,
-0.0771455243229866,
0.0510057657957077,
0.030963705852627754,
0.002824585884809494,
0.014935285784304142,
0.04694139212369919,
0.10714302957057953,
-0.010780636221170425,
0.024911627173423767,
-0.03149457275867462,
-0.14212515950202942,
-0.022379646077752113,
-0.013459905050694942,
-0.04241873323917389,
0.015200181864202023,
0.028510045260190964,
0.0482344813644886,
0.03304554894566536,
0.08420024067163467,
0.05940644070506096,
0.06736583262681961,
-0.04621168598532677,
0.05214693397283554,
0.006920082028955221,
0.0839252844452858,
0.012153932824730873,
-0.024373088032007217,
0.04773308336734772,
-0.04701736569404602,
0.04845106974244118,
0.0007443909416906536,
0.034094396978616714,
-0.04852203279733658,
0.12273237854242325,
0.036268994212150574,
0.06746207177639008,
0.010460940189659595,
0.04909219592809677,
0.03575512021780014,
-0.04307959973812103,
-0.03952724114060402,
0.04532637447118759,
0.028253689408302307,
0.008156424388289452,
0.01710895635187626,
-0.0017827474512159824,
-0.05262111499905586,
-0.06407836824655533,
0.09327053278684616,
-0.000022820606318418868,
0.017262214794754982,
0.008690506219863892,
0.04053908586502075,
0.009044039994478226,
-0.020876720547676086,
0.08335009962320328,
0.03143347427248955,
0.005819672252982855,
-0.02145032025873661,
0.07886750251054764,
0.053950708359479904,
0.008379167877137661,
0.07537644356489182,
0.06853387504816055,
-0.05034921318292618,
-0.038030173629522324,
-0.04144224897027016,
-0.01838752068579197,
0.07552097737789154,
-0.024536453187465668,
-0.015092241577804089,
-0.023541871458292007,
-0.015099856071174145,
-0.004722376819700003,
0.034776899963617325,
-0.07914625853300095,
1.5529457705423163e-33,
0.059932779520750046,
0.04510960727930069,
-0.03011537715792656,
0.0017958641983568668,
0.05607142671942711,
-0.0522461012005806,
-0.010623644106090069,
0.07174510508775711,
-0.05815228819847107,
-0.022186636924743652,
-0.06576754152774811,
-0.059508223086595535,
-0.029537424445152283,
0.08998899906873703,
-0.018100555986166,
-0.02344582974910736,
-0.041191719472408295,
0.04324746131896973,
0.00013294417294673622,
-0.04328548535704613,
0.048692889511585236,
-0.03398199379444122,
-0.02114557847380638,
-0.06451811641454697,
-0.033714570105075836,
0.03230740875005722,
0.07897783815860748,
-0.047644250094890594,
-0.04845146834850311,
0.03038465417921543,
-0.11292979121208191,
0.0333564430475235,
0.023721277713775635,
0.020267479121685028,
0.003837714670225978,
-0.0022576365154236555,
0.011854567565023899,
-0.05077984556555748,
0.018224703148007393,
-0.005217696540057659,
-0.012854162603616714,
0.07321620732545853,
-0.014693103730678558,
-0.05748322233557701,
-0.1252366304397583,
-0.1134372279047966,
0.03373990207910538,
-0.002052535768598318,
-0.028472501784563065,
-0.0909339115023613,
0.028992868959903717,
0.04783031716942787,
0.0070856185629963875,
-0.011712281033396721,
-0.040919866412878036,
-0.043072041124105453,
0.0055367061868309975,
0.00040202122181653976,
-0.023224202916026115,
0.026410743594169617,
0.0559159480035305,
0.03571552410721779,
-0.026062142103910446,
0.06281138956546783,
0.016479000449180603,
0.004548333585262299,
-0.03077239729464054,
0.09638253599405289,
-0.002617551013827324,
0.0067773922346532345,
-0.007757022976875305,
-0.04637826606631279,
-0.045928239822387695,
-0.03456234931945801,
0.0288167092949152,
-0.036430273205041885,
0.09192410111427307,
-0.05675641819834709,
-0.028502557426691055,
0.024122146889567375,
0.04426880180835724,
-0.09320195764303207,
-0.01781398244202137,
-0.025870654731988907,
-0.1125640794634819,
-0.08124092221260071,
0.0575336329638958,
-0.06793215125799179,
-0.030929144471883774,
0.0011269575916230679,
-0.027776317670941353,
-0.04945807531476021,
0.06311836838722229,
0.01033966988325119,
-0.06761915981769562,
-1.2555603934636629e-33,
-0.053773149847984314,
-0.003157052444294095,
-0.06664376705884933,
0.04769757390022278,
-0.0010983541142195463,
-0.009016412310302258,
0.04163013771176338,
0.06588684022426605,
-0.11137248575687408,
-0.10504038631916046,
-0.07330016046762466,
-0.05139092728495598,
0.01895686239004135,
0.00839940831065178,
0.03017861396074295,
-0.10279624164104462,
-0.05946822464466095,
0.021513910964131355,
0.0007927413680590689,
0.019995830953121185,
0.01785070262849331,
0.07142919301986694,
-0.06532517075538635,
0.0035399231128394604,
-0.020036496222019196,
0.040258459746837616,
-0.049002841114997864,
0.08541668951511383,
0.0064064254984259605,
-0.003302769735455513,
-0.011077932082116604,
0.07680882513523102,
0.0036072058137506247,
-0.020733583718538284,
-0.04053482040762901,
0.11566155403852463,
0.03085923381149769,
0.006490550935268402,
0.02345854416489601,
0.04847821593284607,
0.12023723870515823,
0.04508364945650101,
0.019791623577475548,
0.08633437752723694,
0.027530701830983162,
-0.011729937046766281,
-0.06905511021614075,
0.0003467029891908169,
-0.01151793822646141,
0.04084395244717598,
-0.02488250471651554,
0.02658853307366371,
-0.040626998990774155,
0.001658096443861723,
-0.06429816037416458,
-0.11401832103729248,
-0.035045888274908066,
-0.03593659773468971,
0.000013965241123514716,
0.02990056574344635,
-0.09204169362783432,
-0.05403130501508713,
-0.018050432205200195,
-0.06881321966648102,
0.046003133058547974,
-0.010780151933431625,
0.037339042872190475,
0.025169119238853455,
-0.016658253967761993,
-0.008543949574232101,
0.06471209228038788,
0.04064204916357994,
0.0030369439627975225,
0.033449120819568634,
-0.04202263057231903,
-0.03587496280670166,
-0.038928501307964325,
-0.06178189814090729,
-0.028418689966201782,
-0.05213924124836922,
0.016682837158441544,
0.019317906349897385,
-0.029731513932347298,
0.047218888998031616,
0.057799555361270905,
0.15314924716949463,
-0.011488629505038261,
0.11395664513111115,
0.0398990698158741,
0.007700461894273758,
-0.015302210114896297,
-0.01393232960253954,
-0.02824297547340393,
-0.01226618979126215,
-0.04506361111998558,
-3.765272893474503e-8,
-0.061132173985242844,
0.049448929727077484,
-0.03080977313220501,
0.10075165331363678,
0.03762385994195938,
-0.020488658919930458,
-0.03904685005545616,
0.09342486411333084,
0.008842396549880505,
-0.09644051641225815,
0.06551316380500793,
-0.007943611592054367,
-0.04220864921808243,
-0.07262150198221207,
-0.008287247270345688,
0.02830611728131771,
-0.055884331464767456,
0.0916251391172409,
-0.026553118601441383,
-0.016525540500879288,
0.024674024432897568,
0.07487485557794571,
0.02959662675857544,
-0.02410121262073517,
0.030614132061600685,
-0.06823530048131943,
0.01395541150122881,
0.13224150240421295,
0.09148065000772476,
-0.08447977155447006,
-0.020149949938058853,
-0.02255057729780674,
0.03102979063987732,
-0.02471722662448883,
-0.023086247965693474,
0.1537085324525833,
0.009438381530344486,
0.0035299959126859903,
0.03714896738529205,
-0.01713290438055992,
0.00953861977905035,
-0.028719590976834297,
0.04154665768146515,
-0.001684011658653617,
-0.08176619559526443,
-0.0007041087956167758,
-0.07521240413188934,
-0.04891008511185646,
0.06034134328365326,
-0.07987243682146072,
-0.009451904334127903,
0.014898582361638546,
-0.0061528440564870834,
0.12140733748674393,
0.03571116179227829,
-0.014198471792042255,
-0.04152116924524307,
-0.019492320716381073,
0.022174086421728134,
0.05637780949473381,
0.09687864780426025,
-0.041603922843933105,
-0.10285037755966187,
-0.010254879482090473
] |
https://github.com/tianjinyellow/spam-optimizer
| 27 | 0 | 0 | 0 |
Probing the Latent Hierarchical Structure of Data via Diffusion Models
|
https://openreview.net/forum?id=0GzqVqCKns
|
[
"Antonio Sclocchi",
"Alessandro Favero",
"Noam Itzhak Levi",
"Matthieu Wyart"
] |
Poster
|
High-dimensional data must be highly structured to be learnable. Although the compositional and hierarchical nature of data is often put forward to explain learnability, quantitative measurements establishing these properties are scarce. Likewise, accessing the latent variables underlying such a data structure remains a challenge. In this work, we show that forward-backward experiments in diffusion-based models, where data is noised and then denoised to generate new samples, are a promising tool to probe the latent structure of data. We predict in simple hierarchical models that, in this process, changes in data occur by correlated chunks, with a length scale that diverges at a noise level where a phase transition is known to take place. Remarkably, we confirm this prediction in both text and image datasets using state-of-the-art diffusion models. Our results show how latent variable changes manifest in the data and establish how to measure these effects in real data using diffusion models.
|
data structure, hierarchical compositionality, diffusion models, statistical physics, phase transition
|
A hierarchical structure in the data induces a diverging correlation length at a phase transition in diffusion models, which is observed also in text and images.
| 11,698 |
2410.13770
|
[
-0.012889554724097252,
-0.14089040458202362,
0.0722847580909729,
0.04506111890077591,
0.09292486310005188,
-0.05772867798805237,
-0.05472084879875183,
-0.14189015328884125,
0.06075819954276085,
-0.0013920009369030595,
0.011471179313957691,
0.021880239248275757,
0.016101406887173653,
0.03111223876476288,
0.0032252923119813204,
-0.020480027422308922,
0.05510648339986801,
0.0426701158285141,
-0.09253882616758347,
-0.09510371834039688,
-0.027818063274025917,
0.04025368019938469,
-0.03812899813055992,
-0.0023827115073800087,
0.09451285749673843,
0.09440314024686813,
0.052280254662036896,
-0.02219155803322792,
-0.020649906247854233,
-0.05155166611075401,
0.05448609218001366,
-0.026707766577601433,
-0.01900307647883892,
0.03784983977675438,
-0.016789402812719345,
0.0335468128323555,
0.014246262609958649,
0.028396280482411385,
-0.003554555820301175,
0.0323602668941021,
0.015129104256629944,
0.06991373747587204,
-0.007960984483361244,
0.02535005658864975,
0.05640837922692299,
-0.047834910452365875,
0.01872202940285206,
-0.08699063211679459,
-0.031132839620113373,
0.0397430956363678,
0.003895755158737302,
-0.02697724476456642,
-0.039558883756399155,
0.06957193464040756,
-0.015996916219592094,
0.06475736200809479,
0.0755160003900528,
0.05315331742167473,
0.026408374309539795,
0.05036510154604912,
0.0024611284025013447,
-0.054891157895326614,
-0.073823943734169,
-0.05652619153261185,
0.03885166719555855,
0.04000559821724892,
-0.005530386231839657,
0.04285397380590439,
-0.023745140060782433,
0.02950248122215271,
-0.01361816469579935,
0.0664554163813591,
-0.09986555576324463,
0.02575043775141239,
-0.018656499683856964,
-0.0026924151461571455,
0.02152428962290287,
-0.031397029757499695,
0.03821325674653053,
-0.08039753139019012,
-0.0005747228278778493,
0.02632232755422592,
0.03541772812604904,
0.029709307476878166,
0.08918096870183945,
-0.05802486836910248,
0.06115315854549408,
0.05699896812438965,
-0.07858661562204361,
-0.030453085899353027,
-0.033862028270959854,
0.0367666594684124,
-0.09892859309911728,
0.02730623446404934,
-0.0020772365387529135,
0.04056419059634209,
0.05743591487407684,
-0.08588577806949615,
0.09153959155082703,
0.06946355104446411,
-0.02180427312850952,
0.016722870990633965,
0.004965600557625294,
0.08814414590597153,
0.002244828036054969,
-0.07251904159784317,
0.09923118352890015,
0.017624109983444214,
0.0010433554416522384,
-0.02621297538280487,
-0.03020070120692253,
0.019741078838706017,
0.018612278625369072,
0.019553104415535927,
0.062345460057258606,
-0.13574524223804474,
0.08840616792440414,
-0.009422117844223976,
-0.025988534092903137,
0.012713385745882988,
0.031117934733629227,
-0.008563268929719925,
-0.044936180114746094,
-0.024582596495747566,
0.05457460507750511,
-0.01946583017706871,
-0.054237958043813705,
6.784611764900536e-34,
0.02273324504494667,
-0.04075692966580391,
0.04802858456969261,
0.11119057238101959,
0.07508396357297897,
-0.009499926120042801,
-0.049741797149181366,
-0.0727333202958107,
0.02045963890850544,
0.001008681720122695,
-0.0011352295987308025,
0.007919758558273315,
-0.11664930731058121,
0.04553772881627083,
0.013893273659050465,
-0.02329137735068798,
-0.02541610598564148,
0.06165505200624466,
-0.007340842857956886,
0.03886643424630165,
0.006277244072407484,
0.00022060888295527548,
-0.0514775775372982,
-0.06545877456665039,
0.007844829931855202,
-0.036319099366664886,
0.03740817680954933,
-0.07504276931285858,
-0.021322285756468773,
0.011840510182082653,
-0.12623751163482666,
0.04190438985824585,
-0.03736056014895439,
0.04890793561935425,
-0.05552788823843002,
-0.00887971930205822,
0.009581441059708595,
0.00367716490291059,
0.005810460541397333,
-0.06267160177230835,
0.03830290585756302,
0.013641109690070152,
0.020159471780061722,
-0.022379623726010323,
-0.11795614659786224,
0.030978897586464882,
0.029150765389204025,
-0.029588280245661736,
0.017710672691464424,
-0.03291119262576103,
0.00986425206065178,
-0.03266888111829758,
0.002450361615046859,
-0.0893484428524971,
-0.06487422436475754,
0.04968981444835663,
0.06651505827903748,
-0.0653754398226738,
-0.0028411780949681997,
-0.02825036644935608,
0.047026243060827255,
0.020414941012859344,
-0.013184321112930775,
0.06718838959932327,
0.029790354892611504,
0.029110094532370567,
-0.044221729040145874,
0.04396392032504082,
0.021015780046582222,
-0.00029976433143019676,
-0.010517475195229053,
0.07956748455762863,
-0.07123881578445435,
-0.06686655431985855,
0.036316338926553726,
0.002512197010219097,
0.006510495208203793,
-0.01005291473120451,
0.013600023463368416,
-0.016798151656985283,
-0.05690068379044533,
-0.11388170719146729,
0.03304973989725113,
-0.109965980052948,
-0.07867849618196487,
0.022053517401218414,
-0.010074730962514877,
-0.13072983920574188,
-0.09433088451623917,
-0.025416644290089607,
-0.09435347467660904,
-0.018065698444843292,
0.06173497065901756,
0.03516353294253349,
0.03782140091061592,
-2.1345829467230766e-33,
-0.05182287096977234,
0.0010690471390262246,
0.028290459886193275,
0.04461224004626274,
0.001866262755356729,
0.014673952013254166,
-0.019466478377580643,
0.13017451763153076,
0.08916425704956055,
-0.037901196628808975,
0.007761674001812935,
-0.025719106197357178,
-0.014569418504834175,
-0.05433201417326927,
0.015775199979543686,
-0.00565444165840745,
0.08890841156244278,
0.002821306698024273,
0.027803797274827957,
-0.04478908330202103,
-0.027139879763126373,
0.028602944687008858,
0.004191697109490633,
0.035801950842142105,
-0.012295231223106384,
0.002433360554277897,
-0.025461401790380478,
0.05527554824948311,
-0.0171700157225132,
-0.04026596620678902,
-0.06659938395023346,
-0.02083541266620159,
0.08503738790750504,
0.016083037480711937,
-0.10479148477315903,
0.09071578085422516,
-0.03839530050754547,
-0.028943657875061035,
-0.04622407257556915,
-0.013472238555550575,
0.07642082124948502,
-0.014505578204989433,
-0.07410086691379547,
0.00016314579988829792,
0.00697967316955328,
-0.10519044101238251,
0.0008729472756385803,
0.024116884917020798,
0.027334654703736305,
0.07658170163631439,
0.03647135570645332,
0.07422245293855667,
-0.04987252503633499,
0.036744046956300735,
-0.03717153146862984,
0.004003416281193495,
-0.044129278510808945,
-0.03961889073252678,
0.08916415274143219,
0.0442751906812191,
-0.07111289352178574,
0.004547410644590855,
-0.08797973394393921,
-0.011895941570401192,
-0.0648709088563919,
-0.09639187157154083,
-0.08212064951658249,
-0.0393998883664608,
0.07825802266597748,
0.037168003618717194,
0.06893260031938553,
-0.004626317881047726,
-0.02103731781244278,
0.010386896319687366,
0.0462447889149189,
-0.11171311885118484,
-0.0063599105924367905,
0.013452844694256783,
-0.015431021340191364,
-0.007400322239845991,
0.07848700135946274,
0.0351637601852417,
0.015946874395012856,
0.009671325795352459,
0.047925084829330444,
0.07484737783670425,
0.007489986252039671,
-0.037409666925668716,
0.03911743313074112,
0.021161340177059174,
0.004251900129020214,
-0.07237415760755539,
-0.02665105275809765,
0.0022676510270684958,
-0.09986602514982224,
-4.034223621829369e-8,
-0.04855213314294815,
0.02700880356132984,
0.03094450756907463,
0.001197130768559873,
-0.007439922541379929,
-0.032900162041187286,
0.10333317518234253,
0.13343875110149384,
0.014544667676091194,
-0.009086181409657001,
0.1361009180545807,
-0.03175028786063194,
-0.03339604288339615,
-0.05534091219305992,
-0.017455967143177986,
0.06795564293861389,
0.053958628326654434,
-0.012062499299645424,
0.005193168763071299,
-0.002713262801989913,
-0.014177393168210983,
0.05296158418059349,
0.030610566958785057,
0.027392027899622917,
0.05707636848092079,
0.01014777086675167,
-0.041110314428806305,
0.08309560269117355,
0.05500642955303192,
-0.020341094583272934,
0.036362599581480026,
0.012562280520796776,
0.004173559136688709,
0.04045291244983673,
-0.04973940551280975,
0.07502344995737076,
0.0721743106842041,
-0.015595050528645515,
-0.05543062463402748,
-0.06936482340097427,
-0.07928620278835297,
-0.005682996939867735,
-0.0330367311835289,
-0.005693379323929548,
-0.016106106340885162,
0.05449112132191658,
0.015862859785556793,
-0.05185624584555626,
0.050795190036296844,
-0.01614217832684517,
0.05763518065214157,
-0.01413669716566801,
0.016155218705534935,
0.0019131206208840013,
0.08871569484472275,
-0.027246084064245224,
0.006106684450060129,
0.052109602838754654,
0.029409315437078476,
0.04174499586224556,
0.07208509743213654,
-0.005666350480169058,
-0.012396947480738163,
0.028235357254743576
] | 0 | 0 | 0 | 0 |
|
Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustness
|
https://openreview.net/forum?id=daUQ7vmGap
|
[
"Boqian Wu",
"Qiao Xiao",
"Shunxin Wang",
"Nicola Strisciuglio",
"Mykola Pechenizkiy",
"Maurice van Keulen",
"Decebal Constantin Mocanu",
"Elena Mocanu"
] |
Poster
|
It is generally perceived that Dynamic Sparse Training opens the door to a new era of scalability and efficiency for artificial neural networks at, perhaps, some costs in accuracy performance for the classification task. At the same time, Dense Training is widely accepted as being the "de facto" approach to train artificial neural networks if one would like to maximize their robustness against image corruption. In this paper, we question this general practice. Consequently, \textit{we claim that}, contrary to what is commonly thought, the Dynamic Sparse Training methods can consistently outperform Dense Training in terms of robustness accuracy, particularly if the efficiency aspect is not considered as a main objective (i.e., sparsity levels between 10\% and up to 50\%), without adding (or even reducing) resource cost. We validate our claim on two types of data, images and videos, using several traditional and modern deep learning architectures for computer vision and three widely studied Dynamic Sparse Training algorithms. Our findings reveal a new yet-unknown benefit of Dynamic Sparse Training and open new possibilities in improving deep learning robustness beyond the current state of the art.
|
Dynamic Sparse Training, Image Corruption Robustness
| null | 11,692 |
2410.03030
|
[
-0.07594019919633865,
-0.06955000758171082,
-0.004069217015057802,
0.07736477255821228,
0.01890294998884201,
-0.00927697867155075,
-0.0007005665102042258,
-0.048537030816078186,
-0.024579649791121483,
-0.11915452033281326,
-0.05660521611571312,
0.005804777145385742,
0.038588035851716995,
0.06242445856332779,
-0.1245073676109314,
-0.03369584679603577,
0.12525123357772827,
0.11577628552913666,
-0.09885555505752563,
0.0055676898919045925,
-0.014791547320783138,
-0.0660782903432846,
0.04717530310153961,
0.016391422599554062,
0.06036333739757538,
-0.02110459841787815,
-0.0043126000091433525,
-0.05124257132411003,
0.02327030897140503,
-0.038993269205093384,
0.009495274163782597,
-0.07746223360300064,
-0.020251452922821045,
-0.04757661372423172,
0.023364929482340813,
0.004054088145494461,
0.04480785131454468,
0.07183848321437836,
0.0032980162650346756,
0.014680830761790276,
0.010796885937452316,
-0.008081918582320213,
-0.009963041171431541,
-0.0016680354019626975,
0.029196949675679207,
0.018894698470830917,
0.06221282482147217,
-0.04494750499725342,
0.021238718181848526,
0.013406735844910145,
-0.057809390127658844,
0.018123691901564598,
-0.007401514798402786,
0.047932010143995285,
-0.011664343066513538,
-0.017499687150120735,
0.015866518020629883,
0.0380544476211071,
-0.03618738055229187,
0.051166269928216934,
0.04874052107334137,
-0.04553721472620964,
0.029031069949269295,
0.044808123260736465,
0.01393111702054739,
0.04145985469222069,
0.055393118411302567,
0.009588527493178844,
0.030795833095908165,
0.018268916755914688,
0.016896504908800125,
0.13397216796875,
-0.05731586366891861,
-0.0018505832413211465,
0.013135802932083607,
0.042276740074157715,
0.03118443675339222,
-0.01539719756692648,
0.0715101808309555,
-0.00902482494711876,
0.07349242269992828,
-0.0264505073428154,
-0.04294795170426369,
-0.06816694140434265,
0.08741086721420288,
0.01913178339600563,
-0.03127047419548035,
0.045784156769514084,
0.006416027434170246,
-0.04534325376152992,
-0.03772106394171715,
0.02291998825967312,
-0.047320131212472916,
-0.03278962895274162,
0.029493412002921104,
-0.010364431887865067,
0.06489002704620361,
-0.08548329770565033,
-0.04424135386943817,
0.09007097780704498,
-0.04934796690940857,
-0.03881135582923889,
0.01163544598966837,
0.0496692955493927,
0.04849616810679436,
-0.020201077684760094,
0.09903506934642792,
0.050765324383974075,
0.08655399084091187,
-0.04239026457071304,
0.09315484017133713,
-0.0025229784660041332,
-0.03138048201799393,
0.002233328763395548,
0.031784240156412125,
-0.016546381637454033,
-0.016526253893971443,
-0.03240984305739403,
0.002137452829629183,
0.07780909538269043,
-0.0013762672897428274,
-0.028590546920895576,
-0.027624106034636497,
-0.0002527081232983619,
0.05599691718816757,
-0.06702059507369995,
-0.06833365559577942,
1.9329841961891158e-33,
0.05425248667597771,
0.06778883188962936,
0.0993596687912941,
-0.02485322579741478,
-0.005996590945869684,
-0.09739460051059723,
0.015298441052436829,
0.05888285115361214,
-0.010291104204952717,
-0.0066810790449380875,
-0.07276833802461624,
-0.026700137183070183,
-0.03036808781325817,
0.12110565602779388,
0.06167933717370033,
-0.03972401097416878,
-0.004728836473077536,
0.03350706398487091,
0.04013402387499809,
-0.048205431550741196,
0.046764228492975235,
-0.03294241800904274,
0.016242481768131256,
-0.04577207565307617,
0.004084155894815922,
-0.053742341697216034,
0.12882490456104279,
0.0010588931618258357,
0.028405291959643364,
0.019511863589286804,
-0.11943847686052322,
0.014188052155077457,
0.027211766690015793,
0.05106060951948166,
0.04036802425980568,
-0.02700001932680607,
-0.09678569436073303,
0.001788858207874,
0.023486077785491943,
0.007704986725002527,
-0.021389592438936234,
0.015056164935231209,
0.03720906376838684,
-0.040229298174381256,
-0.030069500207901,
-0.023063356056809425,
0.09888587146997452,
-0.008885957300662994,
-0.14737048745155334,
-0.04024423286318779,
0.030466675758361816,
0.0192872267216444,
-0.07458854466676712,
-0.03552741929888725,
-0.05033122003078461,
-0.03659994900226593,
0.02903224341571331,
-0.0025152387097477913,
0.010968598537147045,
0.01614885963499546,
-0.014364794827997684,
0.014285746961832047,
-0.04416873678565025,
0.07959631830453873,
-0.05129273235797882,
0.014245477505028248,
0.06126602739095688,
0.02876989170908928,
-0.023692592978477478,
-0.04311777278780937,
0.02598917856812477,
0.0000039155975173343904,
-0.0285864919424057,
-0.10098865628242493,
0.05772662162780762,
0.02303820662200451,
0.12615306675434113,
-0.05643098056316376,
-0.053818993270397186,
0.03452148288488388,
-0.041987448930740356,
-0.014340060763061047,
0.036824263632297516,
-0.0327129140496254,
-0.08773543685674667,
0.11895667016506195,
0.07061600685119629,
-0.024741576984524727,
0.00787875335663557,
0.006072448566555977,
-0.038732390850782394,
-0.040099095553159714,
-0.040737103670835495,
-0.08204861730337143,
-0.017689518630504608,
-1.5122458398485293e-33,
-0.006876619067043066,
0.05531224235892296,
-0.06784217059612274,
0.07204020023345947,
-0.023295555263757706,
0.020988551899790764,
-0.06724949926137924,
0.07152699679136276,
0.019296426326036453,
-0.11323380470275879,
-0.06167645379900932,
0.017984088510274887,
-0.011029673740267754,
-0.03620145469903946,
-0.028706151992082596,
-0.019785314798355103,
-0.11525392532348633,
0.014670216478407383,
-0.04627849534153938,
0.04060167074203491,
0.06572093814611435,
0.04407845810055733,
-0.029983874410390854,
-0.012757516466081142,
-0.051493894308805466,
0.05582859739661217,
-0.0940377488732338,
0.09681752324104309,
0.016376761719584465,
0.008424273692071438,
-0.01896645501255989,
-0.014105869457125664,
-0.06480071693658829,
-0.06490164995193481,
-0.029009874910116196,
0.06521623581647873,
0.04307803139090538,
-0.025362594053149223,
-0.02120860107243061,
0.04887164384126663,
0.013074318878352642,
0.02710188552737236,
-0.06291844695806503,
0.003523745806887746,
0.006834276486188173,
-0.04663515090942383,
-0.009663118049502373,
-0.016908777877688408,
0.010475361719727516,
0.079034723341465,
0.021336715668439865,
-0.01480952836573124,
-0.1161889061331749,
0.09075096249580383,
-0.03048579767346382,
-0.029310408979654312,
-0.025832613930106163,
0.0445760041475296,
0.07650427520275116,
0.0769587978720665,
-0.06513669341802597,
-0.026136215776205063,
-0.12546029686927795,
-0.0002505260636098683,
-0.01284338440746069,
-0.029826438054442406,
-0.060537293553352356,
-0.0013065218226984143,
0.01425534300506115,
0.08734966814517975,
0.02253223955631256,
-0.0439051054418087,
-0.0030461952555924654,
0.03232342004776001,
-0.04972866550087929,
0.05033068358898163,
0.02120174467563629,
-0.06739559769630432,
0.0004142453835811466,
0.0681973546743393,
0.012036164291203022,
-0.022874390706419945,
0.0191472377628088,
0.02931894175708294,
0.062693290412426,
0.138640359044075,
0.01614329032599926,
-0.032351840287446976,
0.02047767862677574,
0.05721883848309517,
0.0030215876176953316,
0.017707057297229767,
-0.003231094451621175,
-0.02141730859875679,
-0.05722260847687721,
-4.3600049792758e-8,
-0.041832078248262405,
0.09561234712600708,
0.03596937283873558,
-0.019702991470694542,
0.03909856453537941,
-0.08460060507059097,
0.08810514211654663,
0.09076526761054993,
0.020800577476620674,
-0.10160355269908905,
0.07869473099708557,
-0.08117576688528061,
-0.03899575024843216,
-0.06847275048494339,
-0.05149274319410324,
0.06236738711595535,
-0.030225222930312157,
0.006385928485542536,
-0.029189815744757652,
0.03382799029350281,
-0.022886760532855988,
0.006221065763384104,
0.016907405108213425,
-0.010940184816718102,
0.02222405932843685,
-0.03362271562218666,
-0.020428206771612167,
0.06334760040044785,
0.06135081499814987,
0.025643883273005486,
-0.0124850207939744,
0.010395321995019913,
0.08210574835538864,
-0.04781810939311981,
-0.010385904461145401,
0.10854067653417587,
-0.0006376864621415734,
-0.006337994243949652,
-0.07597044110298157,
-0.04596845805644989,
-0.01457543671131134,
0.022498799487948418,
0.035484679043293,
-0.027295084670186043,
-0.03305047005414963,
-0.013996382243931293,
-0.020703069865703583,
-0.04311788082122803,
-0.024277618154883385,
-0.01067392248660326,
-0.01533904206007719,
0.03198067471385002,
-0.0209788978099823,
0.08120381832122803,
0.08404548466205597,
-0.06433780491352081,
0.04294443503022194,
-0.059787798672914505,
-0.01826559007167816,
0.1249515488743782,
0.008173760026693344,
-0.01030941866338253,
-0.03448990359902382,
0.006722385995090008
] | 0 | 0 | 0 | 0 |
|
Unlocking Point Processes through Point Set Diffusion
|
https://openreview.net/forum?id=4anfpHj0wf
|
[
"David Lüdke",
"Enric Rabasseda Raventós",
"Marcel Kollovieh",
"Stephan Günnemann"
] |
Poster
|
Point processes model the distribution of random point sets in mathematical spaces, such as spatial and temporal domains, with applications in fields like seismology, neuroscience, and economics.
Existing statistical and machine learning models for point processes are predominantly constrained by their reliance on the characteristic intensity function, introducing an inherent trade-off between efficiency and flexibility.
In this paper, we introduce Point Set Diffusion, a diffusion-based latent variable model that can represent arbitrary point processes on general metric spaces without relying on the intensity function.
By directly learning to stochastically interpolate between noise and data point sets, our approach effectively captures the distribution of point processes and enables efficient, parallel sampling and flexible generation for complex conditional tasks.
Experiments on synthetic and real-world datasets demonstrate that Point Set Diffusion achieves state-of-the-art performance in unconditional and conditional generation of spatial and spatiotemporal point processes while providing up to orders of magnitude faster sampling.
|
Generative Model, Diffusion Model, Set Model, Point Sets, Forecasting, Density Estimation, Spatial, Temporal, Probabilistic Models
| null | 11,691 |
2410.22493
|
[
-0.03739804029464722,
-0.1149606853723526,
0.051050666719675064,
0.008939164690673351,
0.0325101800262928,
-0.00017316963931079954,
0.026279233396053314,
-0.08971146494150162,
0.12229560315608978,
-0.008424527011811733,
-0.02747945673763752,
-0.024410495534539223,
-0.003921035211533308,
0.08574128150939941,
0.061980511993169785,
-0.02996992878615856,
0.03168214112520218,
-0.02264246717095375,
-0.05544352903962135,
0.004249881953001022,
0.00034487704397179186,
0.05402917042374611,
-0.02170719765126705,
0.01626407541334629,
0.05432037636637688,
0.02239983342587948,
0.1561235934495926,
-0.013768947683274746,
-0.005103624425828457,
-0.0013051419518887997,
0.06866425275802612,
-0.05796800181269646,
-0.030952010303735733,
-0.05021670088171959,
0.0193705502897501,
0.004504202865064144,
-0.0642920434474945,
0.03260283172130585,
-0.040183939039707184,
0.03802227973937988,
0.026364142075181007,
-0.018641406670212746,
0.030396562069654465,
0.08983343094587326,
0.08007586002349854,
-0.003308719489723444,
0.03679442033171654,
-0.1024380698800087,
0.013063047081232071,
0.008003605529665947,
0.034786056727170944,
0.007398327346891165,
-0.030786169692873955,
0.033952537924051285,
-0.0018220223719254136,
0.048677701503038406,
0.026675928384065628,
-0.0462765209376812,
0.0659412071108818,
-0.05467401072382927,
0.013014325872063637,
-0.07904385775327682,
0.015033280476927757,
-0.05096770077943802,
-0.05501476675271988,
-0.05266503617167473,
0.02361219935119152,
0.09153739362955093,
0.025681234896183014,
-0.010702707804739475,
-0.07789016515016556,
0.08798988163471222,
-0.05663343146443367,
-0.004448162857443094,
-0.027224192395806313,
0.032972417771816254,
0.008815648965537548,
-0.0749891847372055,
-0.03405491262674332,
-0.07157377153635025,
-0.01761810854077339,
0.011592356488108635,
0.06542634963989258,
0.014997506514191628,
0.07479193806648254,
0.008901302702724934,
0.04288066551089287,
0.12151145190000534,
-0.005842824932187796,
-0.023147081956267357,
-0.043474260717630386,
0.049388136714696884,
-0.08397751301527023,
-0.06074485182762146,
-0.012627189047634602,
0.0022614742629230022,
0.01167725957930088,
-0.10783509165048599,
0.11738919466733932,
0.07378633320331573,
-0.009796777740120888,
0.06308922916650772,
0.032572560012340546,
0.02020331844687462,
0.03850976750254631,
-0.11723577231168747,
0.021926483139395714,
0.04392252489924431,
-0.010763265192508698,
-0.03703600913286209,
0.006962160114198923,
-0.03557206317782402,
0.04186787083745003,
0.06131550297141075,
0.05957997590303421,
-0.03192281350493431,
0.008903049863874912,
-0.03797087073326111,
0.017797624692320824,
-0.007984809577465057,
-0.06283599883317947,
-0.034912895411252975,
-0.07451121509075165,
0.03617957606911659,
-0.004712880123406649,
-0.013737468980252743,
0.02010762132704258,
7.68603547523128e-34,
0.004072363488376141,
-0.04750122129917145,
0.06119156628847122,
-0.027240192517638206,
0.04304042458534241,
-0.009670359082520008,
-0.05891787260770798,
-0.05927463620901108,
0.06248946115374565,
-0.003563809208571911,
-0.032452542334795,
-0.06542962044477463,
-0.059674493968486786,
0.044716913253068924,
0.0706658586859703,
-0.025025414302945137,
0.019250640645623207,
0.057068075984716415,
0.020075514912605286,
0.03449341654777527,
0.05836738273501396,
-0.04451032727956772,
-0.06338347494602203,
-0.0006502897012978792,
-0.007009199820458889,
0.03631485626101494,
-0.03272661566734314,
-0.022344142198562622,
0.004812418017536402,
-0.007671964354813099,
-0.07415886223316193,
0.07078887522220612,
-0.0397489033639431,
0.030554573982954025,
-0.006788862403482199,
0.018861250951886177,
-0.0024066860787570477,
-0.008829602971673012,
0.03255465254187584,
-0.1051541417837143,
0.05581822991371155,
0.03797302395105362,
0.01821845956146717,
-0.042522870004177094,
-0.0825100690126419,
0.02566494606435299,
0.04340890049934387,
-0.007248491048812866,
0.04467117413878441,
-0.06527578830718994,
0.07839304953813553,
0.013537406921386719,
0.02411232516169548,
-0.10600470006465912,
-0.04363285377621651,
0.05216973274946213,
-0.01253507286310196,
-0.11583089083433151,
0.02574404887855053,
0.03951619565486908,
0.04644423723220825,
0.03320346400141716,
-0.019990289583802223,
-0.0005829390720464289,
0.03621368855237961,
-0.031746748834848404,
0.015957627445459366,
0.06386551260948181,
0.09332979470491409,
0.052075374871492386,
0.024027468636631966,
0.059096816927194595,
-0.07476933300495148,
-0.14929910004138947,
0.01783553510904312,
-0.0024037512484937906,
0.010368862189352512,
-0.02440468966960907,
-0.05802127346396446,
0.04732285812497139,
-0.07815877348184586,
-0.036144744604825974,
-0.018824031576514244,
-0.11228415369987488,
0.05495109781622887,
-0.018003376200795174,
0.011217798106372356,
-0.029138047248125076,
-0.10800797492265701,
-0.10956340283155441,
-0.05952748283743858,
0.021461986005306244,
-0.028223412111401558,
0.02495676465332508,
-0.07986056804656982,
-2.2154158157573712e-33,
-0.08201410621404648,
-0.0204022079706192,
-0.007634319830685854,
0.07584786415100098,
0.018026502802968025,
0.024287475273013115,
-0.07275035232305527,
-0.05752165988087654,
0.09473744034767151,
-0.0024364087730646133,
-0.12001460045576096,
-0.019630899652838707,
0.011778546497225761,
-0.03823911398649216,
-0.005489554721862078,
0.033758774399757385,
0.030860573053359985,
-0.005521298851817846,
-0.00627265265211463,
-0.01707248017191887,
-0.04574701935052872,
0.06150506064295769,
0.010056385770440102,
-0.027084417641162872,
-0.038950446993112564,
0.0058899372816085815,
-0.02773824706673622,
0.013551904819905758,
-0.06659368425607681,
0.02179720625281334,
-0.10131613910198212,
0.01419111993163824,
0.029754839837551117,
-0.05839484557509422,
-0.08988679200410843,
0.07766932249069214,
0.04616831615567207,
0.00190032459795475,
-0.07317949086427689,
0.02397468499839306,
0.10231699049472809,
-0.04989911615848541,
-0.03391370549798012,
0.08126726746559143,
-0.0010827782098203897,
-0.011832836084067822,
0.020435933023691177,
0.01599838212132454,
-0.04307014122605324,
0.03246583044528961,
-0.02401888743042946,
0.09032777696847916,
-0.07693091034889221,
0.03766203671693802,
-0.009599047712981701,
-0.07108417898416519,
-0.03440599516034126,
-0.12807241082191467,
0.010553764179348946,
-0.023721175268292427,
-0.04503059387207031,
-0.05816633999347687,
-0.03133108839392662,
0.10918764024972916,
0.012854965403676033,
-0.004190132021903992,
-0.03393411636352539,
-0.03819623216986656,
0.05249590799212456,
0.030325353145599365,
0.020895222201943398,
0.02296672761440277,
-0.002265208400785923,
0.007425069808959961,
-0.008598499931395054,
-0.048333119601011276,
0.08893124014139175,
-0.034259628504514694,
-0.014710905030369759,
0.011051093228161335,
0.042673591524362564,
0.015524915419518948,
0.018950529396533966,
0.04128196835517883,
-0.005888319108635187,
0.10888230055570602,
0.01343001239001751,
0.019583523273468018,
0.03281279280781746,
-0.014643502421677113,
-0.022557677701115608,
0.012484442442655563,
0.009352847933769226,
0.034126877784729004,
-0.10454192012548447,
-3.833953954313074e-8,
-0.10996586084365845,
0.06299185752868652,
0.10288257896900177,
-0.016925564035773277,
0.030033117160201073,
0.00668366439640522,
0.06880765408277512,
0.06586961448192596,
0.04298750311136246,
-0.0442371740937233,
-0.009251109324395657,
-0.08965209871530533,
0.011914478614926338,
-0.007692284416407347,
0.023993391543626785,
0.04261496290564537,
0.06580671668052673,
0.055563878268003464,
0.014649623073637486,
0.0479789674282074,
-0.05866239592432976,
0.08214942365884781,
-0.018728191033005714,
0.0069786855019629,
0.0887618437409401,
-0.007982106879353523,
-0.004828170407563448,
0.013673617504537106,
0.11127588152885437,
-0.030813192948698997,
0.04892619699239731,
-0.02527167834341526,
0.01575867086648941,
0.03888414055109024,
0.013026442378759384,
0.017125675454735756,
0.034101735800504684,
0.023801691830158234,
-0.048128001391887665,
-0.03506159037351608,
-0.09562548995018005,
0.03531860187649727,
-0.02260562591254711,
-0.010013433173298836,
-0.015114966779947281,
0.09414594620466232,
0.04207947105169296,
-0.03221244364976883,
0.0687815397977829,
-0.008655627258121967,
-0.06166529282927513,
0.023168284446001053,
0.03536650165915489,
-0.026655133813619614,
0.055977266281843185,
0.07629261910915375,
0.007741397246718407,
-0.029161101207137108,
0.06313963234424591,
-0.012968658469617367,
-0.02087569423019886,
-0.01803920418024063,
-0.12133914977312088,
0.03475422039628029
] | 0 | 0 | 0 | 0 |
|
Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity
|
https://openreview.net/forum?id=vVxeFSR4fU
|
[
"Jiachen Jiang",
"Jinxin Zhou",
"Zhihui Zhu"
] |
Poster
|
Analyzing the similarity of internal representations within and across different models has been an important technique for understanding the behavior of deep neural networks. Most existing methods for analyzing the similarity between representations of high dimensions, such as those based on Centered Kernel Alignment (CKA), rely on statistical properties of the representations for a set of data points. In this paper, we focus on transformer models and study the similarity of representations between the hidden layers of individual transformers. In this context, we show that a simple sample-wise cosine similarity metric is capable of capturing the similarity and aligns with the complicated CKA. Our experimental results on common transformers reveal that representations across layers are positively correlated, with similarity increasing when layers get closer. We provide a theoretical justification for this phenomenon under the geodesic curve assumption for the learned transformer, a property that may approximately hold for residual networks. We then show that an increase in representation similarity implies an increase in predicted probability when directly applying the last-layer classifier to any hidden layer representation. This offers a justification for {\it saturation events}, where the model's top prediction remains unchanged across subsequent layers, indicating that the shallow layer has already learned the necessary knowledge. We then propose an aligned training method to improve the effectiveness of shallow layer by enhancing the similarity between internal representations, with trained models that enjoy the following properties: (1) more early saturation events, (2) layer-wise accuracies monotonically increase and reveal the minimal depth needed for the given task, (3) when served as multi-exit models, they achieve on-par performance with standard multi-exit architectures which consist of additional classifiers designed for early exiting in shallow layers. To our knowledge, our work is the first to show that one common classifier is sufficient for multi-exit models. We conduct experiments on both vision and NLP tasks to demonstrate the performance of the proposed aligned training.
|
Representation Similarity, Saturation Event, Early Exit
|
Study how representations propagate across layers in transformers using sample-wise, layer-wise representation similarity; propose aligned training to promote early saturation events design multi-exit models with a single classifier
| 11,690 |
2406.14479
|
[
-0.13600417971611023,
-0.13754506409168243,
0.025506125763058662,
-0.002325051696971059,
0.07334399968385696,
0.03864015266299248,
-0.04650532454252243,
-0.00039539311546832323,
0.09627921134233475,
-0.10926565527915955,
-0.04500238969922066,
-0.015950080007314682,
0.0018621535273268819,
0.055895935744047165,
-0.07724760472774506,
-0.09152685850858688,
0.032546091824769974,
0.0982893854379654,
-0.1402405947446823,
-0.08283229917287827,
0.03378579765558243,
0.011799163185060024,
-0.03055524080991745,
0.0627114474773407,
0.036981817334890366,
0.02579236403107643,
0.10652974247932434,
-0.029301991686224937,
-0.040233809500932693,
-0.01748577691614628,
0.013212691061198711,
-0.03325071558356285,
-0.10435836017131805,
0.021511010825634003,
-0.05849102512001991,
0.06454700231552124,
-0.06477677822113037,
-0.004201166331768036,
0.019295435398817062,
-0.08279202878475189,
0.03978972136974335,
-0.0015021610306575894,
0.0050238254480063915,
0.07159276306629181,
0.033211857080459595,
0.020680254325270653,
-0.022190483286976814,
-0.06465423852205276,
0.012991157360374928,
0.014220953918993473,
0.01404087245464325,
-0.026625612750649452,
-0.032657451927661896,
0.10106587409973145,
0.012708987109363079,
0.0336747020483017,
0.04502570629119873,
-0.009765424765646458,
0.019520020112395287,
0.02082553505897522,
-0.02525443769991398,
-0.07815363258123398,
0.028408139944076538,
-0.011448714882135391,
-0.005034172907471657,
-0.011676594614982605,
0.09563141316175461,
0.03724711388349533,
-0.042180173099040985,
-0.014068080112338066,
-0.06737075746059418,
0.1134965717792511,
-0.09888157993555069,
0.013758612796664238,
0.08324262499809265,
0.06859423965215683,
0.057357754558324814,
0.010741745121777058,
0.0026694645639508963,
-0.09968757629394531,
0.07066784799098969,
0.10349032282829285,
0.08465664833784103,
0.056471940129995346,
0.1016150489449501,
-0.0124189592897892,
-0.01251171063631773,
-0.0005871648318134248,
0.022030359134078026,
-0.06087590381503105,
-0.05786506459116936,
-0.03300120681524277,
-0.031944651156663895,
-0.030457645654678345,
0.024911588057875633,
0.011250076815485954,
-0.011835087090730667,
0.01948690414428711,
-0.038586702197790146,
0.03039102628827095,
-0.039843086153268814,
-0.02013298124074936,
0.07036101073026657,
0.016023239120841026,
0.0785006508231163,
-0.027267038822174072,
0.031249769032001495,
0.0560685396194458,
0.023860281333327293,
-0.0984727144241333,
-0.0015556365251541138,
-0.02458178997039795,
-0.03128259629011154,
0.04040240868926048,
0.06275482475757599,
-0.01734589971601963,
0.06853402405977249,
-0.007000650744885206,
-0.1037399023771286,
0.022699138149619102,
-0.04663027822971344,
0.04388297721743584,
-0.05880590155720711,
0.033554889261722565,
-0.0062036835588514805,
0.003703929716721177,
-0.04800983890891075,
2.414972512198763e-33,
0.007350670639425516,
0.08883162587881088,
0.04957903176546097,
0.040096819400787354,
0.005630494561046362,
-0.05841798707842827,
-0.032378531992435455,
0.05535052344202995,
0.08226793259382248,
0.022472387179732323,
-0.12484391033649445,
0.06462608277797699,
-0.08160172402858734,
0.06443390995264053,
-0.028856784105300903,
0.04655350372195244,
-0.05438695475459099,
0.008641011081635952,
-0.03798430413007736,
-0.05008743703365326,
0.008202238008379936,
0.0061408523470163345,
0.013219536282122135,
-0.02157939039170742,
-0.06896879523992538,
-0.005831032991409302,
-0.011630607768893242,
-0.039681509137153625,
-0.09807336330413818,
-0.044476576149463654,
-0.04359401389956474,
-0.00507183326408267,
0.048417434096336365,
-0.05395796522498131,
0.02066008746623993,
-0.03796154633164406,
0.041961170732975006,
-0.04152899235486984,
-0.016401436179876328,
-0.016866473481059074,
0.01115419901907444,
0.028370583429932594,
-0.042245738208293915,
-0.05097360163927078,
-0.0199297945946455,
-0.025896698236465454,
0.017025507986545563,
-0.028188323602080345,
-0.0008970568305812776,
-0.09532275795936584,
0.08908417075872421,
0.013842190615832806,
-0.05529141426086426,
-0.02526215836405754,
0.028403449803590775,
0.020709311589598656,
0.03726549074053764,
-0.014444543980062008,
0.07489170879125595,
0.027552779763936996,
0.011132468469440937,
0.047469351440668106,
-0.006338454782962799,
0.06572901457548141,
-0.006644496228545904,
-0.013909570872783661,
0.019887587055563927,
0.084427610039711,
0.035956304520368576,
0.04387599974870682,
-0.04102190211415291,
0.056199200451374054,
-0.06117599457502365,
-0.08337904512882233,
0.019886305555701256,
-0.011817943304777145,
-0.033902883529663086,
-0.07145839929580688,
-0.023635171353816986,
0.09213365614414215,
-0.06033014878630638,
-0.03154502809047699,
-0.007065769284963608,
-0.0897873044013977,
-0.08024632930755615,
0.03206407651305199,
0.06085840240120888,
-0.09319067746400833,
-0.0287756510078907,
0.008955086581408978,
0.017691107466816902,
-0.01785849593579769,
0.016986873000860214,
0.004058308433741331,
-0.02207869663834572,
-2.3920804944004682e-33,
-0.0399579294025898,
0.06740705668926239,
0.001190269598737359,
0.005025386810302734,
-0.07080035656690598,
-0.010566772893071175,
-0.020719582214951515,
0.058473531156778336,
-0.04156624153256416,
0.0026304482016712427,
0.035408444702625275,
-0.013001714833080769,
0.006809541489928961,
-0.048272840678691864,
0.07393627613782883,
0.04664752632379532,
-0.04520001262426376,
0.02152032032608986,
0.019883902743458748,
-0.037234604358673096,
0.0717022567987442,
0.021204441785812378,
-0.14472927153110504,
0.009631241671741009,
-0.0577324740588665,
0.041468556970357895,
-0.024441728368401527,
0.07550261914730072,
0.06373833864927292,
0.0049330126494169235,
-0.01702161692082882,
-0.06431186199188232,
0.00025480848853476346,
-0.00076221366180107,
-0.07690516859292984,
0.03271111845970154,
0.04274624213576317,
-0.041150715202093124,
0.023668639361858368,
0.008218257687985897,
-0.039535749703645706,
-0.046918418258428574,
-0.04032249376177788,
-0.015984777361154556,
-0.055881861597299576,
-0.05465138703584671,
0.02217414788901806,
-0.012797932140529156,
-0.027067897841334343,
0.00806186068803072,
0.0517917238175869,
0.07348724454641342,
-0.10773550719022751,
0.005244024097919464,
-0.02984629012644291,
0.023883702233433723,
-0.06103490665555,
0.019407257437705994,
0.1011359691619873,
-0.015710292384028435,
0.00990986730903387,
-0.05143146589398384,
0.007324286736547947,
-0.03396691381931305,
-0.024113638326525688,
-0.03601814806461334,
-0.00738002173602581,
0.008107044734060764,
0.06536895781755447,
0.12353826314210892,
0.03197505697607994,
0.008187984116375446,
-0.01968558318912983,
-0.06644653528928757,
0.003385938238352537,
-0.07192069292068481,
0.05323861911892891,
-0.00034273762139491737,
0.0019794211257249117,
-0.039361607283353806,
-0.058067239820957184,
-0.02213977463543415,
0.016386451199650764,
0.007302830461412668,
0.11105749756097794,
0.10857781022787094,
0.05602141469717026,
0.010722787119448185,
0.05823042616248131,
0.050440043210983276,
-0.027861367911100388,
0.019994981586933136,
-0.027805175632238388,
0.036120764911174774,
-0.06009585037827492,
-5.002446812341077e-8,
-0.0643865093588829,
0.08038906753063202,
0.00046750030014663935,
-0.027710679918527603,
-0.002526820870116353,
-0.050252560526132584,
0.06436800956726074,
0.09879879653453827,
-0.07442891597747803,
0.04644671082496643,
0.060377467423677444,
-0.04461623355746269,
-0.008801736868917942,
0.007840384729206562,
0.004621261265128851,
0.0620255321264267,
-0.019228683784604073,
0.058313481509685516,
0.03564305976033211,
0.07175958901643753,
-0.017920393496751785,
0.0642181932926178,
-0.03373695909976959,
-0.012210526503622532,
0.040238697081804276,
-0.032787930220365524,
-0.009419133886694908,
0.1068093553185463,
0.03983515873551369,
0.05856144428253174,
-0.054161831736564636,
-0.0031279604882001877,
0.031143296509981155,
-0.06370753049850464,
0.014236769638955593,
0.08855485916137695,
0.08409511297941208,
0.006321065593510866,
-0.03179321438074112,
0.04338099807500839,
-0.0035976124927401543,
0.07266287505626678,
-0.023499472066760063,
0.04458501935005188,
0.07866185158491135,
0.030362630262970924,
0.0874854326248169,
-0.043288882821798325,
-0.01739637739956379,
0.010849869810044765,
0.01693514734506607,
-0.023849710822105408,
-0.02866903506219387,
0.07831589132547379,
0.07475356012582779,
-0.04675859585404396,
-0.01636657863855362,
-0.05053326487541199,
0.015933623537421227,
0.02965421974658966,
-0.032781824469566345,
0.0185967069119215,
-0.07530182600021362,
0.032040588557720184
] | 0 | 0 | 0 | 0 |
|
Language Agents Meet Causality -- Bridging LLMs and Causal World Models
|
https://openreview.net/forum?id=y9A2TpaGsE
|
[
"John Gkountouras",
"Matthias Lindemann",
"Phillip Lippe",
"Efstratios Gavves",
"Ivan Titov"
] |
Poster
|
Large Language Models (LLMs) have recently shown great promise in planning and reasoning applications. These tasks demand robust systems, which arguably require a causal understanding of the environment. While LLMs can acquire and reflect common sense causal knowledge from their pretraining data, this information is often incomplete, incorrect, or inapplicable to a specific environment. In contrast, causal representation learning (CRL) focuses on identifying the underlying causal structure within a given environment. We propose a framework that integrates CRLs with LLMs to enable causally-aware reasoning and planning. This framework learns a causal world model, with causal variables linked to natural language expressions. This mapping provides LLMs with a flexible interface to process and generate descriptions of actions and states in text form. Effectively, the causal world model acts as a simulator that the LLM can query and interact with. We evaluate the framework on causal inference and planning tasks across temporal scales and environmental complexities. Our experiments demonstrate the effectiveness of the approach, with the causally-aware method outperforming LLM-based reasoners, especially for longer planning horizons.
|
Large Language Models, Causality, Causal Representation Learning, Language Agents, Planning
|
Improving LLM planning capabilities using learned causal representations
| 11,689 |
2410.19923
|
[
0.008153758943080902,
-0.034155938774347305,
0.03117884136736393,
0.061096519231796265,
0.07825017720460892,
-0.004686316009610891,
-0.04440698027610779,
-0.008210998959839344,
-0.005194021388888359,
0.013786131516098976,
-0.10128840059041977,
-0.06147986650466919,
0.053170088678598404,
0.09251005202531815,
0.013281655497848988,
-0.0014001805102452636,
0.09493809193372726,
-0.03919973596930504,
-0.06226300448179245,
-0.07345274090766907,
0.011257336474955082,
0.04790348559617996,
0.007457527797669172,
0.06565683335065842,
-0.047264229506254196,
0.00595408258959651,
0.047129277139902115,
0.01902277022600174,
0.024522406980395317,
0.05955908074975014,
0.056644972413778305,
0.02338573895394802,
-0.11161186546087265,
-0.06411316990852356,
0.02191288210451603,
0.10841797292232513,
-0.045297179371118546,
0.017208833247423172,
-0.017600413411855698,
-0.08484850078821182,
-0.022390803322196007,
-0.06443235278129578,
-0.03295387700200081,
0.006590080447494984,
0.06949307769536972,
-0.00035868556005880237,
0.012493309564888477,
-0.004082934465259314,
-0.03677555173635483,
0.036886971443891525,
-0.09371168911457062,
-0.05071411654353142,
0.02383788488805294,
-0.032440971583127975,
0.019979164004325867,
0.044196706265211105,
-0.003802613588050008,
-0.015467122197151184,
-0.047457098960876465,
-0.07683371752500534,
-0.10276830941438675,
-0.09391287714242935,
-0.03244352713227272,
-0.01403016410768032,
-0.07283782958984375,
0.00499087804928422,
-0.06944466382265091,
0.1122613474726677,
-0.037850767374038696,
0.014750903472304344,
-0.025207871571183205,
0.051751263439655304,
-0.0013435494620352983,
0.02163221873342991,
-0.02257182076573372,
0.04242772236466408,
-0.017396003007888794,
-0.00679162098094821,
0.03132414072751999,
-0.052633434534072876,
0.008266459219157696,
0.07063718885183334,
-0.02970016747713089,
0.04814295470714569,
0.14142481982707977,
0.011761240661144257,
0.0695134848356247,
0.09440001100301743,
0.05519987270236015,
0.07254227250814438,
0.036614060401916504,
-0.09087401628494263,
0.1426105499267578,
0.030511142686009407,
0.029137862846255302,
0.10967428982257843,
-0.03009740635752678,
-0.06416597962379456,
0.02642582729458809,
0.09395727515220642,
0.016140848398208618,
0.08148767799139023,
0.013375629670917988,
-0.02201126329600811,
-0.01865476928651333,
0.0070204888470470905,
-0.02483445033431053,
0.05399111658334732,
-0.006558557506650686,
-0.10275018215179443,
-0.03224486485123634,
-0.012405775487422943,
0.08012600243091583,
0.004985596053302288,
0.039387013763189316,
-0.035763319581747055,
0.0366843119263649,
-0.010219463147222996,
0.024078747257590294,
-0.00515650724992156,
-0.010677014477550983,
-0.00936303474009037,
-0.00530800549313426,
0.04692908376455307,
0.02489110641181469,
-0.08155877143144608,
0.0219135582447052,
-2.22217417359063e-33,
0.09625405073165894,
0.04394296556711197,
0.060039177536964417,
0.07998230308294296,
0.04955659806728363,
0.022549571469426155,
-0.008550645783543587,
0.06058219447731972,
-0.03396165743470192,
0.006633202079683542,
0.05130323767662048,
-0.04570261016488075,
-0.061053499579429626,
0.02953196130692959,
0.04358386993408203,
0.005887184292078018,
-0.07410876452922821,
0.05163155123591423,
-0.03143725544214249,
-0.07391786575317383,
-0.013721685856580734,
-0.06687798351049423,
0.008609633892774582,
-0.015403371304273605,
0.055223796516656876,
0.03571108728647232,
0.07353164255619049,
-0.007406014017760754,
0.02581479400396347,
-0.013769262470304966,
-0.013192575424909592,
-0.05195489525794983,
0.019423440098762512,
0.05719488859176636,
-0.011702529154717922,
-0.02956455945968628,
-0.14851635694503784,
-0.05512991547584534,
-0.0009191606077365577,
0.016611581668257713,
0.03448601812124252,
-0.018461529165506363,
0.03254077583551407,
0.02168508805334568,
-0.06427557021379471,
-0.091673843562603,
-0.014262761920690536,
0.011364447884261608,
-0.03021189756691456,
-0.0020774605218321085,
0.08039209246635437,
0.03649529069662094,
-0.01858694851398468,
-0.07715284079313278,
-0.011464468203485012,
0.027900010347366333,
0.015055228024721146,
0.004991272930055857,
-0.03961295634508133,
0.0070428200997412205,
0.020536892116069794,
0.014454816468060017,
0.010596392676234245,
0.11383946985006332,
0.0430225245654583,
0.05139658972620964,
-0.03943265229463577,
0.06260165572166443,
0.05106445774435997,
-0.022483495995402336,
-0.01807699352502823,
0.03965560719370842,
0.028137650340795517,
-0.021106000989675522,
0.03777242451906204,
-0.0036648802924901247,
0.0021506629418581724,
-0.09066283702850342,
0.014408592134714127,
0.04926582798361778,
-0.02144952490925789,
-0.06373555213212967,
-0.00044985985732637346,
0.006380670238286257,
-0.0042060683481395245,
-0.07481300085783005,
-0.03553883358836174,
-0.0877862423658371,
0.05321400985121727,
-0.04804645851254463,
-0.02168338932096958,
-0.03783426806330681,
-0.04183319956064224,
0.035245854407548904,
-0.04350349307060242,
9.739373854390786e-34,
-0.00547028100118041,
-0.053479090332984924,
0.011578025296330452,
-0.03126831352710724,
-0.038620252162218094,
-0.07839815318584442,
-0.054234135895967484,
-0.007993604056537151,
0.046988967806100845,
-0.06665752083063126,
-0.09239596128463745,
-0.004095998127013445,
0.045983295887708664,
-0.0006955752614885569,
0.030096882954239845,
-0.012592029757797718,
0.07389882206916809,
-0.029247110709547997,
0.015456811524927616,
0.03141938894987106,
-0.030470427125692368,
0.07046723365783691,
-0.04983969405293465,
-0.010328836739063263,
-0.014079412445425987,
0.017252091318368912,
0.05322111025452614,
0.05161941051483154,
0.027109859511256218,
0.10791625082492828,
-0.04471028968691826,
-0.0574817880988121,
-0.019973592832684517,
-0.046356648206710815,
-0.10171768814325333,
0.06469496339559555,
0.04050063341856003,
-0.02303394488990307,
-0.09148959070444107,
-0.0034062392078340054,
0.09450516104698181,
-0.037105005234479904,
-0.0405946783721447,
-0.03614533320069313,
-0.08726976811885834,
0.05950411036610603,
-0.10438278317451477,
0.06702087074518204,
0.058557309210300446,
0.00792864803224802,
-0.04374092072248459,
0.020461801439523697,
-0.013092747889459133,
-0.00233327760361135,
-0.03327016904950142,
-0.07504606992006302,
0.00024866496096365154,
-0.08409526199102402,
0.05833066254854202,
-0.029163554310798645,
-0.06105024740099907,
0.02679125778377056,
0.06900420039892197,
-0.013926994986832142,
-0.009446943178772926,
-0.03834524005651474,
-0.0600934661924839,
-0.030258556827902794,
0.010525182820856571,
-0.0465979166328907,
0.08557464927434921,
0.00940955150872469,
-0.0951918438076973,
-0.06802631914615631,
0.06003107130527496,
-0.01773722469806671,
-0.01569691114127636,
0.004745160695165396,
-0.0918574333190918,
-0.063234344124794,
0.10506241023540497,
0.03244655579328537,
0.045946814119815826,
0.06696701049804688,
0.009681430645287037,
0.019122285768389702,
0.017452305182814598,
0.1151370257139206,
0.04815240949392319,
0.03957672789692879,
-0.06526865810155869,
0.03662504255771637,
-0.03179371729493141,
0.08180556446313858,
-0.07459370791912079,
-3.65562762283389e-8,
-0.06802254170179367,
0.022260654717683792,
-0.015085858292877674,
0.03841793164610863,
0.03301031515002251,
-0.062450192868709564,
-0.0282606091350317,
0.027094051241874695,
0.027373403310775757,
0.0001957701751962304,
0.016424471512436867,
-0.009616715833544731,
-0.03209851682186127,
-0.04285089671611786,
-0.04151847958564758,
0.06602278351783752,
0.03788599744439125,
-0.03615134581923485,
-0.015777532011270523,
-0.003516998840495944,
0.07034706324338913,
0.012861764058470726,
-0.04495401307940483,
0.06854359805583954,
0.08288685977458954,
-0.08838449418544769,
-0.08836615830659866,
0.09690336138010025,
0.056338775902986526,
-0.002176528098061681,
0.04098862037062645,
-0.019019758328795433,
0.028925534337759018,
-0.003582427743822336,
0.026871198788285255,
0.045862458646297455,
0.019983436912298203,
-0.06489124894142151,
0.00033091436489485204,
-0.08472715318202972,
0.017492489889264107,
0.10365395992994308,
-0.008499265648424625,
0.029599741101264954,
0.052422672510147095,
-0.03506651893258095,
-0.06804362684488297,
-0.06166696920990944,
0.00032348540844395757,
-0.03055569902062416,
-0.08205434679985046,
-0.00821809470653534,
0.00697416951879859,
0.016606012359261513,
0.10812531411647797,
0.020969394594430923,
0.05400780588388443,
-0.026613866910338402,
0.039453551173210144,
0.015868591144680977,
0.04796067997813225,
0.07602889835834503,
0.007550685666501522,
0.0367620475590229
] |
https://github.com/j0hngou/LLMCWM
| 10 | 0 | 0 | 0 |
From Models to Microtheories: Distilling a Model's Topical Knowledge for Grounded Question-Answering
|
https://openreview.net/forum?id=JV8zULNh24
|
[
"Nathaniel Weir",
"Bhavana Dalvi Mishra",
"Orion Weller",
"Oyvind Tafjord",
"Sam Hornstein",
"Alexander Sabol",
"Peter Jansen",
"Benjamin Van Durme",
"Peter Clark"
] |
Poster
|
Recent reasoning methods (e.g., chain-of-thought) help users understand how language models (LMs) answer a single question, but they do little to reveal the LM’s overall understanding, or “theory,” about the question’s topic, making it still hard to trust the model. Our goal is to materialize such theories - here called microtheories (a linguistic analog of logical microtheories) - as a set of sentences encapsulating an LM’s core knowledge about a topic. These statements systematically work together to entail answers to a set of questions to both engender trust and improve performance. Our approach is to first populate a knowledge store with (model-generated) sentences that entail answers to training questions, and then distill those down to a core microtheory which is concise, general, and non-redundant. We show that, when added to a general corpus (e.g., Wikipedia), microtheories can supply critical information not necessarily present in the corpus, improving both a model’s ability to ground its answers to verifiable knowledge (i.e., show how answers are systematically entailed by documents in the corpus, grounding up to +8% more answers), and the accuracy of those grounded answers (up to +8% absolute). We also show that, in a human evaluation in the medical domain, our distilled microtheories contain a significantly higher concentration of topically critical facts than the non-distilled knowledge store. Finally, we show we can quantify the coverage of a microtheory for a topic (characterized by a dataset) using a notion of p-relevance. Together, these suggest that microtheories are an efficient distillation of an LM’s topic-relevant knowledge, that they can usefully augment existing corpora, and can provide both performance gains and an interpretable, verifiable window into the model’s knowledge of a topic.
|
microtheory, textual entailment, knowledge representation, natural language reasoning, text retrieval, automatic knowledge base construction
|
We introduce a question-driven algorithm to generate a list of LM-believed knowledge statements most relevant to answering questions in a dataset.
| 11,685 |
2412.17701
|
[
0.028329672291874886,
-0.08463452756404877,
0.020286085084080696,
0.041134774684906006,
0.08792006969451904,
0.013228200376033783,
0.06874764710664749,
0.08794260025024414,
0.007786833681166172,
-0.001583325327374041,
-0.01293132547289133,
-0.08020146936178207,
0.08541087806224823,
0.048804618418216705,
0.026688771322369576,
0.006233599502593279,
0.07117734104394913,
0.08097332715988159,
-0.13075105845928192,
0.0036191591061651707,
0.08245529979467392,
-0.033783670514822006,
-0.0028591996524482965,
0.04409929737448692,
-0.03519465774297714,
-0.04602551460266113,
0.010929105803370476,
-0.004355676472187042,
-0.03731025755405426,
0.0002785244432743639,
0.05337478592991829,
0.0453537181019783,
-0.01064149010926485,
0.04636569321155548,
0.061764221638441086,
0.09710677713155746,
-0.004057040438055992,
0.03155426308512688,
0.003481629304587841,
-0.07042752206325531,
-0.05385403335094452,
-0.014787117950618267,
0.055733051151037216,
0.011957230977714062,
0.0811295211315155,
0.025462985038757324,
0.018124932423233986,
0.021076953038573265,
-0.0613308809697628,
-0.009394855238497257,
-0.09576331079006195,
-0.07424584031105042,
-0.045818887650966644,
0.020154112949967384,
-0.012021547183394432,
-0.010553085245192051,
-0.037223927676677704,
0.016957029700279236,
-0.05435740202665329,
-0.07983969151973724,
-0.039599183946847916,
-0.10509830713272095,
-0.07442113757133484,
0.010102108120918274,
-0.006641003768891096,
-0.08606971055269241,
-0.017898129299283028,
0.040536582469940186,
-0.00958205759525299,
0.08609383553266525,
0.029967021197080612,
0.013788562268018723,
-0.03916623443365097,
0.017470955848693848,
-0.0090049859136343,
0.053178973495960236,
0.013809992000460625,
-0.00797836109995842,
-0.007440552115440369,
-0.0256354920566082,
0.013899448327720165,
0.07067650556564331,
0.07255610078573227,
0.007475416641682386,
0.02793280966579914,
0.007769438438117504,
0.060714807361364365,
0.017645323649048805,
-0.017056794837117195,
-0.028002658858895302,
-0.04688992351293564,
-0.05640290677547455,
0.04499766230583191,
-0.030091656371951103,
0.11642053723335266,
0.04703149199485779,
0.04536188766360283,
-0.10388406366109848,
0.004429991357028484,
0.10038349032402039,
0.051704589277505875,
0.10423563420772552,
-0.028441380709409714,
-0.10313171148300171,
0.008148826658725739,
-0.0083544896915555,
-0.038162555545568466,
0.06749161332845688,
0.05266696587204933,
-0.06686286628246307,
0.014010239392518997,
0.0906648263335228,
0.008460646495223045,
0.02744070626795292,
0.060085151344537735,
-0.02659885212779045,
0.09959776699542999,
0.01585918292403221,
0.03776615113019943,
-0.010138984769582748,
0.004457399249076843,
0.045392755419015884,
0.053844984620809555,
-0.015534351579844952,
0.003416123567149043,
-0.04879043623805046,
-0.04642587900161743,
2.526534548437979e-34,
0.08644693344831467,
0.034460194408893585,
0.06954114884138107,
0.06257767975330353,
-0.03380708023905754,
0.02734411135315895,
-0.02461957186460495,
0.04993009939789772,
0.01521073468029499,
-0.035442255437374115,
-0.019860178232192993,
0.05461619421839714,
-0.052240412682294846,
0.00247731851413846,
0.035702478140592575,
0.029320552945137024,
-0.0949314534664154,
-0.010307329706847668,
0.010949455201625824,
-0.022644151002168655,
0.04968731105327606,
-0.03949087858200073,
-0.0190526582300663,
-0.008110431022942066,
-0.012162440456449986,
-0.029187576845288277,
0.08896537125110626,
-0.03695426136255264,
-0.05461966618895531,
-0.012840175069868565,
-0.10893691331148148,
-0.09238429367542267,
-0.033459316939115524,
0.03184041380882263,
-0.005219185724854469,
-0.04266074299812317,
-0.010962731204926968,
-0.09436824172735214,
0.017633715644478798,
-0.10035604238510132,
-0.001832671114243567,
0.018685443326830864,
0.03406887874007225,
0.05191539227962494,
-0.010001570917665958,
0.0027016433887183666,
-0.052614543586969376,
-0.05368545278906822,
-0.01590452902019024,
-0.05493872985243797,
0.09009356796741486,
0.022854311391711235,
0.017447352409362793,
-0.07024899125099182,
0.004040894564241171,
0.008227365091443062,
-0.03930547833442688,
0.011312587186694145,
0.015889110043644905,
-0.002352995565161109,
0.036952514201402664,
0.005100593436509371,
0.02588099241256714,
0.06251821666955948,
0.015322119928896427,
0.028147757053375244,
-0.08049602806568146,
0.03385918587446213,
0.021092137321829796,
0.01058857049793005,
-0.05338129773736,
-0.012877401895821095,
0.011336900293827057,
-0.07014796882867813,
0.013243813067674637,
0.042893555015325546,
-0.0027144309133291245,
-0.05491441488265991,
0.02864629030227661,
0.00021050797658972442,
0.054715584963560104,
-0.07158548384904861,
-0.04870026186108589,
-0.04589223116636276,
-0.05015907809138298,
-0.03545992821455002,
0.08089470118284225,
-0.04977528750896454,
0.029396183788776398,
-0.056096162647008896,
-0.042290132492780685,
-0.09307932108640671,
-0.01409201230853796,
0.004293912090361118,
-0.022154193371534348,
-6.494892355863671e-34,
-0.01930728368461132,
-0.08688443154096603,
-0.04232632741332054,
0.11101020127534866,
-0.005980548448860645,
-0.03201186656951904,
-0.07145558297634125,
0.01806326024234295,
-0.07681073993444443,
-0.04121611267328262,
-0.06398578733205795,
-0.03384348005056381,
0.009780862368643284,
0.02421795390546322,
-0.020015032961964607,
-0.01924729160964489,
0.030599061399698257,
0.00014598284906242043,
0.02177976630628109,
0.13552716374397278,
-0.0047352551482617855,
0.06278981268405914,
-0.12485809624195099,
0.05945158004760742,
-0.024354232475161552,
0.04398543760180473,
-0.04342234879732132,
0.07798761129379272,
-0.07221728563308716,
0.028361858800053596,
-0.059694282710552216,
-0.04856076091527939,
0.014395496807992458,
-0.09978258609771729,
-0.11242487281560898,
0.038666289299726486,
0.05323801934719086,
-0.027016127482056618,
-0.0074311550706624985,
0.05719751492142677,
0.019193455576896667,
-0.04973596706986427,
-0.05124218016862869,
-0.03966940566897392,
-0.0607653446495533,
-0.03807707130908966,
-0.09452982246875763,
-0.014018576592206955,
0.021684881299734116,
0.020582212135195732,
-0.034668367356061935,
-0.08407819271087646,
-0.05406003072857857,
-0.04877112805843353,
-0.000018981018001795746,
-0.024869509041309357,
-0.051655881106853485,
0.04206417500972748,
0.050591252744197845,
0.044828783720731735,
-0.03496531397104263,
0.008655764162540436,
-0.03169352933764458,
-0.010105514898896217,
0.04694966971874237,
-0.005060641560703516,
-0.06497634947299957,
-0.01961452327668667,
-0.04295891895890236,
0.04088051989674568,
0.0019313724478706717,
0.0029075322672724724,
0.05513510853052139,
-0.005962500348687172,
0.04238395765423775,
0.01130027137696743,
0.050338830798864365,
-0.05305692180991173,
-0.04433523863554001,
-0.0738091692328453,
0.058695778250694275,
-0.03027070313692093,
0.0750935822725296,
0.036621853709220886,
0.07298417389392853,
-0.04641760513186455,
0.052787475287914276,
0.08604098856449127,
-0.009066787548363209,
0.06950787454843521,
-0.04798083379864693,
-0.019915523007512093,
0.013349797576665878,
0.0349607840180397,
-0.08659138530492783,
-4.4727332948468757e-8,
-0.009650000371038914,
0.09557025879621506,
-0.007287594024091959,
0.0809471383690834,
-0.01489612739533186,
-0.031714748591184616,
-0.03576456755399704,
0.016729798167943954,
-0.02375231496989727,
-0.0036256485618650913,
0.07050567865371704,
-0.033190131187438965,
-0.05824313312768936,
-0.0019480051705613732,
0.05620737746357918,
0.17140977084636688,
0.011280435137450695,
-0.01339374016970396,
-0.044990792870521545,
-0.018988830968737602,
0.10495686531066895,
0.04565068334341049,
-0.034934017807245255,
0.10445303469896317,
0.0701485350728035,
0.04206172376871109,
-0.005158062092959881,
0.13963288068771362,
0.05079912766814232,
0.023674864321947098,
-0.03548450395464897,
-0.034631386399269104,
-0.04099827632308006,
0.040070269256830215,
-0.021128378808498383,
0.10175997018814087,
-0.0013949662679806352,
-0.039804745465517044,
0.033132154494524,
-0.10385256260633469,
-0.012729134410619736,
0.035060517489910126,
-0.07561787217855453,
0.043234068900346756,
0.06644505262374878,
-0.027424335479736328,
-0.06272684782743454,
-0.01836305856704712,
-0.028837278485298157,
0.06632008403539658,
-0.02625984698534012,
-0.0752258449792862,
-0.000992615707218647,
0.07174848020076752,
0.008336023427546024,
-0.003704447764903307,
0.0030444241128861904,
0.06207067519426346,
-0.01676206849515438,
0.03940652683377266,
0.014227167703211308,
-0.009073887020349503,
0.05915825441479683,
0.027014048770070076
] |
https://github.com/nweir127/microtheories
| 0 | 0 | 0 | 0 |
Enabling Realtime Reinforcement Learning at Scale with Staggered Asynchronous Inference
|
https://openreview.net/forum?id=fXb9BbuyAD
|
[
"Matthew Riemer",
"Gopeshh Subbaraj",
"Glen Berseth",
"Irina Rish"
] |
Poster
|
Realtime environments change even as agents perform action inference and learning, thus requiring high interaction frequencies to effectively minimize regret. However, recent advances in machine learning involve larger neural networks with longer inference times, raising questions about their applicability in realtime systems where reaction time is crucial. We present an analysis of lower bounds on regret in realtime reinforcement learning (RL) environments to show that minimizing long-term regret is generally impossible within the typical sequential interaction and learning paradigm, but often becomes possible when sufficient asynchronous compute is available. We propose novel algorithms for staggering asynchronous inference processes to ensure that actions are taken at consistent time intervals, and demonstrate that use of models with high action inference times is only constrained by the environment's effective stochasticity over the inference horizon, and not by action frequency. Our analysis shows that the number of inference processes needed scales linearly with increasing inference times while enabling use of models that are multiple orders of magnitude larger than existing approaches when learning from a realtime simulation of Game Boy games such as Pokemon and Tetris.
|
Realtime Environments, Asynchronous Algorithms, Time Discretization, Real World Deployment, Deep Reinforcement Learning
| null | 11,682 |
2412.14355
|
[
-0.09132906049489975,
-0.0446452721953392,
-0.021639926359057426,
0.043989147990942,
0.034853532910346985,
0.04576491191983223,
-0.033875007182359695,
0.006063457112759352,
0.06627000868320465,
-0.023389816284179688,
-0.044613342732191086,
-0.03374297916889191,
-0.04248243570327759,
0.04021444544196129,
0.0007437000167556107,
-0.019828028976917267,
0.04227916896343231,
0.030729319900274277,
-0.050366148352622986,
-0.04959888756275177,
-0.0028373056557029486,
0.011177758686244488,
0.06676112860441208,
-0.009211618453264236,
0.007195532787591219,
-0.047943804413080215,
-0.05274621397256851,
-0.02091728337109089,
-0.02201405167579651,
0.017508365213871002,
0.05073739215731621,
-0.01869458518922329,
0.015054017305374146,
-0.07194916903972626,
-0.06937597692012787,
0.07723221182823181,
-0.07752738147974014,
-0.028448928147554398,
-0.01624269038438797,
-0.055863045156002045,
0.0003685799892991781,
-0.018978269770741463,
-0.015263466164469719,
0.03208911046385765,
0.07581775635480881,
-0.02308543212711811,
0.024964092299342155,
-0.044933952391147614,
0.020852116867899895,
-0.015181707218289375,
-0.034221500158309937,
-0.03359365835785866,
0.026336602866649628,
0.006161229684948921,
-0.03102875128388405,
0.0453830324113369,
0.010533272288739681,
0.05921833962202072,
0.023341821506619453,
0.019244592636823654,
-0.024087948724627495,
-0.1442645937204361,
-0.06260258704423904,
-0.03324967250227928,
-0.12054967880249023,
0.027382194995880127,
0.015008199028670788,
0.031992025673389435,
0.07290413230657578,
0.0958377867937088,
0.019709507003426552,
0.04892750829458237,
-0.06160943582653999,
-0.0036969580687582493,
-0.014934465289115906,
-0.023871319368481636,
-0.024073796346783638,
-0.076331228017807,
0.026033200323581696,
0.01115893293172121,
-0.03538617119193077,
-0.044165145605802536,
-0.04678426682949066,
-0.009982039220631123,
0.10456397384405136,
0.017617274075746536,
0.023654170334339142,
0.061254728585481644,
0.06232397258281708,
-0.006866951938718557,
-0.006112842354923487,
-0.012977088801562786,
0.01544109359383583,
-0.0010774603579193354,
0.006459699012339115,
0.04218699038028717,
-0.032028235495090485,
-0.08510538935661316,
-0.026047581806778908,
0.085035040974617,
-0.01296463143080473,
0.06022781878709793,
-0.04464513808488846,
0.028363997116684914,
0.036668285727500916,
-0.016855819150805473,
-0.002451034262776375,
0.028275202959775925,
0.031645361334085464,
-0.09209603071212769,
0.027254248037934303,
-0.05673632025718689,
0.04196951165795326,
0.02728729322552681,
0.0003625134122557938,
0.00208943709731102,
0.03680414706468582,
-0.09968157112598419,
-0.04730777069926262,
0.04741990193724632,
0.00782583188265562,
0.047696638852357864,
0.030031098052859306,
-0.005029160995036364,
0.06224462389945984,
-0.04740091785788536,
-0.06924225389957428,
3.957068837224974e-33,
0.0231335386633873,
-0.019002458080649376,
0.09333504736423492,
-0.06432496756315231,
0.04282780736684799,
0.023983947932720184,
0.07273279875516891,
-0.013388163410127163,
0.008536637760698795,
-0.029994212090969086,
0.040230002254247665,
-0.015352452173829079,
-0.07014013826847076,
0.008147518150508404,
0.05056851729750633,
-0.012313754297792912,
-0.04518262296915054,
0.02203371748328209,
0.07026087492704391,
-0.04722476378083229,
0.037790969014167786,
-0.09165631979703903,
-0.03608943149447441,
-0.00440965173766017,
-0.024546172469854355,
-0.009138870052993298,
-0.0004952464951202273,
0.032958317548036575,
0.03711182624101639,
-0.0177144818007946,
-0.016530847176909447,
0.04546517878770828,
-0.09009275585412979,
0.09700477868318558,
0.04188324511051178,
-0.017728112637996674,
-0.07319997996091843,
-0.015086139552295208,
-0.004112679045647383,
0.03897610679268837,
-0.025468939915299416,
0.008818813599646091,
-0.0010509990388527513,
0.06739277392625809,
-0.07701791077852249,
-0.09880920499563217,
0.06082424521446228,
0.005977464374154806,
-0.06594990938901901,
0.004822575952857733,
0.01802455447614193,
0.07468530535697937,
0.04231131076812744,
-0.0796271339058876,
-0.033025357872247696,
0.025651289150118828,
0.04384110867977142,
0.04548893868923187,
-0.013306308537721634,
0.07654517889022827,
0.022547287866473198,
-0.0579521581530571,
0.03485168144106865,
0.04084339365363121,
-0.02562595345079899,
0.0920136347413063,
-0.053083617240190506,
0.09588994830846786,
0.03508016839623451,
-0.009809290990233421,
0.02464580349624157,
0.026024911552667618,
0.0798330307006836,
-0.13373050093650818,
0.10080966353416443,
0.04685499146580696,
0.0758475661277771,
-0.06299334019422531,
0.006069779396057129,
0.014855894260108471,
-0.04328405112028122,
-0.052984267473220825,
-0.04298393800854683,
0.014802210032939911,
0.011458988301455975,
-0.0504261814057827,
0.03523003309965134,
-0.08831828832626343,
0.018311714753508568,
-0.008027439936995506,
-0.186909481883049,
-0.039616912603378296,
0.014231935143470764,
-0.016133634373545647,
0.017062483355402946,
-4.7588008002078994e-33,
0.05702275410294533,
-0.10955929756164551,
-0.015707796439528465,
0.044443774968385696,
0.024936622008681297,
-0.004851992707699537,
-0.04217596352100372,
-0.010084385983645916,
0.01533390674740076,
-0.09157123416662216,
-0.03206958994269371,
0.0969686284661293,
0.07827633619308472,
0.016369925811886787,
-0.023340275511145592,
-0.0596596822142601,
0.08040479570627213,
-0.01743960753083229,
0.04865216836333275,
-0.0022313776426017284,
0.02454501762986183,
0.04206009581685066,
-0.06406763941049576,
0.04067560285329819,
-0.04497077688574791,
-0.053149230778217316,
-0.008205932565033436,
0.013191195204854012,
-0.0063032652251422405,
0.018349668011069298,
-0.060099393129348755,
-0.06772813200950623,
0.011609051376581192,
0.00960451364517212,
0.0075040291994810104,
0.06296361982822418,
0.12056499719619751,
0.020891442894935608,
-0.09292839467525482,
0.015463034622371197,
0.10788343846797943,
-0.1374918818473816,
-0.07099533081054688,
-0.007042244076728821,
0.047304917126894,
0.03278624638915062,
-0.08820652961730957,
0.023979537189006805,
-0.05778757855296135,
0.004874511156231165,
-0.009848236106336117,
0.09612949937582016,
-0.032068584114313126,
-0.03496551513671875,
-0.03518075868487358,
-0.09210220724344254,
-0.024508723989129066,
-0.03919201344251633,
0.10332643240690231,
0.041683658957481384,
-0.03742830082774162,
-0.03864796459674835,
-0.07516045868396759,
-0.00677261920645833,
0.00193660834338516,
0.01903696544468403,
-0.03070903569459915,
-0.010047079995274544,
0.030173247680068016,
0.008643193170428276,
0.019268319010734558,
0.04323551431298256,
-0.006725845392793417,
0.018834641203284264,
-0.016487540677189827,
-0.06441885232925415,
0.015491808764636517,
-0.019970795139670372,
0.031079668551683426,
-0.012602352537214756,
0.040183499455451965,
-0.005467296577990055,
0.07779799401760101,
0.03295362740755081,
0.017639311030507088,
0.04871615767478943,
0.0039663296192884445,
0.07101336121559143,
0.059033192694187164,
-0.017954619601368904,
-0.03947768360376358,
0.012123079970479012,
0.033123429864645004,
0.026633847504854202,
-0.07281677424907684,
-4.065851300083523e-8,
-0.02767828106880188,
0.09310944378376007,
0.12155784666538239,
0.05009178817272186,
0.11859547346830368,
0.01209845021367073,
-0.0009271163726225495,
-0.004778078757226467,
0.010279029607772827,
0.00740521727129817,
0.08179076761007309,
0.03684226796030998,
0.10727588087320328,
-0.08396143466234207,
-0.030183961614966393,
0.04472268000245094,
-0.010364088229835033,
-0.12841852009296417,
0.021604666486382484,
-0.0058612157590687275,
0.020287727937102318,
0.022250844165682793,
0.01954466663300991,
-0.09416306018829346,
0.01923949085175991,
-0.059808049350976944,
0.030246593058109283,
0.10810976475477219,
0.040970586240291595,
-0.006520334165543318,
0.019895656034350395,
-0.007239857222884893,
0.027985963970422745,
-0.011396159417927265,
0.05390362814068794,
-0.02874944545328617,
0.0766587182879448,
-0.027835914865136147,
0.045869775116443634,
-0.09403129667043686,
-0.09721500426530838,
-0.007640539668500423,
-0.050567787140607834,
0.022515488788485527,
-0.024196742102503777,
0.004803127143532038,
-0.06278016418218613,
-0.1274532973766327,
0.061639152467250824,
-0.017145443707704544,
0.05475438013672829,
0.017136452719569206,
0.01259960513561964,
0.07473823428153992,
0.07263538986444473,
0.03613777458667755,
0.013531939126551151,
-0.08312248438596725,
-0.0203876793384552,
0.09110409021377563,
0.019119229167699814,
0.00007049750274745747,
-0.06716376543045044,
0.018028609454631805
] |
https://github.com/cerc-aai/realtime_rl
| 1 | 0 | 0 | 0 |
On the Transfer of Object-Centric Representation Learning
|
https://openreview.net/forum?id=bSq0XGS3kW
|
[
"Aniket Rajiv Didolkar",
"Andrii Zadaianchuk",
"Anirudh Goyal",
"Michael Curtis Mozer",
"Yoshua Bengio",
"Georg Martius",
"Maximilian Seitzer"
] |
Poster
|
The goal of object-centric representation learning is to decompose visual scenes into a structured representation that isolates the entities into individual vectors. Recent successes have shown that object-centric representation learning can be scaled to real-world scenes by utilizing features from pre-trained foundation models like DINO. However, so far, these object-centric methods have mostly been applied in-distribution, with models trained and evaluated on the same dataset. This is in contrast to the underlying foundation models, which have been shown to be applicable to a wide range of data and tasks. Thus, in this work, we answer the question of whether current real-world capable object-centric methods exhibit similar levels of transferability by introducing a benchmark comprising seven different synthetic and real-world datasets. We analyze the factors influencing performance under transfer and find that training on diverse real-world images improves generalization to unseen scenarios. Furthermore, inspired by the success of task-specific fine-tuning in foundation models, we introduce a novel fine-tuning strategy to adapt pre-trained vision encoders for the task of object discovery. We find that the proposed approach results in state-of-the-art performance for unsupervised object discovery, exhibiting strong zero-shot transfer to unseen datasets.
|
representation learning, object-centric learning, object-centric representation learning, unsupervised learning, transfer, zero-shot, generalization
|
We study the transfer of object-centric representations and show that a finetuning strategy leads to state-of-the-art performance.
| 11,679 | null |
[
-0.009685617871582508,
-0.12139804661273956,
0.02701818011701107,
0.02635692059993744,
0.18142810463905334,
-0.019057078287005424,
-0.045397453010082245,
-0.04477870836853981,
0.0020462546963244677,
-0.028823895379900932,
-0.08119937777519226,
-0.11564565449953079,
-0.020631002262234688,
0.1257810741662979,
-0.021496044471859932,
-0.013232405297458172,
0.1351596713066101,
0.10884346067905426,
-0.12032867223024368,
0.052434515208005905,
-0.013914797455072403,
-0.027165498584508896,
0.055560413748025894,
0.0079440176486969,
0.08662112057209015,
-0.001478237216360867,
0.048033080995082855,
-0.07996419072151184,
-0.009203735738992691,
-0.08013691753149033,
-0.02742820233106613,
-0.011364555917680264,
-0.07159202545881271,
-0.013817256316542625,
-0.05975653976202011,
0.11675476282835007,
0.0025754538364708424,
0.010181826539337635,
-0.009383322671055794,
-0.07232709974050522,
0.004256875719875097,
0.05330248177051544,
-0.00450806925073266,
0.05181438475847244,
0.058229684829711914,
0.05798162519931793,
0.012229520827531815,
-0.037343114614486694,
0.0680907815694809,
-0.028460044413805008,
-0.0525137297809124,
-0.02548384852707386,
-0.08580717444419861,
0.06577829271554947,
-0.027543991804122925,
0.061252396553754807,
0.055654965341091156,
-0.08929996192455292,
0.021436411887407303,
0.012837473303079605,
0.0027215087320655584,
-0.04265185445547104,
0.07098066061735153,
0.008945788256824017,
-0.012798965908586979,
-0.06239204853773117,
0.02517547272145748,
-0.0004740874283015728,
0.01062723621726036,
-0.07146086543798447,
0.0008196001290343702,
0.03552784025669098,
0.017982421442866325,
-0.0062873181886971,
0.1146455779671669,
0.033226411789655685,
0.038349028676748276,
0.03184717893600464,
0.04589853063225746,
-0.08394657075405121,
0.027277512475848198,
0.0739213228225708,
0.027376070618629456,
-0.019470201805233955,
0.14299438893795013,
0.0508725643157959,
0.024638602510094643,
0.09924103319644928,
-0.059519484639167786,
-0.005446030758321285,
-0.028356127440929413,
-0.0014755084412172437,
0.051498472690582275,
-0.033227719366550446,
-0.016456279903650284,
-0.0423286035656929,
0.014703648164868355,
-0.07620862126350403,
0.008534847758710384,
0.04565456882119179,
-0.025170346722006798,
0.03837183862924576,
0.07023444026708603,
-0.008424133993685246,
0.017241233959794044,
0.021482186391949654,
0.0364583358168602,
-0.0060034445486962795,
-0.029676789417862892,
-0.038418468087911606,
0.030894579365849495,
0.04369086027145386,
-0.023431697860360146,
0.007495444267988205,
-0.018208235502243042,
-0.0006787083111703396,
-0.09933535754680634,
-0.006427030544728041,
-0.024358807131648064,
0.013463759794831276,
0.0050284783355891705,
-0.018441224470734596,
-0.026019291952252388,
0.01122293807566166,
-0.0028164335526525974,
-0.007017037365585566,
-0.04290826991200447,
1.213836956165195e-33,
-0.011343195103108883,
0.017803631722927094,
0.04546145349740982,
0.0005538020050153136,
0.022755181416869164,
-0.0760839581489563,
0.015549900941550732,
0.1142062246799469,
-0.025897622108459473,
0.06985406577587128,
-0.042856328189373016,
-0.01364104449748993,
-0.027048904448747635,
0.11479394882917404,
0.0075447047129273415,
-0.0027132381219416857,
-0.067075215280056,
0.0112588070333004,
-0.024484673514962196,
0.05184043571352959,
-0.025545133277773857,
-0.010610796511173248,
-0.0532585009932518,
-0.04438360035419464,
0.036016423255205154,
-0.0035443443339318037,
0.027677498757839203,
-0.039921458810567856,
-0.039391789585351944,
-0.018371393904089928,
-0.04476376622915268,
0.03312751278281212,
0.04641414433717728,
0.08157427608966827,
0.004123677499592304,
-0.059058982878923416,
-0.0094680180773139,
-0.05026445910334587,
0.01827709563076496,
-0.01784953847527504,
0.03272107616066933,
0.11715047806501389,
-0.05586821958422661,
-0.005151771474629641,
-0.042745716869831085,
-0.045288361608982086,
0.056011732667684555,
0.044995978474617004,
-0.09059365838766098,
0.0034145021345466375,
0.02709357626736164,
-0.0049354606308043,
-0.0511600598692894,
-0.09148163348436356,
-0.05545896291732788,
-0.01996556855738163,
0.04562612995505333,
0.012546253390610218,
0.052481744438409805,
-0.04921865463256836,
0.013882331550121307,
0.052088573575019836,
-0.024018194526433945,
0.11958403140306473,
-0.01837967149913311,
0.014095272868871689,
-0.003334564622491598,
0.026881271973252296,
0.0016010794788599014,
0.046057626605033875,
-0.03934242203831673,
0.07010764628648758,
-0.06465483456850052,
-0.06811006367206573,
0.08172065764665604,
-0.0435982309281826,
0.05233941599726677,
-0.04706243798136711,
-0.042016807943582535,
0.1153985857963562,
-0.024234073236584663,
0.05003679543733597,
0.01811682991683483,
-0.07193521410226822,
-0.017184779047966003,
-0.013605182059109211,
0.06026678532361984,
-0.08215495944023132,
0.017304757609963417,
0.009627360850572586,
-0.021621549502015114,
-0.08267420530319214,
-0.06446947902441025,
-0.04824349656701088,
0.005705959163606167,
1.7342634093926281e-34,
0.03150856867432594,
0.05353865027427673,
-0.046142302453517914,
0.051246702671051025,
-0.01097694132477045,
-0.0714210718870163,
-0.02490984834730625,
0.05199010297656059,
-0.025189682841300964,
-0.05187787860631943,
-0.010677398182451725,
-0.03547453507781029,
-0.034717489033937454,
-0.054743003100156784,
-0.007041865028440952,
0.005168262403458357,
-0.019240999594330788,
-0.028019433841109276,
-0.07560259848833084,
-0.01355914305895567,
-0.016392789781093597,
0.06431668251752853,
-0.057481806725263596,
0.02380196936428547,
-0.07673682272434235,
0.0667855441570282,
-0.05234648659825325,
0.026126088574528694,
0.011506929062306881,
0.001362503389827907,
-0.05448857694864273,
-0.032375067472457886,
-0.000804194831289351,
0.0027698036283254623,
-0.09545842558145523,
0.03926801681518555,
0.032806750386953354,
-0.0388374924659729,
-0.05159757286310196,
0.0925658643245697,
0.010476134717464447,
-0.03354691341519356,
-0.10232575982809067,
0.051765602082014084,
-0.08290547877550125,
-0.06059008091688156,
-0.051139190793037415,
-0.013646847568452358,
0.05107530578970909,
0.05892913043498993,
-0.041474245488643646,
0.08891081064939499,
-0.057490747421979904,
-0.052887819707393646,
0.01953243277966976,
-0.058764468878507614,
-0.02257484942674637,
0.005963646341115236,
0.12335392832756042,
0.004449006635695696,
-0.02243591472506523,
-0.05941157788038254,
-0.022230902686715126,
-0.01646052487194538,
0.012093441560864449,
-0.004627502989023924,
-0.10275984555482864,
0.018900923430919647,
0.04231496527791023,
0.05124806612730026,
-0.03297439590096474,
0.0003528862725943327,
0.04388555884361267,
-0.008302275091409683,
-0.05643583834171295,
-0.03674420341849327,
0.06081438064575195,
-0.053969621658325195,
0.060688287019729614,
0.023036740720272064,
0.061306606978178024,
-0.08639132231473923,
0.018461013212800026,
0.09511598944664001,
0.1239546686410904,
0.12874023616313934,
-0.00024750729789957404,
-0.03487296774983406,
0.025869853794574738,
-0.018227532505989075,
-0.025524910539388657,
-0.004936569835990667,
-0.017460975795984268,
0.06378627568483353,
0.013526338152587414,
-4.017795873778596e-8,
-0.05013939365744591,
0.01951148547232151,
0.0374152772128582,
-0.0035837162286043167,
-0.01446606870740652,
0.009579192847013474,
-0.016361091285943985,
0.09597893804311752,
-0.09562863409519196,
0.032173074781894684,
0.01639595255255699,
0.02936881221830845,
-0.011489169672131538,
-0.00026502192486077547,
0.0602559931576252,
0.05831081420183182,
0.002488735830411315,
0.02026938647031784,
0.01304465252906084,
0.08785770833492279,
0.025670532137155533,
0.035833656787872314,
0.006741632707417011,
-0.04279524087905884,
0.033982742577791214,
-0.0533674955368042,
-0.0027442362625151873,
0.07386433333158493,
0.093014657497406,
0.011450027115643024,
-0.028014741837978363,
0.00971242692321539,
0.018095260486006737,
-0.015438638627529144,
0.061871208250522614,
0.045927442610263824,
0.00864010863006115,
0.06245831772685051,
-0.06563813239336014,
-0.06070104241371155,
0.07929953932762146,
0.09380690008401871,
-0.06992675364017487,
0.04443295672535896,
0.011502216570079327,
0.031639568507671356,
-0.023428011685609818,
-0.02095744013786316,
-0.0039178067818284035,
-0.016593389213085175,
0.04338407889008522,
-0.013220921158790588,
-0.02736731991171837,
0.06711848825216293,
0.04426303878426552,
0.0037082829512655735,
0.044614408165216446,
-0.06622454524040222,
0.05168262496590614,
0.08150758594274521,
-0.03258378431200981,
-0.00011455659114290029,
-0.0012646653922274709,
-0.0026736101135611534
] | 0 | 0 | 0 | 0 |
|
ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities
|
https://openreview.net/forum?id=lfPkGWXLLf
|
[
"Ezra Karger",
"Houtan Bastani",
"Chen Yueh-Han",
"Zachary Jacobs",
"Danny Halawi",
"Fred Zhang",
"Philip Tetlock"
] |
Poster
|
Forecasts of future events are essential inputs into informed decision-making. Machine learning (ML) systems have the potential to deliver forecasts at scale, but there is no framework for evaluating the accuracy of ML systems on a standardized set of forecasting questions. To address this gap, we introduce ForecastBench: a dynamic benchmark that evaluates the accuracy of ML systems on an automatically generated and regularly updated set of 1,000 forecasting questions. To avoid any possibility of data leakage, ForecastBench is comprised solely of questions about future events that have no known answer at the time of submission. We quantify the capabilities of current ML systems by collecting forecasts from expert (human) forecasters, the general public, and LLMs on a random subset of questions from the benchmark ($N=200$). While LLMs have achieved super-human performance on many benchmarks, they perform less well here: expert forecasters outperform the top-performing LLM ($p$-value $<0.001$). We display system and human scores in a public leaderboard at www.forecastbench.org.
|
language models, evaluation, benchmark, forecasting, LLM, decision-making, datasets
|
We build a dynamic benchmark to evaluate LLMs' ability to forecast future events
| 11,678 |
2409.19839
|
[
-0.08975613117218018,
-0.05778112635016441,
-0.013164578005671501,
0.0925443172454834,
0.03148604929447174,
-0.031034937128424644,
-0.025547979399561882,
0.039464469999074936,
-0.025362126529216766,
0.004617377649992704,
-0.17625786364078522,
-0.07862204313278198,
0.024788837879896164,
-0.018499065190553665,
-0.08416971564292908,
-0.00006491789827123284,
0.016329031437635422,
-0.03502762317657471,
-0.0489923469722271,
-0.04837358742952347,
-0.06422164291143417,
-0.0007502396474592388,
0.005688865669071674,
-0.0024865204468369484,
0.059508614242076874,
-0.07082292437553406,
-0.004622282460331917,
-0.021904490888118744,
-0.006511288229376078,
0.03498417139053345,
0.028237681835889816,
-0.07912668585777283,
-0.0024808053858578205,
0.010608676820993423,
-0.05292125418782234,
-0.005783459171652794,
-0.027111317962408066,
0.013904862105846405,
0.037789735943078995,
-0.03374196216464043,
0.05845998227596283,
-0.056032463908195496,
-0.0018267204286530614,
-0.031468819826841354,
0.07411187887191772,
0.052626416087150574,
0.012280769646167755,
-0.04405311122536659,
-0.0016408603405579925,
0.025377687066793442,
-0.13839374482631683,
-0.06862290948629379,
0.045163027942180634,
-0.003528106492012739,
-0.046870119869709015,
0.08438212424516678,
-0.014867689460515976,
0.017137372866272926,
-0.003003819612786174,
-0.018716542050242424,
-0.08823536336421967,
-0.10534150898456573,
-0.020355911925435066,
0.013969332911074162,
0.04002321884036064,
0.03923638537526131,
-0.014749142341315746,
0.10678686946630478,
0.007208930794149637,
-0.018614213913679123,
0.008489293046295643,
0.11708318442106247,
-0.02988194301724434,
0.05378253012895584,
0.020748762413859367,
0.000651716545689851,
0.021080052480101585,
-0.02792886272072792,
0.06538074463605881,
0.004715039860457182,
-0.009459532797336578,
-0.020282240584492683,
-0.010467803105711937,
0.0054016802459955215,
0.054242923855781555,
0.03911246359348297,
0.051398865878582,
0.08737456053495407,
0.035086095333099365,
-0.01479079108685255,
-0.03419087082147598,
-0.07539919763803482,
0.04384729638695717,
0.005858723074197769,
0.015881938859820366,
0.09429843723773956,
0.01381831057369709,
-0.07186361402273178,
-0.03529765456914902,
0.07201041281223297,
0.0018472692463546991,
0.08103562891483307,
0.01875384710729122,
-0.025672301650047302,
-0.0019394075497984886,
0.03486304357647896,
-0.02879464626312256,
0.006368725094944239,
0.012221654877066612,
-0.08195221424102783,
0.01374862901866436,
-0.01724388636648655,
0.05657416954636574,
-0.03304753825068474,
0.02953764609992504,
-0.04042709246277809,
-0.004450132604688406,
0.0837971568107605,
-0.03810496628284454,
0.10024966299533844,
-0.008583572693169117,
0.10381539911031723,
0.07390391081571579,
0.04086015745997429,
0.08977053314447403,
0.033326469361782074,
-0.03603857010602951,
1.97316406246816e-33,
0.01040597166866064,
-0.0034180372022092342,
0.0832260400056839,
0.0055651795119047165,
0.028621988371014595,
0.004157663322985172,
-0.007477270904928446,
0.03497710078954697,
0.030299874022603035,
0.002856436651200056,
-0.06400883942842484,
0.023416999727487564,
-0.048615314066410065,
0.024049509316682816,
0.06645505875349045,
0.04547519609332085,
0.014892838895320892,
0.04893134906888008,
-0.06230386719107628,
-0.0009340218384750187,
0.0698799416422844,
-0.0020761603955179453,
-0.02225818671286106,
-0.06330256909132004,
0.020142262801527977,
-0.004834598861634731,
0.07409138232469559,
-0.020044466480612755,
0.028935205191373825,
0.009218004532158375,
-0.052085720002651215,
-0.04002341255545616,
-0.03985036164522171,
0.009664286859333515,
0.003888240084052086,
0.004077823832631111,
-0.11657506227493286,
-0.02626342885196209,
0.0029128333553671837,
0.045184098184108734,
-0.0387427881360054,
0.0256835725158453,
-0.02517649531364441,
-0.022736800834536552,
-0.08624517172574997,
-0.08605235815048218,
-0.016815923154354095,
-0.03565313294529915,
0.012731272727251053,
-0.06818260252475739,
0.013359718024730682,
-0.023046836256980896,
-0.03930599242448807,
-0.030703185126185417,
-0.01506088674068451,
0.008239478804171085,
0.06813351809978485,
-0.03464328125119209,
0.001387020223774016,
0.033155057579278946,
-0.03574783727526665,
0.011364035308361053,
0.004090707283467054,
0.034366682171821594,
-0.047667134553194046,
0.04280265420675278,
0.032019734382629395,
0.110896036028862,
0.03601553291082382,
-0.021893583238124847,
0.026722125709056854,
-0.0038023809902369976,
0.06130316108465195,
-0.06140345707535744,
0.041687898337841034,
0.04442469775676727,
0.05065617337822914,
-0.06848244369029999,
0.04739966243505478,
0.041010499000549316,
0.01713729277253151,
-0.008322721347212791,
0.025071339681744576,
-0.036061111837625504,
0.03700154647231102,
-0.04128632694482803,
0.02909492515027523,
0.023539608344435692,
0.004698571283370256,
-0.00526458490639925,
-0.13079550862312317,
-0.06169378384947777,
-0.05380120873451233,
0.0041243815794587135,
-0.04868023097515106,
-2.9603495447474943e-33,
-0.0727834478020668,
-0.0440773107111454,
-0.05996670573949814,
0.06953155249357224,
0.020032301545143127,
-0.05312852934002876,
-0.10438600927591324,
-0.015709610655903816,
-0.016147920861840248,
-0.036233507096767426,
-0.05939820408821106,
0.011474314145743847,
0.07463307678699493,
0.015172223560512066,
-0.022056857123970985,
-0.05996503308415413,
-0.03348172456026077,
-0.006148910149931908,
-0.01138033252209425,
-0.04170166328549385,
-0.055445246398448944,
0.055386852473020554,
-0.08756302297115326,
0.07083121687173843,
-0.007504899986088276,
-0.02394024096429348,
0.04134131595492363,
0.09471912682056427,
-0.021072985604405403,
-0.03051610104739666,
-0.05215299502015114,
-0.07344978302717209,
-0.03195854276418686,
-0.005014870781451464,
0.01616767793893814,
0.05759008973836899,
0.05143329128623009,
-0.07842113822698593,
-0.01865796558558941,
0.12148977071046829,
0.04660026356577873,
0.002943861298263073,
-0.060123980045318604,
-0.0640680193901062,
-0.08695858716964722,
0.041743047535419464,
-0.07719481736421585,
0.0036062300205230713,
0.05562072992324829,
0.039088718593120575,
-0.017378458753228188,
0.027361925691366196,
-0.09658113867044449,
0.013796220533549786,
-0.045718107372522354,
-0.023699315264821053,
-0.0804835632443428,
0.0022534276358783245,
0.03184229135513306,
0.0853416919708252,
-0.06418122351169586,
0.0008592312806285918,
0.05084396153688431,
0.022435085847973824,
0.03567991405725479,
-0.005165287759155035,
-0.029319610446691513,
0.038336269557476044,
-0.0032989748287945986,
-0.01630748249590397,
0.08751329779624939,
-0.04953679069876671,
-0.014892010018229485,
0.048210665583610535,
-0.011229163967072964,
0.05710919201374054,
-0.005284052807837725,
-0.018954316154122353,
0.027674701064825058,
-0.06603489071130753,
-0.023113952949643135,
-0.008126392960548401,
0.06144222989678383,
0.016022982075810432,
0.05117439851164818,
0.044836342334747314,
0.13159498572349548,
0.02328214794397354,
0.021038271486759186,
-0.00329772150143981,
-0.02492806874215603,
0.029755856841802597,
-0.08947427570819855,
0.03544359281659126,
-0.053555507212877274,
-4.099594264062034e-8,
-0.038757484406232834,
0.09517142921686172,
0.03127296641469002,
0.08619216829538345,
0.029236705973744392,
-0.017102079465985298,
-0.017876358702778816,
0.04156021028757095,
-0.009419841691851616,
-0.03014467842876911,
0.09527154266834259,
-0.10490594059228897,
-0.04341064766049385,
0.021598288789391518,
0.0725802481174469,
0.03690251708030701,
-0.061033718287944794,
0.01611669547855854,
-0.05705394968390465,
-0.12237614393234253,
0.0960334911942482,
0.03792639076709747,
0.033351343125104904,
-0.06994318962097168,
0.028650745749473572,
-0.07782646268606186,
-0.08298782259225845,
0.12170521914958954,
-0.042155761271715164,
0.0753142237663269,
-0.009590763598680496,
-0.0037515389267355204,
0.010419096797704697,
-0.05569019168615341,
0.09965220838785172,
0.06449855864048004,
-0.017622634768486023,
-0.023294394835829735,
0.02539542317390442,
0.03752102330327034,
0.011217840947210789,
0.038177456706762314,
-0.005475026089698076,
0.05016506835818291,
0.021624775603413582,
-0.07572944462299347,
0.00975485797971487,
-0.04628082737326622,
0.04034707695245743,
-0.07619085907936096,
0.0580158606171608,
-0.06709568947553635,
0.051416441798210144,
0.07418930530548096,
0.1546144187450409,
0.03229404613375664,
-0.015596901066601276,
-0.05256232991814613,
0.031512580811977386,
0.04470089077949524,
0.04077945277094841,
-0.04165683314204216,
-0.11089997738599777,
0.06759379804134369
] |
https://github.com/forecastingresearch/forecastbench
| 18 | 0 | 0 | 0 |
Score-based free-form architectures for high-dimensional Fokker-Planck equations
|
https://openreview.net/forum?id=5qg6JPSgCj
|
[
"Feng Liu",
"Faguo Wu",
"Xiao Zhang"
] |
Poster
|
Deep learning methods incorporate PDE residuals as the loss function for solving Fokker-Planck equations, and usually impose the proper normalization condition to avoid a trivial solution. However, soft constraints require careful balancing of multi-objective loss functions, and specific network architectures may limit representation capacity under hard constraints. In this paper, we propose a novel framework: Fokker-Planck neural network (FPNN) that adopts a score PDE loss to decouple the score learning and the density normalization into two stages. Our method allows free-form network architectures to model the unnormalized density and strictly satisfy normalization constraints by post-processing. We demonstrate the effectiveness on various high-dimensional steady-state Fokker-Planck (SFP) equations, achieving superior accuracy and over a 20$\times$ speedup compared to state-of-the-art methods. Without any labeled data, FPNNs achieve the mean absolute percentage error (MAPE) of 11.36\%, 13.87\% and 12.72\% for 4D Ring, 6D Unimodal and 6D Multi-modal problems respectively, requiring only 256, 980, and 980 parameters. Experimental results highlights the potential as a universal fast solver for handling more than 20-dimensional SFP equations, with great gains in efficiency, accuracy, memory and computational resource usage.
|
Fokker-Planck Equations, Normalization Condition, Score Model, Physical Constraints.
|
We propose a score PDE loss to decouple the normalization condition, allowing for free-form architectures in solving high-dimensional Fokker-Planck equations.
| 11,675 | null |
[
-0.03874139115214348,
-0.016632214188575745,
-0.0015361609403043985,
0.05281239002943039,
0.0023412504233419895,
0.035433754324913025,
-0.10564439743757248,
-0.03188612312078476,
0.001122194342315197,
-0.00924394279718399,
-0.037696246057748795,
-0.14291980862617493,
-0.027071425691246986,
0.03849569335579872,
-0.06383038312196732,
0.004787584766745567,
-0.007673345506191254,
0.05176921933889389,
-0.1496085822582245,
0.00847745779901743,
-0.03188303858041763,
0.060429032891988754,
-0.022961735725402832,
0.0420023649930954,
0.05914299562573433,
-0.03696848452091217,
0.07959921658039093,
-0.058686643838882446,
0.017530350014567375,
0.012161016464233398,
0.08012957870960236,
-0.05760824680328369,
0.01609058864414692,
0.04683259502053261,
-0.04177949205040932,
0.05707453191280365,
-0.07102299481630325,
-0.054450392723083496,
-0.0874195247888565,
-0.06066083908081055,
-0.009011708199977875,
0.01865432970225811,
-0.03297531604766846,
0.03604396432638168,
0.03468531742691994,
0.017668919637799263,
0.000286654569208622,
-0.05386466532945633,
0.06042346730828285,
-0.00931638479232788,
0.057841233909130096,
0.02093459479510784,
-0.0722571611404419,
0.09929661452770233,
0.07273539155721664,
0.009215573780238628,
0.007605512626469135,
-0.022133374586701393,
0.0012002539588138461,
0.0063422564417123795,
-0.04445241019129753,
-0.07304927706718445,
-0.020173653960227966,
-0.023647941648960114,
0.05707394331693649,
0.01173478178679943,
-0.06742896884679794,
0.056620318442583084,
-0.03516095504164696,
0.042415767908096313,
-0.0016884240321815014,
0.021984871476888657,
0.041475810110569,
-0.05416741222143173,
0.07544969767332077,
0.044126540422439575,
0.0680270567536354,
0.07637446373701096,
0.0553552582859993,
-0.0038201280403882265,
0.039727043360471725,
0.05271779000759125,
0.08353397995233536,
-0.003196136327460408,
0.14750248193740845,
-0.01676226407289505,
-0.04624580964446068,
0.05186265707015991,
-0.033133938908576965,
-0.04942025616765022,
-0.04102451726794243,
0.031922124326229095,
-0.07011768966913223,
-0.04371555894613266,
0.005962742958217859,
-0.007171241566538811,
0.002788654761388898,
-0.0746581181883812,
-0.05420972779393196,
0.09534496814012527,
-0.07322045415639877,
0.0028334674425423145,
0.004143253900110722,
-0.007388815749436617,
0.07897470891475677,
-0.048681627959012985,
0.11908408999443054,
0.06806524097919464,
0.030709896236658096,
-0.06489542126655579,
0.03965875878930092,
0.013089470565319061,
0.003237774595618248,
0.0524434857070446,
0.030284736305475235,
-0.018656395375728607,
0.005977281369268894,
-0.029938839375972748,
0.03459048271179199,
0.03966827690601349,
-0.16354259848594666,
-0.0025231025647372007,
-0.060659363865852356,
0.0398741252720356,
-0.011069941334426403,
0.03373798355460167,
-0.05676764249801636,
2.2260672639437597e-33,
-0.0026408943813294172,
0.03279661014676094,
-0.018587782979011536,
-0.053865354508161545,
0.0047267405316233635,
-0.04915618523955345,
-0.030133916065096855,
0.012472775764763355,
0.024688882753252983,
0.07315593212842941,
-0.019273746758699417,
0.016442086547613144,
-0.1468108743429184,
0.06487664580345154,
0.04618382826447487,
-0.03221575915813446,
-0.0067019592970609665,
0.05646860972046852,
0.016699783504009247,
-0.09621107578277588,
0.04750906676054001,
-0.05214230343699455,
0.028246570378541946,
-0.022134050726890564,
0.013798301108181477,
-0.0012500544544309378,
-0.022095955908298492,
-0.04605419561266899,
-0.017570653930306435,
-0.02186410315334797,
-0.12945903837680817,
-0.016698487102985382,
-0.04642173647880554,
-0.016844160854816437,
-0.010482522659003735,
0.011813094839453697,
-0.022503014653921127,
0.011144508607685566,
-0.037074435502290726,
-0.0073308562859892845,
-0.014458912424743176,
0.0168156661093235,
-0.005405740812420845,
-0.0318647064268589,
-0.052631765604019165,
0.018690943717956543,
0.004927977919578552,
-0.0419788733124733,
0.015670906752347946,
-0.008929518982768059,
-0.016977302730083466,
-0.03042525425553322,
-0.008329088799655437,
-0.013658923096954823,
0.030993901193141937,
-0.023996805772185326,
0.13423247635364532,
-0.020519034937024117,
0.019745951518416405,
0.02956804633140564,
-0.026338202878832817,
0.014107337221503258,
-0.015437623485922813,
0.0440489687025547,
0.011781604960560799,
0.0329611636698246,
-0.0525221973657608,
0.12712620198726654,
0.015850957483053207,
-0.06354162842035294,
-0.028515568003058434,
0.06528804451227188,
0.06705007702112198,
-0.09819559752941132,
0.05840152129530907,
0.023901108652353287,
0.020442143082618713,
-0.13209496438503265,
-0.055067285895347595,
0.02461288869380951,
-0.007985532283782959,
0.09419218450784683,
-0.09005480259656906,
-0.011946619488298893,
-0.0022412326652556658,
-0.011322272010147572,
-0.011136195622384548,
-0.06953790783882141,
-0.007225734647363424,
-0.026670463383197784,
-0.07207923382520676,
-0.06927573680877686,
-0.028952402994036674,
-0.04737763851881027,
-0.04383641108870506,
-6.524084564204892e-34,
-0.06095670163631439,
0.05802840366959572,
-0.13416393101215363,
0.013848777860403061,
0.07234670966863632,
-0.026998775079846382,
-0.062002867460250854,
0.03559807315468788,
0.04218810051679611,
-0.04569997638463974,
0.03708360716700554,
-0.06416036933660507,
0.013175168074667454,
-0.009120417758822441,
0.047484613955020905,
0.020476730540394783,
-0.13085786998271942,
0.06280045956373215,
0.0909406766295433,
-0.0402238629758358,
0.03655994310975075,
0.06505364179611206,
-0.03560025990009308,
0.07767527550458908,
-0.04039445519447327,
0.018756622448563576,
-0.02039380557835102,
0.08802971243858337,
0.03035666048526764,
0.02039823681116104,
-0.03654838725924492,
-0.020946893841028214,
0.012897268868982792,
0.028822073712944984,
-0.004504323471337557,
0.010661250911653042,
0.014934544451534748,
0.05091756209731102,
-0.006085371598601341,
-0.02682379074394703,
0.048828378319740295,
-0.03704426810145378,
-0.1006125956773758,
0.04216749593615532,
0.017771748825907707,
0.09194726496934891,
-0.0007943434757180512,
-0.07155664265155792,
0.0585528165102005,
0.06293145567178726,
0.02828941121697426,
0.0053299288265407085,
-0.045961782336235046,
0.013237875886261463,
-0.1373860239982605,
-0.005109719466418028,
-0.013740436173975468,
0.006295214407145977,
0.08366589993238449,
0.0014407216804102063,
-0.035551104694604874,
-0.058646511286497116,
0.011471402831375599,
0.027405228465795517,
-0.046762820333242416,
-0.05307449772953987,
0.048565573990345,
0.044679056853055954,
0.06906002014875412,
0.11450579762458801,
-0.008550125174224377,
-0.017623091116547585,
0.04508598893880844,
0.05459476634860039,
0.009538223966956139,
-0.03628717362880707,
0.06541799753904343,
-0.010910450480878353,
0.024349307641386986,
-0.050790511071681976,
0.053351834416389465,
0.056327641010284424,
0.04709303006529808,
0.055406805127859116,
0.0974571630358696,
0.03076769784092903,
0.06038263812661171,
-0.021287983283400536,
0.028300896286964417,
0.036743029952049255,
0.01565956138074398,
0.11088631302118301,
0.028545286506414413,
0.018534738570451736,
-0.05314231663942337,
-4.400481401489742e-8,
-0.05457067862153053,
0.03993422910571098,
-0.012424238957464695,
-0.014549982734024525,
0.054374173283576965,
-0.01609116606414318,
0.007476251106709242,
-0.0070191603153944016,
-0.02680971845984459,
-0.02019944228231907,
0.02207774668931961,
-0.05125042796134949,
-0.10523231327533722,
-0.08023713529109955,
-0.09290888160467148,
0.03766482323408127,
0.022119708359241486,
-0.006891613360494375,
0.03625297546386719,
-0.012546701356768608,
-0.021489284932613373,
0.07054734230041504,
0.012956107035279274,
-0.01999921351671219,
0.026081344112753868,
-0.005948969628661871,
-0.020684050396084785,
0.06716906279325485,
0.06130904704332352,
0.02120569348335266,
0.02338552102446556,
0.0179129745811224,
0.07075940072536469,
0.09091401100158691,
-0.015972964465618134,
0.08091431111097336,
-0.0007494539022445679,
0.02050418220460415,
-0.03638715296983719,
0.05041394755244255,
-0.046193115413188934,
0.14386117458343506,
-0.010016120038926601,
-0.057854775339365005,
-0.036703307181596756,
0.0326153002679348,
0.06566963344812393,
-0.018733803182840347,
0.09221716970205307,
0.031203219667077065,
0.02484656684100628,
0.030452685430645943,
0.00614696741104126,
0.04088471457362175,
0.022372009232640266,
0.008341205306351185,
-0.0011438338551670313,
-0.005263812839984894,
0.03367767855525017,
0.04411841928958893,
-0.01118888147175312,
-0.028591398149728775,
-0.04897268861532211,
0.017194714397192
] | 0 | 0 | 0 | 0 |
|
ReGen: Generative Robot Simulation via Inverse Design
|
https://openreview.net/forum?id=EbCUbPZjM1
|
[
"Phat Tan Nguyen",
"Tsun-Hsuan Wang",
"Zhang-Wei Hong",
"Erfan Aasi",
"Andrew Silva",
"Guy Rosman",
"Sertac Karaman",
"Daniela Rus"
] |
Poster
|
Simulation plays a key role in scaling robot learning and validating policies, but constructing simulations remains labor-intensive. In this paper, we introduce ReGen, a generative simulation framework that automates this process using inverse design. Given an agent's behavior (such as a motion trajectory or objective function) and its textual description, we infer the underlying scenarios and environments that could have caused the behavior.
Our approach leverages large language models to construct and expand a graph that captures cause-and-effect relationships and relevant entities with properties in the environment, which is then processed to configure a robot simulation environment. Our approach supports (i) augmenting simulations based on ego-agent behaviors, (ii) controllable, counterfactual scenario generation, (iii) reasoning about agent cognition and mental states, and (iv) reasoning with distinct sensing modalities, such as braking due to faulty GPS signals.
We demonstrate our method in autonomous driving and robot manipulation tasks, generating more diverse, complex simulated environments compared to existing simulations with high success rates, and enabling controllable generation for corner cases. This approach enhances the validation of robot policies and supports data or simulation augmentation, advancing scalable robot learning for improved generalization and robustness.
|
generative simulation, robot, autonomous driving, large language model, inverse design
|
ReGen generates simulations from behavior by inferring plausible simulated environment where the behavior could have occurred through inverse design.
| 11,673 | null |
[
-0.039968036115169525,
-0.07000230997800827,
0.036427296698093414,
0.008613476529717445,
0.03889406844973564,
-0.05468948557972908,
0.008939350955188274,
-0.008924408815801144,
-0.005722775589674711,
0.03316545486450195,
-0.020942572504281998,
-0.07537288218736649,
0.03971780464053154,
0.060041941702365875,
0.05108730122447014,
-0.007146165240556002,
0.09305588901042938,
0.07765670120716095,
-0.06134602054953575,
-0.0024906417820602655,
0.0366215705871582,
0.03131350129842758,
0.04551011696457863,
-0.01566966250538826,
-0.04672524705529213,
0.020678063854575157,
0.04220462217926979,
-0.04731254652142525,
-0.007541770115494728,
-0.019827693700790405,
0.10163076967000961,
-0.019169973209500313,
-0.024354541674256325,
-0.013210944831371307,
0.0022070989944040775,
0.08634161949157715,
-0.055688872933387756,
0.026965277269482613,
0.01843477226793766,
-0.045607790350914,
-0.043491803109645844,
-0.03825083374977112,
0.07123196125030518,
-0.010110494680702686,
0.09392163902521133,
0.015084358863532543,
-0.040101177990436554,
-0.0413668155670166,
0.0572892464697361,
0.014523041434586048,
-0.08669034391641617,
-0.01900552399456501,
0.017680754885077477,
-0.03271379694342613,
0.029241180047392845,
0.06099700182676315,
0.062265679240226746,
-0.020916417241096497,
0.05581948161125183,
-0.03392326086759567,
0.0074638379737734795,
-0.054862167686223984,
-0.025200631469488144,
-0.06209825724363327,
-0.14935164153575897,
-0.022514240816235542,
-0.039640672504901886,
0.049878913909196854,
0.0044191316701471806,
0.05527472496032715,
0.051762789487838745,
0.029268596321344376,
0.0302665363997221,
-0.007335326634347439,
0.03652530536055565,
0.05592627823352814,
-0.09439429640769958,
-0.02045709826052189,
0.009150613099336624,
-0.07182390242815018,
0.016933010891079903,
-0.010591267608106136,
0.06739245355129242,
0.07214636355638504,
0.04804728180170059,
-0.029360728338360786,
0.03734594210982323,
0.09848868101835251,
0.09836876392364502,
0.06061778962612152,
-0.06099824234843254,
-0.024387981742620468,
0.03241966664791107,
0.007226936984807253,
-0.009069708175957203,
0.033914655447006226,
0.006399545818567276,
-0.10328098386526108,
0.008582239970564842,
0.0461655929684639,
0.008998501114547253,
0.041230496019124985,
-0.013079863972961903,
0.03150852024555206,
-0.02180410362780094,
-0.006544193718582392,
0.01050194539129734,
-0.002434109104797244,
0.03512253239750862,
-0.010038545355200768,
0.05331720784306526,
-0.04413396120071411,
0.03881213441491127,
0.01816985383629799,
0.03976347669959068,
-0.035116296261548996,
-0.0018176279263570905,
0.0018189090769737959,
-0.042573582381010056,
0.0502263642847538,
-0.008604741655290127,
0.015997936949133873,
0.07321446388959885,
0.06120362877845764,
0.1248873844742775,
-0.05269832909107208,
0.051998019218444824,
-2.8823729433147056e-33,
0.04939914867281914,
0.03531631827354431,
0.022347556427121162,
0.11084124445915222,
0.12830905616283417,
0.021478543058037758,
-0.05468931794166565,
0.033454593271017075,
-0.013054459355771542,
-0.023964587599039078,
-0.006186367012560368,
-0.0134192556142807,
-0.0855306014418602,
0.08355624973773956,
0.038389213383197784,
0.05306468904018402,
-0.037545278668403625,
0.03813323378562927,
-0.06861356645822525,
-0.018048135563731194,
0.022074300795793533,
-0.0396929606795311,
-0.045597635209560394,
0.0019065048545598984,
-0.016328880563378334,
0.039447180926799774,
0.031823694705963135,
0.009881136007606983,
0.03616318106651306,
-0.016208691522479057,
-0.10185330361127853,
0.028106408193707466,
-0.006048408802598715,
0.10758675634860992,
-0.06107334792613983,
0.061548538506031036,
-0.041428230702877045,
-0.07987028360366821,
-0.04422534629702568,
0.0389518141746521,
-0.026780996471643448,
0.025427842512726784,
0.0185205340385437,
0.005326502025127411,
-0.08260045200586319,
-0.12616655230522156,
0.056257229298353195,
0.025446539744734764,
0.011178378015756607,
-0.06008710712194443,
0.013316256925463676,
0.02973075956106186,
0.029932009056210518,
-0.07742072641849518,
-0.06803051382303238,
-0.05364995077252388,
0.01553942821919918,
0.05606494098901749,
0.011488568037748337,
0.03713138774037361,
0.010012596845626831,
0.0019616330973803997,
-0.015323974192142487,
0.07609283179044724,
0.005846808198839426,
0.0006451747030951083,
0.006470776628702879,
0.011648095212876797,
0.09586046636104584,
0.05888902768492699,
-0.03066645748913288,
0.027938444167375565,
-0.012667222879827023,
-0.03377358987927437,
-0.006384868174791336,
0.04234272241592407,
0.018831893801689148,
-0.058524757623672485,
-0.06184033676981926,
0.024254079908132553,
-0.07079929858446121,
-0.020419659093022346,
-0.06222908943891525,
0.04744376987218857,
0.032651063054800034,
-0.095002181828022,
0.024481769651174545,
-0.06784084439277649,
-0.013068636879324913,
-0.031007882207632065,
-0.07510382682085037,
-0.0833696573972702,
-0.053635451942682266,
0.057411473244428635,
-0.02443966083228588,
8.122424436827111e-34,
-0.04563199728727341,
0.006686881184577942,
0.013607040978968143,
0.030054807662963867,
-0.021561967208981514,
-0.0731964036822319,
-0.003552329260855913,
-0.0043091000989079475,
0.006472193636000156,
-0.09847308695316315,
-0.126246839761734,
-0.05781478434801102,
0.10280823707580566,
0.040822584182024,
-0.016105536371469498,
-0.06219683960080147,
0.006804957520216703,
-0.06965355575084686,
-0.03439009189605713,
0.021107789129018784,
-0.017427949234843254,
0.10674211382865906,
-0.04854558780789375,
-0.0017628847854211926,
0.0036528673954308033,
0.049377333372831345,
-0.01001376286149025,
0.029696928337216377,
0.06534700095653534,
0.08345182240009308,
-0.019912114366889,
0.015908239409327507,
0.030846845358610153,
-0.013402558863162994,
-0.006467747502028942,
0.09920943528413773,
0.02609279751777649,
-0.039107922464609146,
-0.04667144641280174,
-0.01890185847878456,
0.03477217257022858,
-0.09539886564016342,
-0.07840404659509659,
0.004226767458021641,
0.01238839142024517,
-0.03612300753593445,
0.0028111636638641357,
0.06130672246217728,
-0.0032115457579493523,
-0.017980121076107025,
-0.0476033091545105,
0.04608524590730667,
-0.05667415261268616,
-0.04115213826298714,
-0.0609944611787796,
-0.12673555314540863,
0.0018206514650955796,
-0.008284731768071651,
0.060271620750427246,
-0.02135816216468811,
-0.07997676730155945,
-0.015336262062191963,
0.07802924513816833,
0.0034296060912311077,
-0.0687640905380249,
0.0037674782797694206,
-0.05711197480559349,
-0.0017387812258675694,
0.0559963621199131,
-0.02010527066886425,
0.06874018162488937,
0.036948107182979584,
-0.009029433131217957,
0.020951073616743088,
0.08967061340808868,
-0.03426313027739525,
-0.016190897673368454,
0.005288451910018921,
-0.014220049604773521,
-0.026151694357395172,
0.03832240775227547,
0.03134956210851669,
0.03320150822401047,
0.13643236458301544,
-0.08048874884843826,
0.05159084126353264,
-0.06906366348266602,
0.1172330304980278,
0.005061546340584755,
0.07372632622718811,
-0.013735294342041016,
0.05090973898768425,
-0.05759363994002342,
0.12743538618087769,
-0.0927371010184288,
-3.889393695999388e-8,
-0.13045968115329742,
0.06660629063844681,
0.09925585985183716,
0.005378569010645151,
-0.016688909381628036,
-0.017187312245368958,
0.021516047418117523,
-0.05138906463980675,
-0.04067207872867584,
-0.042774349451065063,
0.07367554306983948,
0.0026718275621533394,
0.0023646685294806957,
-0.03528185933828354,
-0.03306998312473297,
0.04963526129722595,
0.051274217665195465,
-0.003289908869192004,
-0.007860061712563038,
0.02047707885503769,
-0.0018532847752794623,
-0.000043841366277774796,
-0.01380931492894888,
-0.09207680076360703,
-0.0026985788717865944,
-0.08718899637460709,
-0.06066340208053589,
-0.0031732686329632998,
0.021548880264163017,
-0.01880829408764839,
0.0194960068911314,
-0.04459841549396515,
0.057696953415870667,
0.05167870596051216,
-0.012325882911682129,
0.025841088965535164,
0.030308788642287254,
-0.048733312636613846,
0.04908433556556702,
-0.11292906850576401,
0.040787722915410995,
0.07208279520273209,
-0.033983539789915085,
-0.004284741822630167,
-0.01957428827881813,
0.0030002978164702654,
-0.05599985644221306,
-0.09624415636062622,
0.024870172142982483,
-0.01630345731973648,
-0.074202761054039,
-0.09208317846059799,
-0.02021252177655697,
0.03749589994549751,
0.07194644957780838,
0.0633004903793335,
-0.02106904238462448,
-0.05351286381483078,
0.043007079511880875,
0.06514943391084671,
-0.0076319146901369095,
0.08471610397100449,
-0.01615486666560173,
-0.05101238191127777
] | 0 | 0 | 0 | 0 |
|
Flow-based Variational Mutual Information: Fast and Flexible Approximations
|
https://openreview.net/forum?id=spDUv05cEq
|
[
"Caleb Dahlke",
"Jason Pacheco"
] |
Poster
|
Mutual Information (MI) is a fundamental measure of dependence between random variables, but its practical application is limited because it is difficult to calculate in many circumstances. Variational methods offer one approach by introducing an approximate distribution to create various bounds on MI, which in turn is an easier optimization problem to solve. In practice, the variational distribution chosen is often a Gaussian, which is convenient but lacks flexibility in modeling complicated distributions. In this paper, we introduce new classes of variational estimators based on Normalizing Flows that extend the previous Gaussian-based variational estimators. Our new estimators maintain many of the same theoretical guarantees while simultaneously enhancing the expressivity of the variational distribution. We experimentally verify that our new methods are effective on large MI problems where discriminative-based estimators, such as MINE and InfoNCE, are fundamentally limited. Furthermore, we compare against a diverse set of benchmarking tests to show that the flow-based estimators often perform as well, if not better, than the discriminative-based counterparts. Finally, we demonstrate how these estimators can be effectively utilized in the Bayesian Optimal Experimental Design setting for online sequential decision making.
|
Mutual Information, Variational Methods, Normalizing Flows, Bayesian Optimal Experimental Design
|
We introduce two new flow-based variational Mutual Information estimations that are comparable in expressiveness to critic-based methods, such as MINE or InfoNCE, but can also approximate large MI
| 11,672 | null |
[
-0.042557355016469955,
-0.05792536959052086,
0.11773394048213959,
-0.01936432532966137,
0.06846261024475098,
0.011192995123565197,
0.07483259588479996,
-0.0137956403195858,
0.0718584731221199,
-0.029533980414271355,
-0.0042114281095564365,
-0.032800979912281036,
0.02772226557135582,
0.0431070402264595,
-0.037174027413129807,
-0.04481692239642143,
0.07263409346342087,
0.09226270765066147,
-0.11989343911409378,
0.005025115329772234,
0.02699911594390869,
0.08143150806427002,
-0.05087757855653763,
-0.015752242878079414,
0.08182478696107864,
-0.00292849144898355,
0.04868031665682793,
-0.006214093882590532,
-0.021332424134016037,
-0.030991874635219574,
0.06647232174873352,
0.04105415940284729,
-0.041830796748399734,
-0.002489339094609022,
-0.036261122673749924,
-0.01685953512787819,
-0.035511866211891174,
0.04261864721775055,
0.00214672414585948,
0.034310340881347656,
-0.028628434985876083,
0.044933050870895386,
0.04657397419214249,
0.08103081583976746,
-0.02140507474541664,
0.040330685675144196,
0.028468267992138863,
-0.06739045679569244,
0.025196371600031853,
0.10388845950365067,
-0.00699971616268158,
0.044469837099313736,
-0.010106989182531834,
-0.017855731770396233,
-0.008064968511462212,
-0.0702294260263443,
0.015758488327264786,
-0.05397956818342209,
0.004871697165071964,
0.11712431907653809,
0.00025276668020524085,
-0.13302455842494965,
0.005344343837350607,
0.02726082131266594,
-0.01717829890549183,
-0.05945007875561714,
0.08951611071825027,
0.04372589662671089,
0.020404379814863205,
-0.004446462262421846,
-0.012370821088552475,
0.08761458098888397,
-0.15861956775188446,
-0.021607134491205215,
-0.013747197575867176,
-0.011339657939970493,
0.07086405903100967,
0.006507745012640953,
-0.0037136084865778685,
-0.0741000697016716,
-0.06043655797839165,
0.03296790271997452,
0.015573553740978241,
0.028814706951379776,
0.10452698916196823,
-0.016023268923163414,
-0.06487218290567398,
0.08768662810325623,
0.07228284329175949,
-0.01525660790503025,
-0.09925029426813126,
0.0008662445470690727,
-0.07664049416780472,
0.009791664779186249,
0.10330234467983246,
0.04553483799099922,
-0.018188703805208206,
-0.10891091078519821,
0.024296848103404045,
0.10101345926523209,
-0.011683647520840168,
0.016536356881260872,
0.035435985773801804,
-0.0030683453660458326,
0.03656119480729103,
-0.0820980817079544,
0.06519291549921036,
0.05604426935315132,
-0.01111370138823986,
-0.03329644724726677,
-0.009009627625346184,
0.026640472933650017,
0.013198774307966232,
0.018431883305311203,
-0.04372934252023697,
-0.03744874522089958,
0.03954595699906349,
0.05193489044904709,
0.04087064042687416,
0.04642779752612114,
-0.001853116787970066,
-0.03431553393602371,
-0.005026435945183039,
-0.0371997244656086,
-0.023844368755817413,
0.06931192427873611,
-0.027227533981204033,
4.422116810271334e-33,
0.021235743537545204,
-0.027488568797707558,
0.05640505999326706,
0.06663283705711365,
-0.0428171269595623,
0.03577426075935364,
-0.04387162998318672,
0.002170666353777051,
-0.056335847824811935,
0.014568724669516087,
-0.00621409947052598,
0.013263360597193241,
-0.14136458933353424,
0.000060141042922623456,
-0.011576741933822632,
0.006261457223445177,
-0.09768125414848328,
0.08048547059297562,
-0.0050059049390256405,
0.02506943605840206,
0.11505144834518433,
-0.049125079065561295,
-0.061431385576725006,
-0.039753928780555725,
-0.03802654147148132,
-0.06128057464957237,
0.03743400797247887,
-0.07384628802537918,
-0.014620223082602024,
-0.012874234467744827,
-0.08453048765659332,
0.003872218541800976,
0.018675310537219048,
-0.019643424078822136,
0.0239938423037529,
-0.03181565925478935,
-0.029093274846673012,
0.017624473199248314,
-0.07515471428632736,
0.003764030523598194,
-0.01220021117478609,
0.040226005017757416,
0.05524198338389397,
-0.07467231154441833,
-0.08287650346755981,
-0.07448139041662216,
-0.02483374997973442,
-0.00608769292011857,
0.07909104228019714,
-0.08200842887163162,
0.0068741850554943085,
0.0006335877114906907,
0.0008292868733406067,
0.004209516104310751,
0.019516032189130783,
-0.034824054688215256,
-0.039767976850271225,
0.009438516572117805,
-0.014134597033262253,
0.030377939343452454,
-0.062126222997903824,
0.07540237158536911,
0.03001270443201065,
0.009435033425688744,
0.08261284977197647,
0.06487468630075455,
-0.02357475459575653,
0.0694064050912857,
0.06375794112682343,
-0.017645832151174545,
-0.014843759126961231,
0.009158904664218426,
-0.010685080662369728,
-0.08663562685251236,
0.025711053982377052,
0.02201073057949543,
0.0015305652050301433,
-0.030593641102313995,
0.007594500668346882,
0.008241941221058369,
-0.006530961953103542,
-0.01742919534444809,
0.015280728228390217,
0.049586016684770584,
-0.003099959809333086,
0.046175722032785416,
0.0007065095123834908,
-0.02676425874233246,
-0.08425820618867874,
-0.03376771882176399,
0.0009260440128855407,
0.017015116289258003,
-0.042398758232593536,
0.006426847539842129,
0.00925321877002716,
-4.875284576192711e-33,
-0.13514302670955658,
-0.016088180243968964,
0.06587811559438705,
0.02924707531929016,
0.12428539991378784,
-0.07309065759181976,
0.021035492420196533,
0.04363645240664482,
0.017986424267292023,
-0.06184963509440422,
-0.09171224385499954,
0.007476700935512781,
-0.032301165163517,
-0.011002705432474613,
-0.016919419169425964,
0.047609202563762665,
0.03903118893504143,
0.06406746804714203,
0.049794867634773254,
-0.018144529312849045,
-0.0274355448782444,
0.06990484148263931,
-0.02285374142229557,
-0.009714125655591488,
0.020226461812853813,
0.009080009534955025,
-0.08165110647678375,
0.10620924085378647,
-0.06917645037174225,
0.00321514206007123,
-0.08067770302295685,
-0.016599776223301888,
-0.08536625653505325,
-0.03995520621538162,
-0.02909008041024208,
0.005495601333677769,
-0.007223624270409346,
0.1098746582865715,
-0.048449017107486725,
0.08626042306423187,
0.02707572467625141,
0.053495485335588455,
-0.02230107970535755,
0.014256657101213932,
0.056329596787691116,
0.00185329036321491,
0.008140006102621555,
-0.06550753116607666,
0.04490893706679344,
0.034193407744169235,
-0.03495614975690842,
0.07689015567302704,
-0.10411108285188675,
0.06068539246916771,
-0.025015031918883324,
-0.09833671152591705,
0.0067755067721009254,
-0.0673704519867897,
0.0025364544708281755,
0.03976377099752426,
-0.03048844449222088,
-0.0019530865829437971,
-0.101148322224617,
0.022932328283786774,
-0.04335998371243477,
-0.013611367903649807,
0.021360615268349648,
-0.045053787529468536,
0.02451450563967228,
0.04908176138997078,
0.023774683475494385,
-0.06488621234893799,
0.037199292331933975,
0.023628665134310722,
0.09327087551355362,
0.017164835706353188,
0.10616391152143478,
0.001402893103659153,
-0.027881693094968796,
0.06638135761022568,
0.09489406645298004,
0.027058374136686325,
0.04160242900252342,
-0.06129785627126694,
0.11005937308073044,
0.0752735435962677,
-0.0009514325065538287,
-0.03098156675696373,
0.08373002707958221,
-0.029705079272389412,
-0.05999245494604111,
-0.028149934485554695,
-0.041973233222961426,
-0.011727269738912582,
-0.12849852442741394,
-4.8011358444455254e-8,
-0.02736937254667282,
-0.017397478222846985,
0.013587554916739464,
-0.07740821689367294,
0.024549752473831177,
0.00986683089286089,
-0.026460710912942886,
0.05250338464975357,
0.004964550491422415,
-0.10668341815471649,
0.12273763865232468,
-0.009637863375246525,
0.00531416991725564,
-0.04244603216648102,
0.006888856180012226,
0.044368207454681396,
-0.03002994880080223,
-0.032530173659324646,
0.013128639198839664,
0.00752123212441802,
-0.0069379983469843864,
-0.011335990391671658,
-0.020664414390921593,
-0.015449224039912224,
0.016005637124180794,
-0.07525461167097092,
0.05359065905213356,
0.016956863924860954,
0.10042982548475266,
-0.00849323533475399,
-0.0635756179690361,
0.00013727755867876112,
-0.0074453819543123245,
0.018165936693549156,
-0.057174429297447205,
0.0226746778935194,
-0.03234439343214035,
-0.013267962262034416,
-0.026004455983638763,
-0.02762901782989502,
-0.006937883794307709,
-0.0031304750591516495,
-0.026027614250779152,
0.005124715622514486,
0.11216074228286743,
0.053832247853279114,
0.05032068490982056,
-0.06678146123886108,
0.00994865968823433,
-0.03259500116109848,
0.029316343367099762,
-0.000605976558290422,
0.0375070720911026,
0.07236339896917343,
0.0772705078125,
-0.013009845279157162,
0.014904273673892021,
-0.04710658639669418,
0.019281528890132904,
-0.05054408311843872,
-0.012635692022740841,
0.03586936742067337,
-0.07025712728500366,
-0.021993856877088547
] | 0 | 0 | 0 | 0 |
|
Multi-agent cooperation through learning-aware policy gradients
|
https://openreview.net/forum?id=GkWA6NjePN
|
[
"Alexander Meulemans",
"Seijin Kobayashi",
"Johannes Von Oswald",
"Nino Scherrer",
"Eric Elmoznino",
"Blake Aaron Richards",
"Guillaume Lajoie",
"Blaise Aguera y Arcas",
"Joao Sacramento"
] |
Poster
|
Self-interested individuals often fail to cooperate, posing a fundamental challenge for multi-agent learning. How can we achieve cooperation among self-interested, independent learning agents? Promising recent work has shown that in certain tasks cooperation can be established between ``learning-aware" agents who model the learning dynamics of each other. Here, we present the first unbiased, higher-derivative-free policy gradient algorithm for learning-aware reinforcement learning, which takes into account that other agents are themselves learning through trial and error based on multiple noisy trials. We then leverage efficient sequence models to condition behavior on long observation histories that contain traces of the learning dynamics of other agents. Training long-context policies with our algorithm leads to cooperative behavior and high returns on standard social dilemmas, including a challenging environment where temporally-extended action coordination is required. Finally, we derive from the iterated prisoner's dilemma a novel explanation for how and when cooperation arises among self-interested learning-aware agents.
|
multi-agent learning, reinforcement learning, decentralized training, social dilemmas, cooperation, iterated prisoner's dilemma, melting pot
| null | 11,667 |
2410.18636
|
[
-0.03243999183177948,
-0.06415028870105743,
-0.04364515468478203,
0.03676066920161247,
0.08038877695798874,
0.023465238511562347,
0.05370917171239853,
-0.07348468154668808,
0.04853567108511925,
0.029994633048772812,
-0.022891899570822716,
0.014846607111394405,
-0.043104998767375946,
0.06970791518688202,
-0.007279697339981794,
0.00199854071252048,
0.005072425119578838,
0.015982013195753098,
-0.025547483935952187,
-0.03990401327610016,
-0.052980273962020874,
-0.019005028530955315,
0.09686388075351715,
0.029730932787060738,
-0.005126576405018568,
-0.031806252896785736,
0.019708817824721336,
-0.05991017073392868,
-0.015138184651732445,
0.009697868488729,
0.12521106004714966,
-0.06287325173616409,
0.0264938622713089,
-0.045943066477775574,
-0.04000403359532356,
0.05488482862710953,
-0.10145226866006851,
-0.009024976752698421,
-0.032753005623817444,
-0.010763953439891338,
0.05563608556985855,
0.0641183853149414,
-0.008822116069495678,
0.005557192489504814,
-0.009851625189185143,
-0.015197008848190308,
-0.0077300406992435455,
-0.05636996403336525,
0.035006169229745865,
-0.014883331954479218,
-0.04633728414773941,
0.006681603379547596,
-0.018743019551038742,
0.02617558091878891,
0.07559908181428909,
0.04114597290754318,
-0.018871929496526718,
0.023217372596263885,
-0.028411762788891792,
-0.03792601451277733,
0.012614943087100983,
-0.13678070902824402,
-0.04477298632264137,
-0.01929762028157711,
-0.0589246042072773,
0.025842148810625076,
-0.0038258039858192205,
0.1355830729007721,
-0.04805466905236244,
-0.024194879457354546,
-0.025832829996943474,
0.009440042078495026,
-0.03510129451751709,
-0.02200663834810257,
0.08383205533027649,
0.026773877441883087,
-0.0012594586005434394,
-0.008892115205526352,
0.06565286964178085,
-0.047855962067842484,
-0.055096011608839035,
-0.001281389151699841,
0.003279611701145768,
0.014478863216936588,
0.10301905870437622,
-0.09521497040987015,
-0.000803836970590055,
0.019559191539883614,
0.11989922821521759,
-0.030401285737752914,
-0.02377614565193653,
0.014169150032103062,
-0.0239463709294796,
-0.004880351480096579,
0.006369585171341896,
0.07892176508903503,
-0.007205558475106955,
-0.0336620919406414,
-0.05951697379350662,
0.10123851895332336,
-0.030978776514530182,
0.03205203637480736,
-0.07591899484395981,
0.05303208529949188,
0.013648450374603271,
0.011938933283090591,
0.007348858751356602,
-0.010064952075481415,
0.07273352146148682,
-0.056046999990940094,
0.02706136927008629,
0.042768917977809906,
0.056593019515275955,
0.0844273790717125,
-0.02293490245938301,
0.018082572147250175,
-0.012421919032931328,
-0.012257793918251991,
-0.010701793245971203,
0.07753145694732666,
0.005775478668510914,
-0.12376227229833603,
0.024162113666534424,
0.010408277623355389,
0.01978015899658203,
-0.0332796685397625,
-0.022256456315517426,
-8.049179200156382e-34,
0.04053736478090286,
0.039336685091257095,
0.03429650515317917,
-0.02040601335465908,
0.0041610668413341045,
-0.006786633282899857,
-0.025944514200091362,
0.0014056614600121975,
-0.06515411287546158,
0.008152472786605358,
0.02103479392826557,
0.03384042903780937,
-0.020054221153259277,
0.06941988319158554,
-0.02446398325264454,
-0.008719717152416706,
-0.10812290757894516,
0.01338802371174097,
0.06675177812576294,
-0.05364307761192322,
0.07610374689102173,
0.005445646587759256,
-0.044144947081804276,
-0.04349958896636963,
-0.00011493525380501524,
-0.015059775672852993,
0.022801896557211876,
-0.020281333476305008,
0.07650253176689148,
0.02449214644730091,
0.00433028768748045,
0.10232552886009216,
-0.05570326745510101,
0.08822210878133774,
0.021869216114282608,
0.07171197980642319,
-0.07197548449039459,
0.014748566783964634,
-0.009660414420068264,
0.07354043424129486,
0.005277896765619516,
-0.006835691165179014,
0.08272823691368103,
-0.019074156880378723,
-0.08709557354450226,
-0.1264292150735855,
-0.041844479739665985,
-0.0461522676050663,
-0.10881754010915756,
-0.05815114453434944,
-0.03214245289564133,
0.014429544098675251,
-0.027411065995693207,
-0.10526470094919205,
-0.03339183330535889,
0.0831095427274704,
-0.008669368922710419,
0.05686972290277481,
-0.01002717949450016,
0.03871632739901543,
0.053134605288505554,
0.00007240652485052124,
0.00194094842299819,
0.10737167298793793,
-0.021826837211847305,
0.059137117117643356,
-0.03962709382176399,
0.10232578217983246,
0.07386460155248642,
-0.13363689184188843,
0.011800122447311878,
0.034858960658311844,
-0.03098510019481182,
-0.08393150568008423,
0.05328192189335823,
0.01586025208234787,
0.10926619917154312,
-0.06832422316074371,
-0.0018418923718854785,
0.004010214004665613,
0.005554501432925463,
-0.06713471561670303,
-0.07049740105867386,
0.018669843673706055,
-0.06255669891834259,
-0.032419126480817795,
-0.03802819922566414,
-0.08102931082248688,
0.011008942499756813,
0.05523642897605896,
-0.0700124055147171,
-0.006740038748830557,
-0.007549027446657419,
0.0874798595905304,
0.03834924101829529,
-1.4336974706166609e-33,
0.027667442336678505,
-0.046828482300043106,
-0.04018649831414223,
-0.033380888402462006,
0.0628795325756073,
0.011085488833487034,
-0.02432321384549141,
-0.029400235041975975,
-0.006469095125794411,
0.004349511582404375,
-0.11280597746372223,
0.016621429473161697,
0.08057869225740433,
-0.018541600555181503,
0.0006668552523478866,
-0.0853298231959343,
0.03192213550209999,
0.020686505362391472,
-0.0024070595391094685,
-0.03198017552495003,
0.015631267800927162,
0.03253757208585739,
-0.06153836473822594,
0.009855532087385654,
0.043910037726163864,
0.047613225877285004,
0.0007000093464739621,
0.04987621307373047,
0.003360721981152892,
0.11388394236564636,
-0.07117364555597305,
-0.1075788140296936,
0.014823410660028458,
-0.03204771503806114,
-0.013223305344581604,
0.11420564353466034,
-0.00038125202991068363,
0.00017305718210991472,
-0.0660252496600151,
0.09875833243131638,
0.014422759413719177,
-0.09674707055091858,
-0.029546400532126427,
0.011256644502282143,
0.014721909537911415,
-0.03612807020545006,
-0.013457927852869034,
0.008469042368233204,
0.05580897629261017,
0.05176384001970291,
0.005500982515513897,
0.058542732149362564,
-0.06595324724912643,
-0.013248341158032417,
-0.05507703125476837,
-0.05472608655691147,
0.013725858181715012,
0.07589361071586609,
0.053392019122838974,
-0.010157144628465176,
-0.0135184982791543,
0.04988807067275047,
-0.036083925515413284,
0.06618248671293259,
-0.005660350434482098,
-0.019659196957945824,
-0.03346438333392143,
0.01876472681760788,
0.0978730320930481,
-0.017545197159051895,
0.0799027755856514,
0.0358845554292202,
-0.06394791603088379,
-0.0928594321012497,
0.10843661427497864,
0.019399380311369896,
-0.049996618181467056,
-0.04833972081542015,
-0.05064402148127556,
-0.012528832070529461,
0.007121171336621046,
0.0396995022892952,
0.06399976462125778,
0.03191813454031944,
0.009455668739974499,
-0.00047413705033250153,
0.06762566417455673,
0.11681437492370605,
-0.021340660750865936,
-0.024210290983319283,
0.04347231984138489,
-0.01612468995153904,
0.030251573771238327,
-0.04522094875574112,
-0.08158306032419205,
-4.398241415515258e-8,
-0.03533658757805824,
0.00558234890922904,
0.05591841787099838,
0.007974455133080482,
0.04409318044781685,
0.007368558552116156,
-0.00888586975634098,
-0.035498663783073425,
-0.051729410886764526,
-0.007083195261657238,
0.06353478133678436,
0.07063660025596619,
0.03738556057214737,
-0.09367626905441284,
-0.03458278253674507,
0.05273841693997383,
-0.0030148779042065144,
-0.03870633617043495,
-0.032907482236623764,
0.020508401095867157,
-0.034105751663446426,
-0.06265716999769211,
0.0450262613594532,
-0.07773998379707336,
-0.08969805389642715,
-0.017011798918247223,
-0.02194131538271904,
0.04105854779481888,
0.07293712347745895,
0.021565690636634827,
-0.034103792160749435,
-0.005648524034768343,
0.003617670387029648,
-0.058314234018325806,
-0.016375672072172165,
0.04873478040099144,
0.024490615352988243,
-0.05381650850176811,
0.08027976006269455,
-0.09930779039859772,
-0.0803818553686142,
0.09486833214759827,
0.00236401055008173,
0.03290071338415146,
0.02138167805969715,
0.0285140722990036,
-0.06158126890659332,
-0.07869970798492432,
0.016966165974736214,
-0.020957695320248604,
0.06539705395698547,
-0.048184849321842194,
0.01773977279663086,
0.026349660009145737,
0.10500287264585495,
-0.018462732434272766,
-0.021904191002249718,
-0.027502408251166344,
0.06173839792609215,
0.06039629504084587,
-0.00461215665563941,
-0.00016607905854471028,
-0.0772499144077301,
-0.031132107600569725
] | 0 | 0 | 0 | 0 |
|
Few-Class Arena: A Benchmark for Efficient Selection of Vision Models and Dataset Difficulty Measurement
|
https://openreview.net/forum?id=2ET561DyPe
|
[
"Bryan Bo Cao",
"Lawrence O'Gorman",
"Michael Coss",
"Shubham Jain"
] |
Poster
|
We propose Few-Class Arena (FCA), as a unified benchmark with focus on testing efficient image classification models for few classes. A wide variety of benchmark datasets with many classes (80-1000) have been created to assist Computer Vision architectural evolution. An increasing number of vision models are evaluated with these many-class datasets. However, real-world applications often involve substantially fewer classes of interest (2-10). This gap between many and few classes makes it difficult to predict performance of the few-class applications using models trained on the available many-class datasets. To date, little has been offered to evaluate models in this Few-Class Regime. We conduct a systematic evaluation of the ResNet family trained on ImageNet subsets from 2 to 1000 classes, and test a wide spectrum of Convolutional Neural Networks and Transformer architectures over ten datasets by using our newly proposed FCA tool. Furthermore, to aid an up-front assessment of dataset difficulty and a more efficient selection of models, we incorporate a difficulty measure as a function of class similarity. FCA offers a new tool for efficient machine learning in the Few-Class Regime, with goals ranging from a new efficient class similarity proposal, to lightweight model architecture design, to a new scaling law. FCA is user-friendly and can be easily extended to new models and datasets, facilitating future research work. Our benchmark is available at https://github.com/bryanbocao/fca.
|
Few-Class, lightweight, small neural network, benchmark, scaling law, image similarity, convolutional neural network, CNN, transformer
|
We propose a Few-Class neural network benchmark for model selection with deep analyses.
| 11,666 |
2411.01099
|
[
-0.03465628623962402,
-0.10521043092012405,
-0.00611832132562995,
0.021015550941228867,
0.11734805256128311,
-0.007878169417381287,
-0.03196634352207184,
-0.002911889925599098,
-0.02101750671863556,
-0.08404090255498886,
-0.06522364914417267,
-0.13250210881233215,
0.0029826508834958076,
0.06973396986722946,
-0.07040740549564362,
-0.08667242527008057,
0.08378250896930695,
0.04249099642038345,
-0.117300845682621,
0.007346864324063063,
-0.012151738628745079,
-0.008769594132900238,
0.015534928999841213,
-0.028435777872800827,
0.0011504774447530508,
0.005314890295267105,
0.03465123847126961,
-0.02601315639913082,
-0.028188319876790047,
-0.05108330398797989,
-0.042656414210796356,
-0.0025362572632730007,
-0.06838539242744446,
0.10102364420890808,
0.004332824144512415,
0.023310445249080658,
0.003487837500870228,
0.04330512881278992,
0.016342468559741974,
-0.051850590854883194,
-0.014742420054972172,
0.004714059177786112,
0.057018280029296875,
0.009522153995931149,
0.08274815231561661,
0.01932070031762123,
-0.013810190372169018,
-0.08722043037414551,
0.0657232329249382,
-0.018474163487553596,
-0.013242372311651707,
-0.019185857847332954,
-0.06649376451969147,
0.041298773139715195,
-0.06482639908790588,
0.052698828279972076,
0.05416378378868103,
-0.010612464509904385,
-0.012639278545975685,
-0.0011303296778351068,
0.005996124818921089,
-0.1056997999548912,
0.01647784933447838,
0.04044336453080177,
-0.08983194828033447,
0.006001756526529789,
0.05205809324979782,
-0.019692305475473404,
0.012540671974420547,
-0.05400899797677994,
-0.036841243505477905,
0.10813047736883163,
-0.03325428441166878,
0.0853762999176979,
0.04061465337872505,
0.024750055745244026,
0.11953054368495941,
0.006362696178257465,
0.058931317180395126,
-0.11814195662736893,
0.0144793801009655,
0.010014848783612251,
-0.003614729270339012,
0.03201226517558098,
0.13767553865909576,
0.023081298917531967,
-0.03255084156990051,
0.03584934398531914,
-0.0161738358438015,
0.014315244741737843,
-0.05026938393712044,
0.019594846293330193,
0.07884154468774796,
-0.026507647708058357,
0.05259831249713898,
0.06390631943941116,
0.01892939954996109,
-0.09728290885686874,
-0.07885189354419708,
0.08432971686124802,
-0.0776219591498375,
-0.005022282712161541,
0.05137833207845688,
-0.0305444598197937,
0.01989591121673584,
-0.0137247983366251,
0.0667860209941864,
0.00871364213526249,
0.006280985660851002,
-0.09894491732120514,
0.00044343428453430533,
-0.013101582415401936,
-0.014324696734547615,
-0.02967780828475952,
0.04635123163461685,
0.021562807261943817,
-0.04276903718709946,
-0.033214982599020004,
0.011402427218854427,
0.055574096739292145,
-0.04754132032394409,
0.015391443856060505,
0.019499879330396652,
-0.026798103004693985,
0.06190142408013344,
0.018714996054768562,
-0.1096993088722229,
1.7432935707709694e-33,
0.026173627004027367,
-0.010575133375823498,
0.03036341816186905,
0.01577700674533844,
0.08628930896520615,
-0.050174612551927567,
0.023366188630461693,
0.09894029051065445,
-0.04445507377386093,
0.0232093445956707,
-0.06374287605285645,
-0.0028581880033016205,
-0.05399128049612045,
0.04522969201207161,
0.08174176514148712,
0.028457632288336754,
-0.05491351708769798,
0.017722055315971375,
-0.07140658050775528,
0.024686938151717186,
-0.023326601833105087,
-0.013709235936403275,
0.06736370921134949,
-0.06810687482357025,
-0.010797014459967613,
-0.0038088643923401833,
0.020613867789506912,
0.01143873855471611,
-0.02342219464480877,
0.0024017449468374252,
-0.04329177364706993,
0.0029875151813030243,
0.0008460420067422092,
0.04579621180891991,
0.05737494304776192,
-0.030716199427843094,
0.021328004077076912,
0.006310353521257639,
-0.016828123480081558,
-0.03811229020357132,
-0.04951423779129982,
0.08435248583555222,
0.010574865154922009,
-0.022645793855190277,
-0.037297673523426056,
0.0034956445451825857,
0.05015622079372406,
0.01975362002849579,
-0.03813512995839119,
-0.019613077864050865,
0.05198156088590622,
0.03630722686648369,
-0.058370672166347504,
-0.053424134850502014,
-0.034241143614053726,
-0.04143508896231651,
0.03255779668688774,
0.02901627868413925,
0.009780751541256905,
0.029458610340952873,
-0.01946309022605419,
0.022236982360482216,
-0.02562899701297283,
0.09019305557012558,
-0.02629627287387848,
0.0038677058182656765,
0.0037252656184136868,
0.07130562514066696,
-0.011832100339233875,
0.0034762669820338488,
0.031712982803583145,
-0.010901989415287971,
-0.07387750595808029,
-0.11802292615175247,
0.18027250468730927,
0.04045262560248375,
0.10372092574834824,
-0.06122210994362831,
-0.07243662327528,
0.03443698585033417,
0.00765350041911006,
0.015306252054870129,
0.008493641391396523,
-0.009364472702145576,
-0.07688542455434799,
0.013386333361268044,
0.029916061088442802,
-0.048105236142873764,
-0.012289916165173054,
-0.03015120141208172,
-0.07972271740436554,
0.02539844438433647,
-0.052775606513023376,
-0.007893565110862255,
0.004192705266177654,
-1.1379837750745714e-33,
-0.0040168859995901585,
0.05664339289069176,
-0.020275365561246872,
0.07763560861349106,
0.007420916575938463,
-0.04736810550093651,
0.051194868981838226,
0.034928955137729645,
-0.03243723139166832,
-0.03606486693024635,
0.05086448788642883,
-0.020828966051340103,
-0.02308993600308895,
-0.043598100543022156,
-0.015359977260231972,
-0.035194676369428635,
-0.07008671015501022,
-0.019206233322620392,
-0.0031696511432528496,
0.03454185649752617,
0.03674997389316559,
0.10760529339313507,
-0.024539539590477943,
0.0013202581321820617,
-0.12953750789165497,
0.07360929250717163,
-0.10109782963991165,
0.018791059032082558,
-0.036973804235458374,
-0.06377368420362473,
0.012195366434752941,
0.03707295283675194,
-0.05130591243505478,
-0.034420982003211975,
0.0005330516723915935,
0.08001578599214554,
0.019263647496700287,
-0.018225204199552536,
-0.03285346180200577,
0.11335429549217224,
0.012910798192024231,
0.0031722653657197952,
-0.013819681480526924,
0.05814450979232788,
-0.0454208143055439,
-0.017109420150518417,
0.058480847626924515,
-0.03048667684197426,
-0.0069893281906843185,
0.06539317965507507,
-0.060675766319036484,
-0.0014825607649981976,
-0.09047695249319077,
0.060345668345689774,
0.04752150923013687,
-0.0062552280724048615,
-0.04086272045969963,
0.015706902369856834,
0.051823150366544724,
0.07109671831130981,
-0.050162822008132935,
-0.01878317818045616,
-0.06410294026136398,
0.0370456837117672,
-0.003194924211129546,
0.002057119971141219,
-0.05599461868405342,
0.018058523535728455,
-0.015121396631002426,
0.12102047353982925,
0.0337430015206337,
-0.01774716190993786,
-0.017481138929724693,
-0.005498051643371582,
-0.08793243020772934,
-0.04040738567709923,
0.050398945808410645,
0.010286527685821056,
0.1170300543308258,
-0.005188549868762493,
0.0147376898676157,
-0.07947807013988495,
-0.04680017754435539,
0.03965487331151962,
0.04634808748960495,
0.04995918646454811,
0.09075919538736343,
-0.007677113637328148,
0.1022505834698677,
0.054843880236148834,
-0.08949633687734604,
0.042120933532714844,
0.08260639011859894,
0.016445936635136604,
-0.0008015427738428116,
-3.7578644196401e-8,
-0.044870734214782715,
0.0054010869935154915,
0.013861103914678097,
-0.034304287284612656,
-0.015379086136817932,
-0.026885317638516426,
-0.05588524416089058,
0.16382256150245667,
-0.05109606310725212,
0.011126008816063404,
0.050827205181121826,
0.02138678729534149,
-0.026108307763934135,
0.04203229025006294,
-0.037510309368371964,
0.06268834322690964,
0.020908448845148087,
0.05114831402897835,
0.03982893005013466,
0.008769217878580093,
0.007013915106654167,
-0.00921171810477972,
0.02801649644970894,
-0.01772022806107998,
-0.03449685871601105,
-0.10017954558134079,
-0.031092949211597443,
0.03296778351068497,
0.04920501634478569,
0.006217990070581436,
-0.003237230936065316,
0.031388059258461,
0.0157020166516304,
-0.10637681186199188,
0.03364577889442444,
0.08089115470647812,
-0.06367923319339752,
0.014370135962963104,
-0.041484635323286057,
-0.029866771772503853,
-0.012208535335958004,
-0.0009198064799420536,
-0.030775433406233788,
0.011179843917489052,
0.002033843891695142,
-0.042424820363521576,
0.040650494396686554,
-0.08799056708812714,
0.013102680444717407,
-0.06469570100307465,
0.007789683528244495,
-0.009436548687517643,
-0.06653288006782532,
0.058403972536325455,
-0.027190977707505226,
-0.02029511146247387,
0.05406683683395386,
-0.13332070410251617,
0.023089656606316566,
0.06740008294582367,
0.01608344167470932,
0.05102554336190224,
-0.07929275929927826,
0.008360644802451134
] |
https://github.com/fewclassarena/fca
| 0 | 0 | 0 | 0 |
Transformers Struggle to Learn to Search
|
https://openreview.net/forum?id=9cQB1Hwrtw
|
[
"Abulhair Saparov",
"Srushti Ajay Pawar",
"Shreyas Pimpalgaonkar",
"Nitish Joshi",
"Richard Yuanzhe Pang",
"Vishakh Padmakumar",
"Mehran Kazemi",
"Najoung Kim",
"He He"
] |
Poster
|
Search is an ability foundational in many important tasks, and recent studies have shown that large language models (LLMs) struggle to perform search robustly. It is unknown whether this inability is due to a lack of data, insufficient model parameters, or fundamental limitations of the transformer architecture. In this work, we use the foundational graph connectivity problem as a testbed to generate effectively limitless high-coverage data to train small transformers and test whether they can learn to perform search. We find that, when given the right training distribution, the transformer is able to learn to search.
We analyze the algorithm that the transformer has learned through a novel mechanistic interpretability technique that enables us to extract the computation graph from the trained model. We find that for each vertex in the input graph, transformers compute the set of vertices reachable from that vertex. Each layer then progressively expands these sets, allowing the model to search over a number of vertices exponential in the number of layers.
However, we find that as the input graph size increases, the transformer has greater difficulty in learning the task. This difficulty is not resolved even as the number of parameters is increased, suggesting that increasing model scale will not lead to robust search abilities. We also find that performing search in-context (i.e., chain-of-thought) does not resolve this inability to learn to search on larger graphs.
|
search, reasoning, transformers, scaling laws, mechanistic interpretability, circuit analysis
| null | 11,660 |
2412.04703
|
[
-0.0819910392165184,
0.005522984080016613,
-0.012533293105661869,
0.048604100942611694,
0.04263946786522865,
-0.02250715158879757,
-0.0622548833489418,
0.03903369605541229,
-0.030894054099917412,
-0.029743246734142303,
-0.06988716125488281,
-0.010987000539898872,
0.05433974787592888,
0.08095567673444748,
-0.0164295993745327,
0.010791395790874958,
0.04089103639125824,
0.055474553257226944,
-0.08846104890108109,
-0.10724593698978424,
0.04295450448989868,
0.04441145062446594,
-0.01656980998814106,
-0.029970398172736168,
0.036269258707761765,
-0.02684919349849224,
-0.04659700393676758,
-0.08796809613704681,
0.03476228937506676,
-0.014792296104133129,
0.047257259488105774,
-0.03427160158753395,
-0.0799407809972763,
0.11345912516117096,
-0.010982965119183064,
0.025920556858181953,
-0.06443849951028824,
-0.007178352214396,
0.03165864944458008,
-0.023237550631165504,
0.010951567441225052,
-0.011542439460754395,
0.029747888445854187,
-0.001982747809961438,
0.05763482302427292,
0.021669749170541763,
0.026174470782279968,
-0.04137050732970238,
0.032783254981040955,
-0.05630423128604889,
-0.07650456577539444,
-0.06252891570329666,
-0.03179823234677315,
0.07519948482513428,
-0.034865427762269974,
-0.005187972914427519,
0.07371218502521515,
-0.027115561068058014,
0.017695927992463112,
-0.02468893863260746,
-0.03383225202560425,
-0.07880625873804092,
-0.014297815039753914,
-0.034503791481256485,
-0.08728715032339096,
-0.0007376679568551481,
0.0798133835196495,
0.02122892253100872,
-0.019061392173171043,
-0.021159950643777847,
0.0016583878314122558,
0.059186555445194244,
-0.062463823705911636,
0.04536548629403114,
0.12004876136779785,
0.043857116252183914,
0.06744524836540222,
-0.04978696629405022,
0.05847733095288277,
-0.02209887094795704,
0.0050070397555828094,
-0.01841270737349987,
0.013838338665664196,
0.063987135887146,
0.08372020721435547,
-0.012129603885114193,
-0.008427266962826252,
-0.035438984632492065,
0.035314157605171204,
0.023064741864800453,
-0.03732084110379219,
-0.06715718656778336,
-0.03856854513287544,
0.011861645616590977,
0.031522803008556366,
0.0326363779604435,
0.02746140956878662,
-0.04970389977097511,
-0.0471251904964447,
0.08085786551237106,
-0.0661110058426857,
0.026574905961751938,
0.12629567086696625,
-0.04173405468463898,
-0.07131136208772659,
0.0035147147718816996,
-0.0174090676009655,
0.052814170718193054,
-0.019436663016676903,
-0.09219630062580109,
0.014511974528431892,
0.0417167954146862,
0.04076060652732849,
-0.007948260754346848,
0.04090036824345589,
0.00261601316742599,
-0.004449610598385334,
-0.03785507380962372,
-0.09828481078147888,
0.11062809824943542,
-0.0022214872296899557,
0.06678973138332367,
-0.039581675082445145,
0.046039510518312454,
0.017437521368265152,
0.037664227187633514,
-0.04051724076271057,
1.2139035368999096e-33,
0.036968428641557693,
0.09289585798978806,
0.027851330116391182,
0.05203862115740776,
0.05355400592088699,
-0.005154935643076897,
0.058887943625450134,
0.08657829463481903,
0.004406019113957882,
0.04229523986577988,
-0.0938916727900505,
0.06467951834201813,
-0.03079410456120968,
0.032065361738204956,
0.04796295613050461,
-0.0051538217812776566,
0.03238474205136299,
-0.01836784929037094,
-0.07954009622335434,
-0.050532158464193344,
0.09017257392406464,
-0.060247305780649185,
0.0458773709833622,
-0.09149053692817688,
-0.024113986641168594,
-0.018327368423342705,
0.03681139275431633,
-0.027580393478274345,
-0.017997926101088524,
-0.008873976767063141,
-0.14772774279117584,
0.016915084794163704,
0.027008026838302612,
0.06489508599042892,
0.04558512941002846,
-0.057495322078466415,
-0.003456937614828348,
-0.03541941195726395,
-0.018273452296853065,
-0.09016153961420059,
-0.00681472709402442,
0.00905218068510294,
0.010621817782521248,
-0.03063667193055153,
-0.013579238206148148,
0.021606579422950745,
0.007689641788601875,
-0.0785079300403595,
-0.037601981312036514,
-0.05230154097080231,
0.046879928559064865,
0.05494096130132675,
-0.06735732406377792,
-0.005518071819096804,
0.12300854921340942,
0.0423237681388855,
0.07099965959787369,
0.0266613457351923,
0.024665389209985733,
0.061474017798900604,
-0.009111444465816021,
0.0045105391182005405,
0.007320322096347809,
0.12120918929576874,
0.056375473737716675,
-0.033275388181209564,
-0.03488786518573761,
0.054007936269044876,
0.10337066650390625,
-0.02334778942167759,
-0.03915952891111374,
-0.03402116894721985,
0.017367150634527206,
-0.0804135650396347,
0.03733163699507713,
-0.05976605415344238,
0.05610924959182739,
-0.03422330692410469,
0.00005870581298950128,
-0.029462343081831932,
-0.016066135838627815,
-0.06109561026096344,
0.06000782549381256,
0.004273678176105022,
0.03376676142215729,
-0.04015715420246124,
-0.007384209427982569,
-0.06706355512142181,
-0.003941221162676811,
-0.04886973276734352,
-0.05990901216864586,
-0.010837815701961517,
-0.02105472795665264,
-0.028129927814006805,
0.0238654762506485,
-5.55215966182862e-34,
-0.059936970472335815,
0.01546421181410551,
-0.030873671174049377,
0.07788029313087463,
-0.01386856846511364,
-0.09099699556827545,
0.023404989391565323,
0.03031746670603752,
-0.028411932289600372,
0.031492237001657486,
-0.035767003893852234,
-0.04731420427560806,
0.060000043362379074,
-0.05879054591059685,
0.02973194420337677,
-0.004745326470583677,
-0.017321940511465073,
-0.007211212534457445,
0.032685719430446625,
0.05798779055476189,
-0.017786717042326927,
0.13892553746700287,
-0.17295312881469727,
-0.019601548090577126,
-0.06633130460977554,
0.04330733045935631,
-0.1113273948431015,
0.04254646226763725,
0.05776045843958855,
0.062424950301647186,
-0.05344632267951965,
0.0009640951175242662,
-0.03840334713459015,
0.005956258159130812,
-0.04055103287100792,
0.02659808285534382,
0.012769412249326706,
0.02451251819729805,
-0.03278188034892082,
0.09131468832492828,
0.0077292039059102535,
-0.015780556946992874,
-0.07921659201383591,
0.0201436597853899,
0.017913617193698883,
-0.001882009906694293,
-0.014328599907457829,
-0.035273488610982895,
-0.016000784933567047,
0.012497877702116966,
0.015736645087599754,
0.05008794739842415,
-0.032651763409376144,
-0.06374029815196991,
-0.024976329877972603,
-0.07602813839912415,
0.011835156008601189,
0.03016963042318821,
0.043598175048828125,
-0.0413629449903965,
-0.044141292572021484,
-0.018454590812325478,
0.02458040416240692,
-0.022811735048890114,
-0.026653945446014404,
-0.024807142093777657,
-0.04655032232403755,
0.003690747544169426,
0.05320800095796585,
0.06675422936677933,
0.05764711648225784,
-0.0276136826723814,
0.04056885465979576,
-0.045990366488695145,
0.04619460552930832,
0.0005780116189271212,
-0.03554312139749527,
-0.011930896900594234,
0.011239629238843918,
-0.041326917707920074,
-0.04323049262166023,
0.04911777749657631,
-0.014391223900020123,
0.05564893037080765,
0.06424696743488312,
0.04587310180068016,
0.05330570787191391,
0.06802404671907425,
0.06673268228769302,
0.0066001638770103455,
-0.04434394836425781,
0.03137590363621712,
-0.06135520711541176,
0.04076012969017029,
-0.04892789199948311,
-4.684527254994464e-8,
-0.1297723650932312,
0.055838558822870255,
-0.05177657678723335,
-0.02005436085164547,
0.0868336483836174,
-0.04277154058218002,
0.06771643459796906,
0.13967230916023254,
-0.0597967803478241,
0.06749621778726578,
0.09651663899421692,
-0.036841485649347305,
-0.00436017569154501,
0.04238957539200783,
0.0031387032940983772,
0.08393122255802155,
0.033133406192064285,
-0.0036108812782913446,
-0.03217627480626106,
-0.006894560996443033,
-0.015081665478646755,
0.04313027486205101,
-0.019463147968053818,
0.03255179524421692,
0.06192563846707344,
-0.0902566909790039,
-0.07881273329257965,
0.02026437222957611,
-0.01331740990281105,
0.08897607773542404,
-0.041231416165828705,
-0.011851922608911991,
0.017421698197722435,
0.0011566902976483107,
0.003099519293755293,
0.13708579540252686,
0.03141961246728897,
-0.05836710333824158,
-0.03788077458739281,
-0.034714460372924805,
0.07689949870109558,
0.12052090466022491,
-0.03437281399965286,
0.05639967694878578,
-0.01482151634991169,
-0.00174060242716223,
0.03861548379063606,
-0.06946372240781784,
0.005446267314255238,
-0.026964552700519562,
0.0020494686905294657,
-0.04742613434791565,
-0.07702595740556717,
-0.004564485512673855,
0.02413352020084858,
0.006181003525853157,
-0.005540261045098305,
-0.08448973298072815,
-0.03292842209339142,
0.09416152536869049,
-0.021769646555185318,
0.01701512187719345,
-0.03942305967211723,
-0.0055846902541816235
] |
https://github.com/asaparov/learning_to_search
| 5 | 0 | 0 | 0 |
Is uniform expressivity too restrictive? Towards efficient expressivity of GNNs
|
https://openreview.net/forum?id=lsvGqR6OTf
|
[
"Sammy Khalife",
"Josué Tonelli-Cueto"
] |
Poster
|
Uniform expressivity guarantees that a Graph Neural Network (GNN) can express a query without the parameters depending on the size of the input graphs. This property is desirable in applications in order to have number of trainable parameters that is independent of the size of the input graphs. Uniform expressivity of the two variable guarded fragment (GC2) of first order logic is a well-celebrated result for Rectified Linear Unit (ReLU) GNNs [Barcelo &. Al, 2020]. In this article, we prove that uniform expressivity of GC2 queries is not possible for GNNs with a wide class of Pfaffian activation functions (including the sigmoid and $\tanh$), answering a question formulated by [Grohe, 2021]. We also show that despite these limitations, many of those GNNs can still efficiently express GC2 queries in a way that the number of parameters remains logarithmic on the maximal degree of the input graphs. Furthermore, we demonstrate that a log-log dependency on the degree is achievable for a certain choice of activation function. This shows that uniform expressivity can be successfully relaxed by covering large graphs appearing in practical applications. Our experiments illustrates that our theoretical estimates hold in practice.
|
Graph Neural Networks, Expressivity, Efficiency, Activation Function, Queries, Logic
| null | 11,655 | null |
[
-0.05132853239774704,
-0.09402234107255936,
0.016784250736236572,
0.03818103298544884,
-0.0027051977813243866,
0.0064674923196434975,
0.07038275897502899,
-0.045545630156993866,
-0.0028592783492058516,
0.009655237197875977,
0.016520226374268532,
0.009869681671261787,
0.00917871855199337,
0.05540890246629715,
-0.01075713150203228,
-0.035035084933042526,
0.07860452681779861,
0.053123097866773605,
-0.10755600035190582,
0.012735076248645782,
0.021879546344280243,
-0.03606799989938736,
0.007514696102589369,
0.0059557161293923855,
0.015549714677035809,
-0.07222309708595276,
-0.05305915325880051,
-0.0312986820936203,
0.043021127581596375,
-0.004912173375487328,
-0.017633497714996338,
-0.004868121352046728,
0.007893921807408333,
0.09178754687309265,
-0.05365956574678421,
0.008874058723449707,
-0.031818460673093796,
-0.07810754328966141,
0.0313136950135231,
-0.06297753006219864,
0.008599168621003628,
-0.008941568434238434,
-0.010575351305305958,
0.05093930661678314,
0.1155458316206932,
0.032989662140607834,
0.02840789593756199,
0.006422715727239847,
-0.03137553110718727,
0.013631297275424004,
0.02026534453034401,
0.015606021508574486,
-0.07285985350608826,
0.022093314677476883,
0.044164594262838364,
-0.015845036134123802,
-0.058154232800006866,
0.02911970391869545,
-0.01733078435063362,
0.053736139088869095,
-0.03175367787480354,
-0.06335041671991348,
-0.027914609760046005,
-0.021588357165455818,
0.01213513221591711,
-0.03939024731516838,
0.03638916835188866,
-0.009005794301629066,
0.036988504230976105,
0.10834616422653198,
0.026662830263376236,
0.05169272795319557,
-0.07335708290338516,
-0.029871735721826553,
0.03381076082587242,
0.08564028143882751,
0.017215505242347717,
0.0012632034486159682,
-0.0028184023685753345,
-0.04597514867782593,
0.0408683717250824,
-0.04830760508775711,
0.05532744526863098,
0.026214512065052986,
0.07435919344425201,
-0.08936326950788498,
-0.038447361439466476,
0.017048321664333344,
0.09745848178863525,
-0.058602746576070786,
-0.06448768079280853,
0.02687019482254982,
-0.011540801264345646,
-0.03885862976312637,
0.05160225182771683,
0.004306199494749308,
-0.045471154153347015,
-0.010238801129162312,
-0.05568350851535797,
0.08588138967752457,
-0.040814220905303955,
0.041118357330560684,
0.04503558576107025,
-0.05611267313361168,
-0.022379698231816292,
0.044120579957962036,
0.0022111614234745502,
0.005298434291034937,
0.03173034265637398,
-0.10353189706802368,
-0.03863411024212837,
0.08398135751485825,
0.04167957976460457,
0.007644840516149998,
0.018498441204428673,
0.0073530045337975025,
0.020330771803855896,
-0.04265046492218971,
0.04833803325891495,
0.010538598522543907,
-0.0335700549185276,
0.04830151051282883,
-0.05774854123592377,
0.02291363663971424,
-0.027587702497839928,
-0.006908725015819073,
-0.007270331960171461,
4.023972832885994e-33,
0.05891026183962822,
0.09594056755304337,
0.06220144033432007,
-0.04380512982606888,
0.0749102234840393,
0.08095770329236984,
-0.0037479589227586985,
-0.005628949962556362,
0.009663174860179424,
0.02255132794380188,
-0.1476280838251114,
0.031917914748191833,
-0.011280753649771214,
0.062179893255233765,
0.038883037865161896,
0.051352448761463165,
0.04031158238649368,
0.019663479179143906,
0.011105330660939217,
-0.05944359302520752,
0.0728643462061882,
-0.023786647245287895,
-0.0030479198321700096,
-0.014489663764834404,
0.015885168686509132,
-0.012189850211143494,
-0.004742381628602743,
0.00817190296947956,
-0.06680026650428772,
-0.024919675663113594,
-0.09758612513542175,
-0.030448731034994125,
-0.015444668009877205,
0.06327565014362335,
0.010042794980108738,
-0.024008208885788918,
-0.030380545184016228,
-0.019309842959046364,
0.06016743928194046,
-0.08646758645772934,
-0.024401506409049034,
0.032532237470149994,
0.00933354627341032,
-0.010850274935364723,
-0.024503832682967186,
-0.0666588619351387,
-0.05546226352453232,
0.03980673477053642,
0.016710320487618446,
-0.06466514617204666,
0.025787629187107086,
0.04204519838094711,
-0.06052953377366066,
-0.0630417913198471,
0.07644813507795334,
-0.06834335625171661,
0.07185385376214981,
0.10061303526163101,
0.030407138168811798,
0.1271069347858429,
-0.014941603876650333,
-0.005368763115257025,
-0.01848759688436985,
0.012982706539332867,
-0.013499530032277107,
0.06807456910610199,
-0.1286507099866867,
0.058450546115636826,
0.1263948529958725,
0.018472731113433838,
0.012029112316668034,
0.031847357749938965,
-0.03176077455282211,
-0.10083921998739243,
0.07014106959104538,
0.004104445688426495,
0.022094110026955605,
-0.10471618920564651,
-0.001249179127626121,
0.031799521297216415,
-0.06242533400654793,
0.017124632373452187,
-0.028721299022436142,
0.04233832284808159,
-0.09704669564962387,
-0.09787258505821228,
0.030217228457331657,
-0.005821129307150841,
-0.003498245496302843,
-0.03430300951004028,
-0.04035656154155731,
-0.02964986301958561,
0.026473183184862137,
-0.007562636397778988,
0.020463373512029648,
-3.400246714132057e-33,
-0.029547888785600662,
0.07090388983488083,
-0.05988023430109024,
0.03488273173570633,
-0.06845887005329132,
-0.044373515993356705,
0.009638067334890366,
-0.0684291422367096,
-0.04582871124148369,
0.001989471260458231,
0.026036838069558144,
-0.0437968447804451,
0.044081248342990875,
-0.012649810872972012,
0.044696658849716187,
-0.001274762791581452,
-0.10821309685707092,
0.026640741154551506,
0.02583237923681736,
-0.007885339669883251,
0.06484882533550262,
0.12656071782112122,
-0.13276374340057373,
0.05848770588636398,
0.018416395410895348,
0.011523067019879818,
-0.08014976978302002,
0.06694424897432327,
0.00043516859295777977,
0.06066128984093666,
-0.09748893231153488,
-0.08246681094169617,
-0.11361255496740341,
-0.02350936457514763,
0.09253745526075363,
-0.05038518086075783,
0.060681141912937164,
0.05838007479906082,
-0.07345470786094666,
0.06267086416482925,
0.04446199908852577,
-0.025590544566512108,
-0.028648436069488525,
0.08293896168470383,
0.07367558032274246,
-0.01626467891037464,
-0.024035600945353508,
-0.043510615825653076,
-0.07128726691007614,
0.011580834165215492,
-0.01661221869289875,
0.0032594860531389713,
-0.02556486986577511,
0.02076132781803608,
-0.029491854831576347,
0.018479924649000168,
-0.05079459398984909,
0.07382083684206009,
0.022150211036205292,
-0.012441628612577915,
-0.023263167589902878,
-0.0782681554555893,
0.003428054042160511,
-0.018895328044891357,
-0.00883130356669426,
-0.008220462128520012,
0.016185810789465904,
0.055701225996017456,
0.04437077417969704,
0.04446769505739212,
0.003278350457549095,
-0.07617215812206268,
0.014435580931603909,
0.04761296510696411,
-0.04891962185502052,
0.015887295827269554,
0.04316548630595207,
-0.01961299031972885,
-0.03883085027337074,
0.025572028011083603,
0.02464885637164116,
0.02445848658680916,
0.03141212835907936,
0.04742608964443207,
0.06655816733837128,
-0.03350012004375458,
0.029691999778151512,
0.03283286094665527,
0.06143227219581604,
0.027396637946367264,
-0.018534863367676735,
0.016856662929058075,
-0.047083839774131775,
0.05798356235027313,
-0.014501954428851604,
-4.641546880179703e-8,
-0.12731806933879852,
0.06060060113668442,
-0.020294662564992905,
0.02730579301714897,
0.1243382915854454,
-0.028104525059461594,
0.025245927274227142,
0.054974671453237534,
-0.047880932688713074,
0.005195959005504847,
0.07986965775489807,
0.04208514094352722,
0.0012931723613291979,
-0.10455683618783951,
-0.028092656284570694,
0.058088261634111404,
-0.002842646325007081,
-0.10025905817747116,
0.011986824683845043,
-0.013462166301906109,
-0.04950582981109619,
0.05778024345636368,
-0.05799547955393791,
0.0039564766921103,
-0.07315564155578613,
-0.09756948798894882,
-0.05686115473508835,
0.05691177025437355,
-0.028911277651786804,
0.1266653835773468,
0.03088078461587429,
0.002886140253394842,
0.12500624358654022,
-0.04394884034991264,
0.02515948936343193,
0.07446257770061493,
-0.0034070180263370275,
0.00976557657122612,
-0.03930417448282242,
-0.08944655954837799,
-0.014930253848433495,
0.046585727483034134,
0.006602450739592314,
0.00027933254023082554,
0.008537200279533863,
-0.07552189379930496,
0.048887357115745544,
-0.08472360670566559,
0.05135379359126091,
0.032244518399238586,
0.029022496193647385,
-0.03292849287390709,
-0.0694417878985405,
0.02646530792117119,
-0.03758002817630768,
-0.041711919009685516,
-0.05612936615943909,
-0.07652401179075241,
-0.01930311881005764,
0.043807633221149445,
0.0172727033495903,
0.06919506937265396,
0.01884264498949051,
-0.04708002507686615
] | 0 | 0 | 0 | 0 |
|
Training Neural Networks as Recognizers of Formal Languages
|
https://openreview.net/forum?id=aWLQTbfFgV
|
[
"Alexandra Butoi",
"Ghazal Khalighinejad",
"Anej Svete",
"Josef Valvoda",
"Ryan Cotterell",
"Brian DuSell"
] |
Poster
|
Characterizing the computational power of neural network architectures in terms of formal language theory remains a crucial line of research, as it describes lower and upper bounds on the reasoning capabilities of modern AI. However, when empirically testing these bounds, existing work often leaves a discrepancy between experiments and the formal claims they are meant to support. The problem is that formal language theory pertains specifically to recognizers: machines that receive a string as input and classify whether it belongs to a language. On the other hand, it is common instead to evaluate language models on proxy tasks, e.g., language modeling or sequence-to-sequence transduction, that are similar in only an informal sense to the underlying theory. We correct this mismatch by training and evaluating neural networks directly as binary classifiers of strings, using a general method that can be applied to a wide variety of languages. As part of this, we extend an algorithm recently proposed by Snæbjarnarson et al. (2025) for efficient length-controlled sampling of strings from regular languages. We provide results on a variety of languages across the Chomsky hierarchy for three neural architectures: a simple RNN, an LSTM, and a causally-masked transformer. We find that the RNN and LSTM often outperform the transformer, and that auxiliary training objectives such as language modeling can help, although no single objective uniformly improves performance across languages and architectures. Our contributions will facilitate theoretically sound empirical testing of language recognition claims in future work. We have released our datasets as a benchmark called FLaRe (Formal Language Recognition), along with our code.
|
neural network, formal language theory, transformer, rnn, lstm
|
We train neural networks as recognizers (binary classifiers) on formal languages to fix a disconnect with formal results
| 11,654 |
2411.07107
|
[
-0.1124834418296814,
-0.05888139829039574,
0.05960214510560036,
0.027011485770344734,
0.020176373422145844,
0.03104156255722046,
0.004927495494484901,
0.002458933275192976,
-0.0449305884540081,
-0.06146934628486633,
-0.0954156443476677,
-0.06068260222673416,
-0.012788267806172371,
0.05318335071206093,
-0.06778857856988907,
0.011014733463525772,
0.020926179364323616,
0.05680825188755989,
-0.04470640793442726,
-0.03848164528608322,
0.04021069034934044,
-0.011096039786934853,
0.008445973508059978,
-0.002541471039876342,
0.05942494049668312,
-0.020221346989274025,
-0.025095034390687943,
-0.041791412979364395,
-0.03087761253118515,
0.08139683306217194,
0.05012235417962074,
-0.006957177072763443,
-0.017118874937295914,
0.034167319536209106,
0.0025309312622994184,
0.04599407687783241,
-0.041916195303201675,
-0.07127407938241959,
0.043158985674381256,
-0.029384631663560867,
0.00708831287920475,
-0.020785672590136528,
0.05950405076146126,
0.004950880538672209,
0.1435110867023468,
0.07018964737653732,
0.026777924969792366,
-0.011632108129560947,
-0.1003006100654602,
-0.005538118537515402,
-0.059355661273002625,
-0.028816083446145058,
-0.005415042396634817,
0.0993465781211853,
-0.03612643480300903,
-0.040782034397125244,
0.05375069007277489,
0.03935835137963295,
-0.04566791281104088,
0.007475757505744696,
-0.05412742868065834,
-0.10403580963611603,
-0.07435770332813263,
-0.022263178601861,
-0.03658350557088852,
0.015777520835399628,
-0.0012386050075292587,
0.05455360934138298,
-0.022542783990502357,
0.06722062081098557,
-0.04236318916082382,
0.09001514315605164,
-0.10451827198266983,
0.11604674160480499,
0.009609968401491642,
0.03142668679356575,
0.07613494992256165,
0.041493628174066544,
-0.005829656962305307,
-0.10913950949907303,
-0.036509811878204346,
-0.02031586691737175,
0.05798661336302757,
-0.0032156873494386673,
0.048077598214149475,
0.050547193735837936,
-0.049169912934303284,
0.061882585287094116,
0.015098829753696918,
0.02511557750403881,
-0.03058413416147232,
-0.1025298684835434,
0.04353983327746391,
-0.014790425077080727,
0.05045301094651222,
0.05919112637639046,
-0.006069987080991268,
-0.020309124141931534,
-0.019325094297528267,
0.11036361753940582,
-0.012372957542538643,
0.07170300930738449,
0.022239938378334045,
-0.07124064117670059,
0.02015388011932373,
0.06367579847574234,
0.03886467590928078,
-0.014009066857397556,
0.07086647301912308,
-0.09027685970067978,
0.007193157449364662,
0.05850251019001007,
0.02645917423069477,
-0.008115677163004875,
0.0022303650621324778,
0.0029715157579630613,
-0.039626818150281906,
-0.05391034111380577,
0.08361238986253738,
0.0706522986292839,
-0.047950420528650284,
0.04029655084013939,
0.006438514683395624,
0.03313048556447029,
0.06652170419692993,
-0.011596454307436943,
-0.052138131111860275,
-2.5093094236011055e-34,
0.03724884241819382,
0.037508655339479446,
0.05557989329099655,
0.011091773398220539,
0.04808127507567406,
-0.03991909697651863,
0.02357596345245838,
0.014769664034247398,
0.08028686791658401,
-0.031488873064517975,
-0.02961476892232895,
0.0380001924932003,
-0.034199271351099014,
-0.0017554418882355094,
0.04470342397689819,
0.0635424479842186,
0.05532669275999069,
0.030657442286610603,
-0.031501661986112595,
-0.012001266703009605,
0.01707039773464203,
0.00980053935199976,
-0.006902037654072046,
-0.013490344397723675,
-0.025390977039933205,
-0.03369239345192909,
0.00538497744128108,
-0.04305576905608177,
-0.03676033020019531,
-0.003839152166619897,
-0.09427352994680405,
-0.025860842317342758,
0.060927681624889374,
0.06519366055727005,
0.024818576872348785,
-0.06764549762010574,
0.034088023006916046,
-0.03190607205033302,
0.01863831840455532,
-0.016400136053562164,
0.0006552188424393535,
-0.011639665812253952,
0.0516929030418396,
0.1008567214012146,
-0.007436962798237801,
-0.06590031087398529,
-0.0327938050031662,
-0.027015160769224167,
0.002359624020755291,
-0.08705747872591019,
0.08670550584793091,
0.019223347306251526,
-0.03304718807339668,
-0.022289182990789413,
0.05943528935313225,
0.03993496671319008,
0.0036690416745841503,
0.0392821803689003,
0.035540252923965454,
0.11657097190618515,
-0.012230266816914082,
0.07026063650846481,
0.01722712814807892,
0.09208020567893982,
0.0072898645885288715,
0.047373026609420776,
-0.05201699584722519,
-0.002377909142524004,
0.0793977826833725,
-0.034554243087768555,
-0.0426909364759922,
-0.03154360130429268,
-0.059119563549757004,
0.041914064437150955,
0.06431727856397629,
-0.03146851807832718,
0.03498018905520439,
-0.16233310103416443,
-0.046451386064291,
0.05842963606119156,
0.027500003576278687,
0.005532523151487112,
0.009432793594896793,
0.0249937754124403,
-0.055380914360284805,
-0.07292910665273666,
0.03126820549368858,
-0.0960875153541565,
0.02612292394042015,
-0.08369423449039459,
-0.10104287415742874,
-0.08101139217615128,
-0.03866643086075783,
-0.0494711734354496,
0.021322103217244148,
-1.801287135581764e-33,
0.0019778399728238583,
0.0038558992091566324,
-0.08499135822057724,
0.05845874547958374,
-0.09602022916078568,
-0.06357841938734055,
-0.009429541416466236,
0.03864588961005211,
-0.02678845264017582,
-0.010960335843265057,
-0.005035965703427792,
-0.023628389462828636,
0.0686931386590004,
-0.009368804283440113,
0.028236277401447296,
-0.05222726985812187,
-0.04898481070995331,
-0.0245746448636055,
0.08060277998447418,
0.06657823920249939,
0.059236880391836166,
0.1128864735364914,
-0.11389973759651184,
0.034233614802360535,
-0.019621649757027626,
0.006487270817160606,
-0.018956147134304047,
0.09668902307748795,
-0.01428140327334404,
0.047954261302948,
-0.056734297424554825,
-0.046556685119867325,
-0.05235490947961807,
-0.012559444643557072,
-0.00878505315631628,
0.035606469959020615,
0.08854089677333832,
0.03298833593726158,
-0.007844775915145874,
0.035780083388090134,
0.08721429109573364,
-0.05452794209122658,
-0.03554832935333252,
0.000025407000066479668,
-0.020217863842844963,
-0.032121751457452774,
-0.0973646491765976,
0.052143681794404984,
-0.04235275462269783,
-0.010877433232963085,
0.029948975890874863,
-0.025916941463947296,
-0.02894412912428379,
-0.02075864188373089,
-0.024271251633763313,
-0.05823560059070587,
-0.04915354773402214,
-0.01497135404497385,
0.047992102801799774,
0.030182627961039543,
-0.07752036303281784,
-0.022705545648932457,
0.0473376139998436,
-0.0033425772562623024,
-0.007137912791222334,
0.007655210793018341,
-0.03403224050998688,
-0.002170225838199258,
0.09002745151519775,
-0.06701387465000153,
0.04313923418521881,
-0.024285906925797462,
-0.05098297819495201,
0.11021905392408371,
-0.02418188750743866,
-0.012906325049698353,
0.02060556411743164,
-0.06820248067378998,
-0.03312879055738449,
-0.04806924983859062,
0.01294000819325447,
-0.018611326813697815,
0.07390352338552475,
-0.004591384436935186,
0.08642224222421646,
0.017058663070201874,
-0.005745472386479378,
0.037084415555000305,
0.029300259426236153,
0.029141681268811226,
0.05592479556798935,
0.03500554710626602,
0.013612663373351097,
0.02605750411748886,
-0.0891188532114029,
-4.7604160613445856e-8,
-0.016224607825279236,
0.03780758008360863,
0.0647430568933487,
0.045081473886966705,
0.041753605008125305,
-0.0440790094435215,
-0.03253332898020744,
-0.029338272288441658,
-0.053206175565719604,
-0.07819389551877975,
0.0354950912296772,
0.0322696827352047,
-0.04915701225399971,
-0.08507528156042099,
-0.04228638485074043,
0.04336797818541527,
0.03255370259284973,
-0.0026787519454956055,
-0.04293057322502136,
0.06359174102544785,
0.06086773797869682,
0.05739546939730644,
-0.026998907327651978,
-0.017681362107396126,
0.005460337270051241,
-0.05326232314109802,
-0.06180848181247711,
0.10681405663490295,
0.04935934394598007,
0.029009154066443443,
0.013218940235674381,
0.014375929720699787,
-0.02589857578277588,
-0.08766183257102966,
0.05851706862449646,
0.08063856512308121,
-0.01382829062640667,
-0.09476161748170853,
0.0038606622256338596,
-0.038219206035137177,
0.03993756324052811,
0.03583700954914093,
-0.05920251086354256,
-0.007488960865885019,
0.004401304759085178,
-0.08183547109365463,
-0.016722217202186584,
-0.10928978025913239,
-0.0123435752466321,
-0.012069317512214184,
0.07162976264953613,
-0.03513690456748009,
-0.022990096360445023,
0.011864613741636276,
-0.05022844672203064,
-0.011826589703559875,
-0.060344669967889786,
-0.06587888300418854,
-0.10517162829637527,
0.038484230637550354,
0.025111226364970207,
0.10765641927719116,
0.06972631067037582,
-0.02045314572751522
] |
https://github.com/rycolab/flare
| 3 | 0 | 0 | 0 |
Nonconvex Stochastic Optimization under Heavy-Tailed Noises: Optimal Convergence without Gradient Clipping
|
https://openreview.net/forum?id=NKotdPUc3L
|
[
"Zijian Liu",
"Zhengyuan Zhou"
] |
Poster
|
Recently, the study of heavy-tailed noises in first-order nonconvex stochastic optimization has gotten a lot of attention since it was recognized as a more realistic condition as suggested by many empirical observations. Specifically, the stochastic noise (the difference between the stochastic and true gradient) is considered to have only a finite $\mathfrak{p}$-th moment where $\mathfrak{p}\in\left(1,2\right]$ instead of assuming it always satisfies the classical finite variance assumption. To deal with this more challenging setting, people have proposed different algorithms and proved them to converge at an optimal $\mathcal{O}(T^{\frac{1-\mathfrak{p}}{3\mathfrak{p}-2}})$ rate for smooth objectives after $T$ iterations. Notably, all these new-designed algorithms are based on the same technique – gradient clipping. Naturally, one may want to know whether the clipping method is a necessary ingredient and the only way to guarantee convergence under heavy-tailed noises. In this work, by revisiting the existing Batched Normalized Stochastic Gradient Descent with Momentum (Batched NSGDM) algorithm, we provide the first convergence result under heavy-tailed noises but without gradient clipping. Concretely, we prove that Batched NSGDM can achieve the optimal $\mathcal{O}(T^{\frac{1-\mathfrak{p}}{3\mathfrak{p}-2}})$ rate even under the relaxed smooth condition. More interestingly, we also establish the first $\mathcal{O}(T^{\frac{1-\mathfrak{p}}{2\mathfrak{p}}})$ convergence rate in the case where the tail index $\mathfrak{p}$ is unknown in advance, which is arguably the common scenario in practice.
|
Stochastic Optimization, Heavy-Tailed Noises
| null | 11,651 |
2412.19529
|
[
-0.06923831254243851,
-0.07625699788331985,
0.061495281755924225,
-0.003923209384083748,
0.035463426262140274,
-0.08698728680610657,
-0.0031535436864942312,
-0.02991594187915325,
0.033170994371175766,
-0.03596235066652298,
-0.03777099400758743,
0.07313738018274307,
-0.0030140895396471024,
-0.07999279350042343,
-0.009389471262693405,
-0.03953878581523895,
0.09834988415241241,
0.049962639808654785,
0.06623400747776031,
-0.013826446607708931,
0.031166929751634598,
0.003804475534707308,
0.05407727509737015,
0.024951571598649025,
0.06147821992635727,
-0.06798824667930603,
-0.014260069467127323,
-0.0035087892320007086,
0.041980184614658356,
-0.008418741635978222,
0.09164702147245407,
-0.017413610592484474,
0.032689012587070465,
-0.10865581780672073,
0.060007061809301376,
0.05938391759991646,
-0.04642533138394356,
0.057122208178043365,
0.0003050779632758349,
0.040421634912490845,
-0.034125566482543945,
-0.0028981997165828943,
-0.058550767600536346,
-0.04648646339774132,
0.014275670982897282,
-0.10326636582612991,
0.02321041375398636,
-0.09784820675849915,
-0.0010852691484615207,
-0.011792155914008617,
-0.01806594431400299,
-0.01073264330625534,
-0.059051621705293655,
-0.03213176503777504,
-0.014716179110109806,
-0.014636115171015263,
0.0849798396229744,
0.047289181500673294,
0.027564099058508873,
-0.009332343935966492,
0.039268236607313156,
-0.1271447390317917,
-0.0023183480370789766,
-0.0185566283762455,
0.02616732195019722,
-0.013972735963761806,
0.04639551788568497,
-0.07847775518894196,
0.024461477994918823,
0.09242673218250275,
-0.06601133942604065,
0.08498818427324295,
-0.07827643305063248,
0.08803095668554306,
0.030987290665507317,
0.028484901413321495,
0.07071135938167572,
-0.026160189881920815,
-0.050984255969524384,
-0.04663686081767082,
-0.0010644588619470596,
-0.03337438032031059,
0.025987448170781136,
-0.10407206416130066,
-0.045042913407087326,
-0.0009383199503645301,
-0.07802237570285797,
-0.013876572251319885,
0.06681844592094421,
-0.09602746367454529,
0.019052254036068916,
0.02543525956571102,
-0.13939610123634338,
0.08788567036390305,
0.09494983404874802,
-0.03797908499836922,
0.01796458102762699,
0.011286410503089428,
0.052066620439291,
0.09221713244915009,
0.012733062729239464,
0.007337897550314665,
0.0003894562541972846,
-0.013313938863575459,
0.02927277982234955,
-0.0676385685801506,
0.015363248065114021,
-0.0718659907579422,
0.009655511006712914,
-0.03942190855741501,
0.025608805939555168,
0.044363051652908325,
0.05262482166290283,
0.023352336138486862,
-0.023620568215847015,
-0.03646164759993553,
0.007067949045449495,
-0.03631288558244705,
0.0034233564510941505,
0.040657781064510345,
-0.052045587450265884,
-0.035743143409490585,
-0.03317464888095856,
0.05987478047609329,
0.0686023011803627,
0.09264149516820908,
-0.002003150060772896,
3.635397746858332e-33,
-0.053393904119729996,
0.05870143696665764,
0.05934073030948639,
-0.1018640398979187,
-0.013794675469398499,
0.04194606468081474,
0.06049719452857971,
-0.0873744785785675,
-0.045080628246068954,
0.022963209077715874,
0.05697932094335556,
-0.07892438769340515,
-0.023729149252176285,
-0.05265563353896141,
0.10693779587745667,
-0.10655120015144348,
-0.04125470295548439,
0.044309310615062714,
0.01795363612473011,
-0.010295492596924305,
0.11040604114532471,
-0.04504577815532684,
0.0039265756495296955,
0.05308046191930771,
0.0025244273710995913,
-0.07559122145175934,
0.07257533073425293,
-0.006598315667361021,
-0.0022320495918393135,
0.07985235005617142,
-0.08802255243062973,
0.001662998110987246,
0.03600817173719406,
-0.008339467458426952,
-0.018959645181894302,
0.007286690175533295,
-0.0177605003118515,
-0.030133116990327835,
-0.03233372047543526,
-0.014687702991068363,
-0.021558817476034164,
0.0423949733376503,
0.010022154077887535,
-0.0025589095894247293,
-0.10056093335151672,
-0.14960969984531403,
0.048382967710494995,
0.0729924812912941,
-0.04655676335096359,
-0.07050920277833939,
0.04618337005376816,
0.022560523822903633,
-0.0556923970580101,
0.0031689489260315895,
0.027277827262878418,
0.006399907637387514,
0.07824225723743439,
-0.055703435093164444,
0.04000027850270271,
0.026067016646265984,
0.02043616771697998,
-0.027196133509278297,
-0.010998657904565334,
-0.005785421468317509,
0.02966872602701187,
0.05313602089881897,
-0.0028081913478672504,
0.05910932645201683,
-0.038680143654346466,
-0.06604908406734467,
0.017510470002889633,
0.05279022455215454,
-0.06902013719081879,
-0.10692480206489563,
-0.006829855497926474,
-0.07044749706983566,
0.04381175339221954,
0.04523244500160217,
0.011142613366246223,
-0.050382036715745926,
-0.017381219193339348,
0.035900749266147614,
-0.007170724682509899,
-0.015631087124347687,
-0.0115822684019804,
-0.03665466234087944,
0.029479337856173515,
-0.03294594585895538,
-0.0930681824684143,
0.012776033952832222,
-0.048049941658973694,
-0.004523439798504114,
-0.02048933319747448,
0.05706494674086571,
-0.031967129558324814,
-3.266259827263502e-33,
-0.008168861269950867,
0.007754504680633545,
0.018534500151872635,
0.11245869100093842,
0.04342955723404884,
0.03211776539683342,
-0.0533357709646225,
-0.05304891988635063,
-0.07065549492835999,
-0.09775064140558243,
-0.01837378740310669,
0.020030925050377846,
-0.07244256138801575,
0.003993424586951733,
0.011554904282093048,
-0.05545004829764366,
0.07098729908466339,
0.12045426666736603,
-0.007056012749671936,
-0.02271200530230999,
0.09276751428842545,
-0.028419235721230507,
0.03122570551931858,
0.022784961387515068,
-0.02453623153269291,
0.006075321696698666,
-0.11943849921226501,
0.063239686191082,
-0.02909242734313011,
0.024699905887246132,
-0.0794534906744957,
0.010069954209029675,
-0.05371364578604698,
0.006571408361196518,
-0.023484474048018456,
0.10060066729784012,
-0.029535481706261635,
0.10856914520263672,
0.06894142180681229,
0.11631929874420166,
0.017366312444210052,
0.015883497893810272,
0.022848429158329964,
-0.045764535665512085,
0.022161239758133888,
-0.05170706287026405,
-0.022827468812465668,
-0.00017343927174806595,
0.0018190426053479314,
0.043348412960767746,
0.047981102019548416,
0.012773776426911354,
-0.03275252506136894,
0.14676053822040558,
-0.023228999227285385,
0.043717216700315475,
-0.058413755148649216,
-0.058181170374155045,
0.024219904094934464,
-0.01671013981103897,
-0.03361397609114647,
0.03625528886914253,
-0.09036696702241898,
0.07927500456571579,
0.014594130218029022,
-0.009272879920899868,
-0.030941827222704887,
-0.0027559478767216206,
0.08239997923374176,
0.01124555990099907,
-0.021552346646785736,
0.03806071728467941,
-0.007274925708770752,
0.07141538709402084,
-0.018615780398249626,
0.048065658658742905,
0.048554807901382446,
0.04520076885819435,
-0.03653265908360481,
-0.04127131775021553,
0.004922033287584782,
0.06905253976583481,
0.04603561758995056,
-0.05719342827796936,
0.04944370687007904,
-0.000735119974706322,
-0.029452675953507423,
-0.0168926864862442,
0.041485633701086044,
-0.047312602400779724,
0.020136913284659386,
-0.0015875828685238957,
0.06851315498352051,
-0.03685249760746956,
-0.0039876182563602924,
-4.895928285009177e-8,
-0.019535288214683533,
-0.011592685244977474,
0.04023456946015358,
-0.010200466960668564,
0.13010500371456146,
-0.011912317015230656,
-0.021338779479265213,
-0.04334752634167671,
0.01413783524185419,
-0.07481968402862549,
0.07754743844270706,
-0.03832028806209564,
0.00626137712970376,
0.00963498279452324,
-0.06188958138227463,
-0.01172668393701315,
-0.020853061228990555,
0.020061174407601357,
-0.0003385781601537019,
-0.03791292384266853,
-0.06745985895395279,
0.04597129300236702,
0.01504785567522049,
-0.01064328197389841,
0.06495895981788635,
-0.01090474147349596,
-0.021432077512145042,
-0.02492229826748371,
0.1344643533229828,
0.042092494666576385,
-0.04534691572189331,
0.0033239417243748903,
-0.03898857533931732,
-0.0021127089858055115,
0.011038058437407017,
0.0554945170879364,
-0.025137096643447876,
0.04909266158938408,
-0.04437146335840225,
0.0010047727264463902,
0.03474007919430733,
0.07452671229839325,
0.06262855976819992,
-0.030377613380551338,
0.0020289362873882055,
0.01935836300253868,
-0.05445127189159393,
-0.009560049511492252,
0.025191567838191986,
0.0007095821783877909,
-0.02385886386036873,
0.043200161308050156,
-0.013620268553495407,
0.06354931741952896,
0.03871661797165871,
0.024239903315901756,
0.003752123098820448,
-0.07269328832626343,
-0.009155475534498692,
-0.03310015797615051,
0.046647898852825165,
-0.06097736582159996,
-0.06999751925468445,
0.03581751137971878
] | 0 | 0 | 0 | 0 |
|
Can We Trust Embodied Agents? Exploring Backdoor Attacks against Embodied LLM-Based Decision-Making Systems
|
https://openreview.net/forum?id=S1Bv3068Xt
|
[
"Ruochen Jiao",
"Shaoyuan Xie",
"Justin Yue",
"TAKAMI SATO",
"Lixu Wang",
"Yixuan Wang",
"Qi Alfred Chen",
"Qi Zhu"
] |
Poster
|
Large Language Models (LLMs) have shown significant promise in real-world decision-making tasks for embodied artificial intelligence, especially when fine-tuned to leverage their inherent common sense and reasoning abilities while being tailored to specific applications. However, this fine-tuning process introduces considerable safety and security vulnerabilities, especially in safety-critical cyber-physical systems. In this work, we propose the first comprehensive framework for **B**ackdoor **A**ttacks against **L**LM-based **D**ecision-making systems (BALD) in embodied AI, systematically exploring the attack surfaces and trigger mechanisms. Specifically, we propose three distinct attack mechanisms: *word injection*, *scenario manipulation*, and *knowledge injection*, targeting various components in the LLM-based decision-making pipeline. We perform extensive experiments on representative LLMs (GPT-3.5, LLaMA2, PaLM2) in autonomous driving and home robot tasks, demonstrating the effectiveness and stealthiness of our backdoor triggers across various attack channels, with cases like vehicles accelerating toward obstacles and robots placing knives on beds. Our word and knowledge injection attacks achieve nearly 100\% success rate across multiple models and datasets while requiring only limited access to the system. Our scenario manipulation attack yields success rates exceeding 65\%, reaching up to 90\%, and does not require any runtime system intrusion. We also assess the robustness of these attacks against defenses, revealing their resilience. Our findings highlight critical security vulnerabilities in embodied LLM systems and emphasize the urgent need for safeguarding these systems to mitigate potential risks.
|
Backdoor attacks, Large language models, Autonomous agents, Robotics
|
We propose a comprehensive framework on backdoor attacks against embodi LLM for decision making during fine-tuning, including three different attack mechanisms targeting various channels of the systems.
| 11,647 |
2405.20774
|
[
0.031227782368659973,
-0.06984502077102661,
0.03135267272591591,
-0.007436132989823818,
0.0796344131231308,
-0.025975817814469337,
0.057316120713949203,
-0.033385615795850754,
0.03964720293879509,
0.09791101515293121,
0.025087973102927208,
-0.053927719593048096,
0.09774404764175415,
0.010285326279699802,
0.07634088397026062,
0.0067982142791152,
0.12149394303560257,
0.03823580592870712,
-0.1032600849866867,
0.0620669424533844,
0.062331050634384155,
0.03210139647126198,
0.02983078546822071,
0.00526762381196022,
-0.06241607666015625,
-0.0245341919362545,
0.0096893897280097,
-0.09772700816392899,
-0.0730988085269928,
-0.0798039361834526,
0.03313088044524193,
-0.06479694694280624,
-0.03577594459056854,
-0.024032356217503548,
-0.03662905842065811,
0.028383616358041763,
-0.023483259603381157,
0.0036922579165548086,
-0.030794549733400345,
-0.12091155350208282,
-0.061789363622665405,
-0.10612700134515762,
0.05731455981731415,
0.025444766506552696,
0.015309364534914494,
0.04761495068669319,
-0.00562975462526083,
-0.030684031546115875,
-0.01002193707972765,
-0.06618369370698929,
-0.09655299037694931,
-0.03168491646647453,
0.06403911858797073,
-0.036348938941955566,
-0.07276146858930588,
-0.012841403484344482,
0.003638023743405938,
0.04584285989403725,
0.02343951165676117,
-0.002655204152688384,
0.029856478795409203,
-0.03428739309310913,
0.017288895323872566,
0.022284263744950294,
-0.004233216866850853,
0.006268318276852369,
0.026213672012090683,
-0.017521405592560768,
0.03635689243674278,
0.06468982249498367,
0.06926968693733215,
-0.014533493667840958,
0.002248086268082261,
-0.04009077697992325,
0.039539534598588943,
0.010000405833125114,
-0.017614364624023438,
-0.022891076281666756,
0.055661071091890335,
-0.06086109206080437,
0.042418669909238815,
0.00885063223540783,
0.028512299060821533,
0.05933773145079613,
0.029788551852107048,
0.04691247269511223,
-0.014653529971837997,
0.05988558381795883,
0.059486858546733856,
0.046839792281389236,
-0.021438904106616974,
-0.14353881776332855,
0.08438263088464737,
0.047756992280483246,
0.07750793546438217,
-0.06411398202180862,
-0.03945150598883629,
-0.053686078637838364,
-0.031926341354846954,
0.060497723519802094,
-0.06350118666887283,
0.00113197963219136,
-0.05706057325005531,
0.008923552930355072,
0.0376710407435894,
0.034779489040374756,
-0.004732756409794092,
-0.08441849052906036,
0.023581836372613907,
-0.027971040457487106,
0.00846504233777523,
-0.021927539259195328,
0.07084231078624725,
0.05267679691314697,
0.02293938398361206,
-0.026420820504426956,
0.032732196152210236,
0.01727289892733097,
0.05528803542256355,
0.028836704790592194,
0.037386760115623474,
-0.02190519869327545,
0.07516317814588547,
0.007423478178679943,
0.06831824779510498,
-0.03942262753844261,
-0.031344007700681686,
-3.2487471654881575e-33,
-0.0035306515637785196,
0.03138747811317444,
-0.0033399120438843966,
0.009457685053348541,
0.06090811640024185,
-0.08034945279359818,
-0.002737536560744047,
0.0805896744132042,
-0.005403271876275539,
0.07020485401153564,
-0.0730598121881485,
-0.01226888969540596,
-0.014274606481194496,
0.09436466544866562,
0.10469994693994522,
0.07939714193344116,
-0.002968879183754325,
0.001588484738022089,
-0.05430053174495697,
-0.04440118744969368,
0.054417070001363754,
-0.012221574783325195,
-0.035698484629392624,
-0.040554214268922806,
0.015613270923495293,
0.02248232252895832,
0.03122040070593357,
-0.015883242711424828,
0.043436162173748016,
0.026135146617889404,
-0.12975212931632996,
0.029743025079369545,
-0.01474873535335064,
0.06511258333921432,
-0.009351990185678005,
-0.038978006690740585,
-0.10952155292034149,
-0.07383768260478973,
-0.02482006326317787,
-0.013898304663598537,
0.027956971898674965,
-0.02709891088306904,
-0.053434718400239944,
-0.008781805634498596,
-0.03669871389865875,
-0.045039765536785126,
-0.01456359215080738,
0.050227854400873184,
-0.054147861897945404,
-0.07095886021852493,
0.03989921510219574,
0.08431905508041382,
0.036270108073949814,
0.019292453303933144,
-0.045731157064437866,
-0.05961025506258011,
-0.03355487808585167,
0.0016297845868393779,
0.0260658860206604,
0.049011651426553726,
-0.018065668642520905,
0.013786006718873978,
0.017457280308008194,
0.04794589802622795,
-0.007022193167358637,
-0.016288526356220245,
-0.033407989889383316,
0.05835476517677307,
0.043160125613212585,
0.005181056447327137,
-0.05325046554207802,
-0.03103998675942421,
0.032158076763153076,
-0.03629307076334953,
-0.03933427110314369,
0.03080143593251705,
0.06646016240119934,
-0.03729776293039322,
-0.009322090074419975,
-0.00410058256238699,
0.025032564997673035,
0.021912291646003723,
-0.01506972499191761,
0.021983958780765533,
0.0012786693405359983,
-0.05318184569478035,
-0.054060470312833786,
-0.14308835566043854,
-0.060148097574710846,
0.07266925275325775,
-0.1050780788064003,
-0.10215514153242111,
-0.03130121901631355,
0.07687712460756302,
-0.08762073516845703,
8.924083115084102e-34,
-0.06365469098091125,
-0.03785074129700661,
-0.03158758580684662,
0.0944073349237442,
-0.058045193552970886,
-0.08447191119194031,
-0.023353951051831245,
-0.006535759195685387,
0.01688818819820881,
-0.007862570695579052,
-0.08126428723335266,
0.000823358481284231,
0.08679354190826416,
0.013219119980931282,
0.0903954729437828,
-0.08226039260625839,
-0.0010810852982103825,
-0.04280858114361763,
-0.04769473522901535,
0.0020496021024882793,
0.017559360712766647,
0.06945621967315674,
-0.05466049164533615,
0.037119872868061066,
0.039702579379081726,
0.07301415503025055,
0.004986583720892668,
0.03856516629457474,
0.01317750196903944,
0.08950809389352798,
-0.030869511887431145,
0.0061547779478132725,
0.04357802867889404,
0.06349500268697739,
-0.029508084058761597,
0.04765159636735916,
0.050109703093767166,
-0.010959424078464508,
-0.028167089447379112,
-0.02217676304280758,
0.06409366428852081,
-0.05756877735257149,
-0.054216139018535614,
-0.009674908593297005,
-0.014960815198719501,
-0.028041409328579903,
-0.08343315869569778,
0.013891209848225117,
0.027722511440515518,
-0.03363273665308952,
0.0849301666021347,
0.009537739679217339,
-0.020023945719003677,
-0.015741361305117607,
-0.071062371134758,
-0.032103247940540314,
0.05824510008096695,
-0.004941598046571016,
0.0723620280623436,
0.04690387099981308,
-0.024273933842778206,
0.001133336452767253,
0.0010971149895340204,
0.03492351248860359,
-0.0020268913358449936,
-0.01925857923924923,
-0.081245057284832,
0.03313225507736206,
0.035164497792720795,
-0.051930636167526245,
0.08549404889345169,
-0.0054781194776296616,
0.005370893515646458,
0.03917096182703972,
0.030423259362578392,
0.05165991187095642,
0.003714012447744608,
-0.11887399107217789,
-0.013088644482195377,
0.024167295545339584,
0.03975418210029602,
-0.0400368869304657,
0.07659463584423065,
0.10505371540784836,
0.023914681747555733,
0.04756716638803482,
-0.01979614421725273,
0.04660244658589363,
-0.014440170489251614,
0.02754831872880459,
-0.010658537968993187,
0.025357702746987343,
-0.019535839557647705,
0.07826712727546692,
-0.03964607045054436,
-4.316068213938706e-8,
0.002835396444424987,
0.012107810005545616,
0.06391932815313339,
0.0154196098446846,
0.004227857105433941,
0.03671359270811081,
-0.061493322253227234,
-0.08180925250053406,
-0.044192854315042496,
-0.03790859505534172,
0.07691037654876709,
-0.03701715171337128,
-0.01335748191922903,
-0.013287235051393509,
-0.004992441739886999,
0.08027651160955429,
-0.03860882297158241,
-0.013253548182547092,
-0.06632673740386963,
-0.0648113489151001,
0.06052177771925926,
-0.0013319216668605804,
-0.016231711953878403,
0.008344624191522598,
0.02318190224468708,
-0.06152784079313278,
-0.10724950581789017,
0.10554052889347076,
0.019254114478826523,
0.11139555275440216,
-0.04179809242486954,
-0.0667080506682396,
0.006514665205031633,
0.10023165494203568,
-0.010043739341199398,
0.08157195150852203,
-0.019359176978468895,
-0.09128246456384659,
0.0018254353199154139,
-0.0014140873681753874,
0.029325563460588455,
0.03954429179430008,
-0.03742238134145737,
0.007923522032797337,
-0.06119286268949509,
-0.011885310523211956,
-0.047510966658592224,
-0.13056497275829315,
-0.019704697653651237,
0.01891871728003025,
-0.0072442167438566685,
-0.0140910092741251,
-0.03106352686882019,
0.16084839403629303,
0.05170316994190216,
0.025007138028740883,
-0.019343966618180275,
-0.08516699820756912,
0.026669932529330254,
0.08125945180654526,
0.07125799357891083,
-0.007728622294962406,
-0.008958318270742893,
-0.048445478081703186
] | 0 | 0 | 0 | 0 |
|
Self-Attention-Based Contextual Modulation Improves Neural System Identification
|
https://openreview.net/forum?id=JeLqFpFzwX
|
[
"Isaac Lin",
"Tianye Wang",
"Shang Gao",
"Tang Shiming",
"Tai Sing Lee"
] |
Poster
|
Convolutional neural networks (CNNs) have been shown to be state-of-the-art models for visual cortical neurons. Cortical neurons in the primary visual cortex are sensitive to contextual information mediated by extensive horizontal and feedback connections. Standard CNNs integrate global contextual information to model contextual modulation via two mechanisms: successive convolutions and a fully connected readout layer. In this paper, we find that self-attention (SA), an implementation of non-local network mechanisms, can improve neural response predictions over parameter-matched CNNs in two key metrics: tuning curve correlation and peak tuning. We introduce peak tuning as a metric to evaluate a model's ability to capture a neuron's top feature preference. We factorize networks to assess each context mechanism, revealing that information in the local receptive field is most important for modeling overall tuning, but surround information is critically necessary for characterizing the tuning peak. We find that self-attention can replace posterior spatial-integration convolutions when learned incrementally, and is further enhanced in the presence of a fully connected readout layer, suggesting that the two context mechanisms are complementary. Finally, we find that decomposing receptive field learning and contextual modulation learning in an incremental manner may be an effective and robust mechanism for learning surround-center interactions.
|
self-attention, incremental learning, neural prediction, contextual modulation
|
Self-Attention-Based Contextual Modulation Improves Neural System Identification
| 11,639 |
2406.07843
|
[
0.0075385672971606255,
-0.10179407149553299,
0.0789722353219986,
0.04023918882012367,
0.03979305550456047,
0.022233914583921432,
0.02742057666182518,
-0.007974726147949696,
0.013020496815443039,
-0.08170377463102341,
-0.09889809787273407,
-0.048786334693431854,
-0.011634426191449165,
0.056632932275533676,
-0.0634312629699707,
-0.018726883456110954,
0.1701733022928238,
0.08091463893651962,
-0.07940656691789627,
-0.025941072031855583,
0.04820555821061134,
-0.05954361334443092,
0.023239679634571075,
0.009034095332026482,
0.04288116842508316,
0.03010011464357376,
-0.03054722771048546,
-0.036915212869644165,
0.010065400041639805,
-0.0018547080690041184,
0.03035125322639942,
0.02047833427786827,
-0.0010579939698800445,
0.02900511957705021,
-0.0648960992693901,
0.0577959269285202,
0.008260222151875496,
-0.021575462073087692,
-0.00037234657793305814,
-0.05846357345581055,
0.0023829611018300056,
0.006386009976267815,
0.025873733684420586,
0.1001734808087349,
0.08258764445781708,
0.03224286437034607,
0.02701001614332199,
-0.0000046548470891139004,
-0.016520248726010323,
-0.03451533988118172,
-0.0382634699344635,
-0.004309951327741146,
-0.06358195096254349,
0.014086593873798847,
0.02351793460547924,
0.08819980174303055,
0.05973521247506142,
0.04962838068604469,
0.016048002988100052,
0.04030989110469818,
0.022853268310427666,
-0.04336543753743172,
-0.023113269358873367,
0.009287375956773758,
-0.011137404479086399,
0.021096229553222656,
0.03444097936153412,
0.014165490865707397,
-0.057367954403162,
-0.04785792902112007,
0.04077792540192604,
0.08207758516073227,
0.024228263646364212,
-0.025575065985322,
0.10705415159463882,
0.05184391140937805,
-0.017542332410812378,
-0.016323890537023544,
0.03329284489154816,
-0.07206916809082031,
0.044511470943689346,
0.05395948141813278,
0.027683250606060028,
-0.04804623872041702,
0.1542741060256958,
-0.01125948503613472,
-0.008335706777870655,
0.034479979425668716,
0.015691597014665604,
0.028101861476898193,
-0.06444964557886124,
-0.09024424850940704,
-0.10125325620174408,
-0.056808724999427795,
0.03442729637026787,
0.06532073020935059,
-0.005604517180472612,
-0.10380064696073532,
-0.043975211679935455,
-0.02144089713692665,
-0.02129189483821392,
-0.038612525910139084,
0.03583480045199394,
0.020918983966112137,
0.07236557453870773,
-0.032072413712739944,
0.06451353430747986,
0.13220000267028809,
0.012128915637731552,
-0.0562637522816658,
-0.0016887590754777193,
0.011594090610742569,
0.0450449213385582,
-0.05769428238272667,
0.06656018644571304,
0.027452189475297928,
0.01855378970503807,
0.024690164253115654,
0.09016067534685135,
-0.02797759510576725,
-0.04527704045176506,
-0.06035952642560005,
-0.0835360586643219,
-0.045265115797519684,
-0.004967023152858019,
-0.011059106327593327,
-0.1041017547249794,
1.210750365139325e-33,
-0.01911003515124321,
0.028532356023788452,
0.083954818546772,
0.009952510707080364,
-0.04095417633652687,
-0.05567464232444763,
-0.041999414563179016,
-0.04553886502981186,
-0.007000163197517395,
-0.022095782682299614,
-0.030331797897815704,
0.06441794335842133,
-0.06503644585609436,
0.12069569528102875,
0.026536721736192703,
0.03775013983249664,
-0.029177885502576828,
-0.01285274513065815,
-0.01154601201415062,
-0.031239526346325874,
-0.019640963524580002,
0.02869410254061222,
-0.02114604040980339,
-0.0554463230073452,
-0.010535512119531631,
-0.03870101273059845,
0.028149954974651337,
-0.07616938650608063,
-0.08807361125946045,
-0.00824445765465498,
-0.05846894904971123,
0.07168895751237869,
0.006603956688195467,
0.003897454822435975,
-0.020568761974573135,
-0.08621258288621902,
-0.0455416701734066,
0.004767337813973427,
0.00984682235866785,
0.009773828089237213,
0.021349899470806122,
0.06679943203926086,
-0.06215036287903786,
0.05422234907746315,
-0.07338637113571167,
-0.09106161445379257,
-0.027527017518877983,
0.05160598084330559,
-0.06462904065847397,
-0.031851138919591904,
0.03508005291223526,
0.07667100429534912,
-0.027896258980035782,
-0.10710354149341583,
-0.04947070777416229,
-0.04792647808790207,
0.06568869948387146,
0.057774003595113754,
-0.012118467129766941,
0.046702466905117035,
0.011368395760655403,
-0.008041738532483578,
-0.010470135137438774,
0.09273062646389008,
0.024383671581745148,
0.0097495187073946,
0.018787775188684464,
0.07911675423383713,
0.006329888943582773,
0.023304026573896408,
-0.06159977614879608,
0.07527396082878113,
-0.050629425793886185,
-0.051565904170274734,
0.0011927918531000614,
-0.03633682802319527,
0.04058028385043144,
-0.03738102689385414,
0.013037953525781631,
0.028560219332575798,
0.076717309653759,
-0.017886610701680183,
-0.01956464909017086,
0.04477107152342796,
-0.0034927506931126118,
0.03981509432196617,
0.019209230318665504,
-0.10102793574333191,
-0.0301240012049675,
-0.06044892594218254,
0.017270388081669807,
-0.03324497491121292,
0.04506300762295723,
-0.04357369989156723,
-0.05185125023126602,
-1.2881189237520517e-33,
0.05320577323436737,
-0.000006125850632088259,
-0.008352452889084816,
-0.004368696827441454,
-0.046966977417469025,
0.013765262439846992,
0.023982709273695946,
0.008294552564620972,
-0.03920895606279373,
-0.02941027842462063,
-0.005175044760107994,
-0.05067203938961029,
0.004123134538531303,
0.056128449738025665,
-0.006207399535924196,
-0.04046230390667915,
-0.023817142471671104,
-0.005471831187605858,
-0.038777824491262436,
-0.037086766213178635,
0.053339503705501556,
0.06515300273895264,
-0.04181599244475365,
0.04621407762169838,
-0.09205736964941025,
0.0020585486199706793,
-0.12898477911949158,
0.04075264558196068,
-0.017486685886979103,
-0.038746427744627,
-0.031570836901664734,
0.06632596254348755,
0.022402945905923843,
-0.02667918987572193,
-0.03132312744855881,
0.08957710862159729,
0.032197996973991394,
-0.06883864104747772,
-0.10785150527954102,
0.04729751497507095,
0.14265185594558716,
0.027734944596886635,
0.02007818967103958,
0.05462922900915146,
-0.01583303138613701,
-0.03056574985384941,
-0.015667717903852463,
0.029468603432178497,
-0.06041781231760979,
0.11792208254337311,
-0.06355856359004974,
0.05959400162100792,
-0.08794145286083221,
-0.025013910606503487,
-0.04418016970157623,
-0.004778021480888128,
-0.0019475586013868451,
-0.010830835439264774,
0.16214121878147125,
-0.025742340832948685,
-0.024749821051955223,
-0.08560873568058014,
-0.026834391057491302,
-0.08068875223398209,
-0.011167589575052261,
0.01831461861729622,
0.049763891845941544,
-0.054876625537872314,
0.09767872840166092,
0.029011059552431107,
-0.033995285630226135,
-0.03806748986244202,
0.04343951866030693,
-0.022713012993335724,
-0.005383190233260393,
-0.011919990181922913,
0.018396547064185143,
-0.02138162963092327,
-0.04313204437494278,
0.00836711935698986,
0.014217723160982132,
0.010006684809923172,
-0.023495180532336235,
0.021355655044317245,
0.11051290482282639,
0.047705408185720444,
0.00899416208267212,
0.03768199682235718,
0.07809104025363922,
-0.030736900866031647,
-0.010693199932575226,
0.033142849802970886,
-0.07414748519659042,
0.07033596187829971,
-0.010268516838550568,
-4.320826008097356e-8,
-0.1225196048617363,
0.049420248717069626,
-0.0004861195629928261,
0.04418226704001427,
0.02823127806186676,
-0.009647184982895851,
0.030886223539710045,
0.0860387459397316,
-0.03425455465912819,
-0.010983291082084179,
0.06762946397066116,
-0.011370977386832237,
-0.0351225808262825,
-0.05437879264354706,
0.031562477350234985,
0.030185433104634285,
0.10091786831617355,
-0.014475240372121334,
0.023840995505452156,
0.003612149739637971,
-0.00793188251554966,
0.0007914535817690194,
0.02300644852221012,
-0.03250879794359207,
0.03924781456589699,
-0.04620540142059326,
-0.03987741842865944,
0.1153356283903122,
0.04743796959519386,
-0.03550180420279503,
0.044171418994665146,
-0.041941966861486435,
-0.018072236329317093,
-0.08296990394592285,
-0.07423935830593109,
0.0797802284359932,
0.04313850775361061,
0.029094688594341278,
0.044481322169303894,
-0.034310534596443176,
0.062243591994047165,
0.05973992869257927,
-0.019213935360312462,
0.08819857984781265,
-0.01529955118894577,
0.029880788177251816,
0.08105593174695969,
-0.07353662699460983,
0.0107166962698102,
-0.00577523373067379,
0.02659851498901844,
-0.04297538474202156,
-0.0037648228462785482,
0.04475194215774536,
-0.037479985505342484,
-0.03642822057008743,
0.01882605440914631,
0.01384403184056282,
0.003941959701478481,
0.050327446311712265,
-0.050749052315950394,
0.08035968244075775,
-0.07395108044147491,
-0.018526585772633553
] | 0 | 0 | 0 | 0 |
|
Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks
|
https://openreview.net/forum?id=c61unr33XA
|
[
"Siddharth Joshi",
"Jiayi Ni",
"Baharan Mirzasoleiman"
] |
Poster
|
Dataset distillation (DD) generates small synthetic datasets that can efficiently train deep networks with a limited amount of memory and compute. Despite the success of DD methods for supervised learning, DD for self-supervised pre-training of deep models has remained unaddressed. Pre-training on unlabeled data is crucial for efficiently generalizing to downstream tasks with limited labeled data. In this work, we propose the first effective DD method for SSL pre-training. First, we show, theoretically and empirically, that naiive application of supervised DD methods to SSL fails, due to the high variance of the SSL gradient. Then, we address this issue by relying on insights from knowledge distillation (KD) literature. Specifically, we train a small student model to match the representations of a larger teacher model trained with SSL. Then, we generate a small synthetic dataset by matching the training trajectories of the student models. As the KD objective has considerably lower variance than SSL, our approach can generate synthetic datasets that can successfully pre-train high-quality encoders. Through extensive experiments, we show that our distilled sets lead to up to 13% higher accuracy than prior work, on a variety of downstream tasks, in the presence of limited labeled data. Code at https://github.com/BigML-CS-UCLA/MKDT.
|
dataset distillation, self-supervised learning
|
We present the first effective method for dataset distillation (i.e. creating a small synthetic dataset to summarize a large real dataset) for self-supervised learning.
| 11,630 |
2410.02116
|
[
-0.10053145885467529,
-0.07983249425888062,
0.07882174849510193,
0.046146124601364136,
-0.001992722973227501,
0.010061048902571201,
-0.003936597146093845,
-0.09007593244314194,
0.027385428547859192,
-0.1383311152458191,
-0.03333671763539314,
0.019496669992804527,
0.018740041181445122,
0.03467600420117378,
-0.04253072291612625,
0.006789762061089277,
0.12044646590948105,
0.14498735964298248,
-0.1235746219754219,
-0.048569582402706146,
0.03765532001852989,
0.002074663992971182,
-0.012814649380743504,
0.06038907170295715,
0.07128243893384933,
-0.033100761473178864,
0.04143364354968071,
-0.05928143858909607,
0.009297751821577549,
-0.04868120700120926,
0.07640517503023148,
-0.020052233710885048,
-0.08859056234359741,
0.08729308098554611,
-0.04552951082587242,
0.03168277442455292,
-0.03387846052646637,
0.10151736438274384,
0.005202224478125572,
-0.03477425128221512,
0.03846602886915207,
-0.015396800823509693,
0.0029731791000813246,
0.07451515644788742,
0.010940623469650745,
0.015151776373386383,
0.004698867443948984,
-0.057933658361434937,
0.025878261774778366,
-0.003004341619089246,
-0.05423901230096817,
-0.00304094678722322,
-0.026151027530431747,
0.09428145736455917,
-0.025231176987290382,
-0.0022447947412729263,
0.06509201973676682,
-0.006339959800243378,
-0.04048483446240425,
0.07423342764377594,
0.00382796349003911,
-0.07097773253917694,
-0.043765150010585785,
-0.01732807233929634,
-0.06526120007038116,
-0.005183646455407143,
0.01990063302218914,
0.07215284556150436,
0.03469102084636688,
0.00426057493314147,
0.015648072585463524,
0.1025785282254219,
-0.024726474657654762,
0.055473752319812775,
0.07315932959318161,
-0.037437669932842255,
0.017233064398169518,
-0.0030670722480863333,
0.06659022718667984,
-0.09776246547698975,
0.03226557746529579,
0.027306299656629562,
0.04138268530368805,
0.010520011186599731,
0.09496244788169861,
0.04899205267429352,
0.03535538539290428,
0.015292328782379627,
0.034346845000982285,
0.01046749483793974,
-0.06710473448038101,
-0.017495637759566307,
-0.02019868977367878,
-0.034441933035850525,
0.03050064854323864,
0.030191071331501007,
0.04248218983411789,
-0.06744296103715897,
0.0009552983101457357,
0.05115082859992981,
0.005654243752360344,
0.1145729199051857,
-0.0029721204191446304,
-0.032641079276800156,
0.004672923590987921,
-0.006294010207056999,
0.09971088916063309,
0.048001788556575775,
-0.0005135759129188955,
-0.08471015095710754,
0.06533215939998627,
0.061606016010046005,
-0.0070235771127045155,
-0.017398085445165634,
0.01674804463982582,
-0.003110605524852872,
-0.02245979569852352,
-0.022245870903134346,
-0.061755258589982986,
0.03745424374938011,
-0.08229178935289383,
0.030198050662875175,
0.010783568024635315,
0.06158657371997833,
-0.07569309324026108,
-0.028811486437916756,
-0.09849360585212708,
1.6163971188897803e-33,
0.07540343701839447,
0.05003489553928375,
-0.01769254170358181,
-0.046168211847543716,
-0.010571137070655823,
-0.06073480099439621,
-0.002260204404592514,
0.049354903399944305,
-0.060931093990802765,
0.043839048594236374,
-0.02392144687473774,
-0.04516861215233803,
-0.11836739629507065,
0.0031831238884478807,
0.04965487867593765,
0.004713567905128002,
-0.08160645514726639,
0.026224380359053612,
0.01425658818334341,
-0.03467166796326637,
0.083103246986866,
-0.025748994201421738,
-0.002369045978412032,
-0.018642734736204147,
0.008057666011154652,
0.008028627373278141,
0.03392700105905533,
0.022580694407224655,
0.0027304054237902164,
0.016241662204265594,
-0.05728510767221451,
-0.057627927511930466,
0.059493646025657654,
0.02702530473470688,
0.058560434728860855,
-0.0011509362375363708,
0.02264227531850338,
0.03062012605369091,
0.04071792587637901,
-0.02791343256831169,
0.030364254489541054,
0.07869365066289902,
0.0077438391745090485,
0.012272181920707226,
-0.1091257780790329,
-0.08486344665288925,
-0.00498245470225811,
-0.0013971454463899136,
0.0015463695162907243,
-0.042593374848365784,
-0.005883254110813141,
0.0003969609388150275,
-0.0696411207318306,
-0.08024778962135315,
-0.07780701667070389,
-0.001058870111592114,
0.06785891205072403,
0.06281951814889908,
0.02847573533654213,
0.05932248756289482,
-0.07681708037853241,
0.018463147804141045,
-0.07722540199756622,
0.11060873419046402,
0.0036452924832701683,
-0.028991492465138435,
0.0333014577627182,
0.05942258611321449,
0.028305834159255028,
-0.07563967257738113,
-0.10526454448699951,
-0.0008974816300906241,
-0.017485560849308968,
-0.07431130856275558,
0.0668710470199585,
-0.017829280346632004,
0.018113335594534874,
-0.03969879448413849,
0.011184577830135822,
0.03252173215150833,
0.06960850208997726,
-0.012608281336724758,
-0.05283172056078911,
-0.0021889552008360624,
-0.06049831956624985,
0.05920347198843956,
0.002641564467921853,
-0.09582512825727463,
0.011568281799554825,
0.008667710237205029,
-0.033199697732925415,
-0.04072064533829689,
-0.0002642919425852597,
0.002044757828116417,
0.007410339079797268,
-3.829099379406675e-34,
-0.007419760338962078,
0.04757051169872284,
-0.004581465385854244,
0.10492765158414841,
0.02873922511935234,
-0.04671172797679901,
0.0197150856256485,
0.09093605726957321,
-0.02064330317080021,
0.01267366111278534,
0.016365187242627144,
-0.06614869087934494,
0.024555277079343796,
0.010115978308022022,
-0.019796324893832207,
0.01798657700419426,
-0.05168227106332779,
0.03646926209330559,
-0.03838668763637543,
0.029263626784086227,
-0.02752130664885044,
0.10344269871711731,
-0.07753665745258331,
0.07174850255250931,
0.01669330894947052,
0.05303280055522919,
-0.04821840301156044,
0.12089923769235611,
-0.023615337908267975,
0.03423675149679184,
0.0067336102947592735,
-0.0019564463291317225,
0.018295496702194214,
-0.07655700296163559,
-0.1316426396369934,
0.03778766468167305,
0.010501958429813385,
-0.023279285058379173,
-0.012849051505327225,
0.028384126722812653,
0.06217222288250923,
-0.03046058677136898,
-0.07734595984220505,
-0.024638649076223373,
-0.005400095134973526,
-0.04786614328622818,
-0.06298334896564484,
0.025304002687335014,
0.08887013792991638,
0.03651420772075653,
0.003422253765165806,
0.03508181869983673,
-0.06907948851585388,
0.0038218337576836348,
-0.00033615861320868134,
-0.11942523717880249,
-0.0027008699253201485,
0.015495034866034985,
0.009289616718888283,
0.031641166657209396,
-0.08279561996459961,
-0.04038102924823761,
0.07048157602548599,
-0.046762075275182724,
-0.02093018963932991,
-0.09246014058589935,
-0.05886557698249817,
0.10101628303527832,
0.03982964903116226,
0.0031555164605379105,
-0.024651208892464638,
-0.015316134318709373,
0.019935043528676033,
-0.02132272534072399,
-0.025148579850792885,
-0.0859178751707077,
0.0005743015208281577,
-0.0014884751290082932,
-0.014808295294642448,
0.025037096813321114,
0.03644610568881035,
0.037363335490226746,
-0.01828981377184391,
0.015303133055567741,
0.16490723192691803,
0.05934393033385277,
0.04231147840619087,
0.04113250598311424,
-0.0033680747728794813,
0.03070724382996559,
-0.04792988300323486,
0.036730147898197174,
-0.03863826394081116,
0.04985188692808151,
-0.014497230760753155,
-3.602086451337527e-8,
-0.09545333683490753,
0.053165894001722336,
0.05388341099023819,
0.05464697256684303,
0.05387395992875099,
-0.02616347000002861,
0.0027602878399193287,
0.13250397145748138,
-0.004500746726989746,
-0.007757583167403936,
0.06414782255887985,
0.02854178659617901,
-0.06521777808666229,
-0.04861399158835411,
-0.020577479153871536,
0.07729622721672058,
0.01924964040517807,
0.010285614058375359,
0.015952521935105324,
-0.023562993854284286,
0.08937359601259232,
-0.005784673150628805,
0.03423270955681801,
0.002439843723550439,
0.09328363835811615,
-0.0742507129907608,
-0.00421547656878829,
0.09211865067481995,
0.027631720528006554,
0.041915103793144226,
-0.06668687611818314,
-0.03567245975136757,
-0.011391905136406422,
-0.06377492845058441,
0.06355760991573334,
0.10876353830099106,
-0.021227659657597542,
-0.024631783366203308,
-0.0128471152856946,
-0.004154952708631754,
-0.010889269411563873,
0.08548320084810257,
-0.08621873706579208,
0.01619827374815941,
-0.03504746034741402,
-0.025114521384239197,
-0.051848623901605606,
-0.011404668912291527,
0.03398574888706207,
0.007020614109933376,
-0.005228986963629723,
-0.07368757575750351,
0.01942453905940056,
0.025302929803729057,
0.003256828524172306,
0.0304487943649292,
-0.020632633939385414,
-0.029993774369359016,
-0.006756660994142294,
0.012939020060002804,
0.019379451870918274,
-0.009930403903126717,
-0.0122221065685153,
0.0005476531223393977
] |
https://github.com/bigml-cs-ucla/mkdt
| 2 | 0 | 0 | 0 |
Robotouille: An Asynchronous Planning Benchmark for LLM Agents
|
https://openreview.net/forum?id=OhUoTMxFIH
|
[
"Gonzalo Gonzalez-Pumariega",
"Leong Su Yean",
"Neha Sunkara",
"Sanjiban Choudhury"
] |
Poster
|
Effective asynchronous planning, or the ability to efficiently reason and plan over states and actions that must happen in parallel or sequentially, is essential for agents that must account for time delays, reason over diverse long-horizon tasks, and collaborate with other agents. While large language model (LLM) agents show promise in high-level task planning, current benchmarks focus primarily on short-horizon tasks and do not evaluate such asynchronous planning capabilities. We introduce Robotouille, a challenging benchmark environment designed to test LLM agents' ability to handle long-horizon asynchronous scenarios. Our synchronous and asynchronous datasets capture increasingly complex planning challenges that go beyond existing benchmarks, requiring agents to manage over-
lapping tasks and interruptions Our results show that ReAct (gpt-4o) achieves 47% on synchronous tasks but only 11% on asynchronous tasks, highlighting significant room for improvement. We further analyze failure modes, demonstrating the need for LLM agents to better incorporate long-horizon feedback and self-audit their reasoning during task execution.
|
benchmark, llm, agents, planning
|
Robotouille is an LLM agent benchmark for stress testing stress testing long-horizon synchronous and asynchronous planning capabilities.
| 11,625 |
2502.05227
|
[
-0.0343463197350502,
-0.034438855946063995,
0.011905936524271965,
0.06045681610703468,
0.07149780541658401,
-0.02522147446870804,
-0.05800578370690346,
-0.02648196741938591,
0.018927527591586113,
-0.010955248959362507,
-0.07823922485113144,
-0.056927524507045746,
-0.027654480189085007,
0.05423915758728981,
0.047216929495334625,
-0.031904932111501694,
0.09695183485746384,
-0.026001770049333572,
-0.06720319390296936,
-0.05320840701460838,
0.0008681833278387785,
-0.030990995466709137,
0.08573702722787857,
0.036293353885412216,
0.004690555389970541,
0.04028356447815895,
-0.004366945009678602,
-0.11685264855623245,
-0.004089285619556904,
0.007685970515012741,
0.014376395381987095,
0.009268867783248425,
-0.0505349226295948,
0.01144252996891737,
-0.0037933075800538063,
0.10271997004747391,
-0.07345208525657654,
0.011275415308773518,
0.008115816861391068,
-0.10062658041715622,
-0.019388001412153244,
-0.009900086559355259,
0.005982265807688236,
0.005898066330701113,
0.029980741441249847,
-0.007951113395392895,
0.014397398568689823,
-0.05393596366047859,
0.013565032742917538,
0.006878544110804796,
-0.0931096002459526,
-0.013708447106182575,
0.02439778484404087,
-0.04999612644314766,
-0.016770819202065468,
0.08653632551431656,
0.09034407883882523,
0.03523769974708557,
0.01725665107369423,
-0.05812749266624451,
0.013615324161946774,
-0.1311032474040985,
-0.025155942887067795,
-0.00276842899620533,
-0.029125239700078964,
0.0723685547709465,
0.03269471600651741,
-0.01334003172814846,
-0.03445769473910332,
0.023088665679097176,
-0.01758367381989956,
0.00834230612963438,
0.019931931048631668,
0.012732994742691517,
0.0026370645500719547,
0.007456018589437008,
0.0035108732990920544,
-0.007712365593761206,
0.07686582952737808,
-0.031664855778217316,
-0.023257404565811157,
-0.013104881159961224,
-0.02453291043639183,
0.06528421491384506,
0.07751473784446716,
0.01432759314775467,
0.08223136514425278,
0.08947432786226273,
0.07590000331401825,
-0.007646429818123579,
0.0267608892172575,
-0.04609549418091774,
0.0028006264474242926,
0.028521880507469177,
0.020815253257751465,
0.06435439735651016,
-0.0031698960810899734,
-0.0501042902469635,
-0.03728508949279785,
0.00858361553400755,
-0.010797729715704918,
0.06996074318885803,
-0.0269702710211277,
0.006898263934999704,
-0.01738084852695465,
-0.03385370224714279,
0.016588451340794563,
-0.0287021454423666,
0.0006514551932923496,
-0.04581577330827713,
0.021235428750514984,
-0.032127898186445236,
0.12958045303821564,
0.05920746549963951,
0.019320083782076836,
0.008176609873771667,
-0.021621394902467728,
0.01583971455693245,
0.0016301830764859915,
0.05109866335988045,
0.0528845451772213,
0.01436566561460495,
0.057199813425540924,
0.01620703563094139,
0.053832195699214935,
0.024536993354558945,
0.02784697152674198,
3.547105631484101e-33,
0.06283882260322571,
0.05729706212878227,
0.040363069623708725,
-0.0013408864615485072,
0.08434006571769714,
-0.000513735634740442,
0.0591735877096653,
0.02989177778363228,
-0.07999931275844574,
-0.0408923514187336,
0.025258198380470276,
-0.05813821405172348,
-0.0727102980017662,
0.03598317503929138,
0.026469314470887184,
-0.029956074431538582,
0.01472304854542017,
0.03865781053900719,
-0.07257021218538284,
-0.0016309989150613546,
0.06620814651250839,
-0.015360841527581215,
0.00576037960126996,
-0.06496135890483856,
0.09137875586748123,
0.010869321413338184,
0.023475002497434616,
-0.0021258804481476545,
-0.0009528323425911367,
0.0028674686327576637,
-0.12039503455162048,
0.015714727342128754,
-0.07420241087675095,
0.12854789197444916,
-0.041114192456007004,
-0.008894775994122028,
-0.11294174194335938,
-0.1199393942952156,
0.020919635891914368,
-0.010293023660779,
-0.04834809899330139,
0.03573109582066536,
-0.02160210721194744,
-0.03167755529284477,
-0.06483318656682968,
-0.10574883967638016,
0.03430347889661789,
-0.005234309937804937,
-0.01794252172112465,
-0.013092361390590668,
0.00812485720962286,
0.10490720719099045,
-0.025881873443722725,
-0.06372234970331192,
-0.005393504165112972,
-0.0522269643843174,
0.03501689434051514,
-0.017775125801563263,
0.020506976172327995,
0.11645123362541199,
0.07775089889764786,
-0.01805633306503296,
-0.04222903028130531,
0.09334123134613037,
0.011360786855220795,
0.017705751582980156,
-0.07818450778722763,
0.13122639060020447,
0.10702108591794968,
-0.02357896789908409,
0.00604977086186409,
-0.02412923239171505,
0.05212434381246567,
0.005365508608520031,
0.05356259271502495,
-0.011058889329433441,
0.03183656558394432,
-0.06276550889015198,
0.003994639031589031,
-0.043776556849479675,
-0.0033978093415498734,
-0.03771243616938591,
-0.05691837519407272,
0.02470555528998375,
-0.017796974629163742,
-0.03216729313135147,
-0.031028911471366882,
-0.044628072530031204,
-0.00889683049172163,
0.0011436390923336148,
-0.09560158103704453,
-0.060000985860824585,
0.0044011021964251995,
0.0051673962734639645,
-0.04920921474695206,
-4.0791303339044374e-33,
-0.021404990926384926,
-0.12607961893081665,
-0.033127561211586,
0.07157625257968903,
0.024030230939388275,
-0.07110434025526047,
-0.04988478869199753,
-0.0910828486084938,
-0.010231632739305496,
-0.027833200991153717,
-0.10403299331665039,
-0.061125654727220535,
0.029864082112908363,
0.02260063961148262,
0.020358286798000336,
-0.06885935366153717,
0.100863978266716,
-0.08654443174600601,
0.03803963214159012,
0.023107700049877167,
0.015184974297881126,
0.08835785835981369,
-0.08626122027635574,
0.03827459365129471,
-0.02175842970609665,
0.04195855185389519,
-0.015050232410430908,
-0.05693679302930832,
0.011919345706701279,
-0.009927015751600266,
-0.00432093720883131,
-0.01415332779288292,
-0.03355273976922035,
-0.015852972865104675,
0.02486410364508629,
0.0874079242348671,
0.004245664458721876,
0.009277429431676865,
-0.07468904554843903,
0.09460266679525375,
0.13864536583423615,
-0.1066470816731453,
-0.015182815492153168,
0.008229490369558334,
0.009749793447554111,
0.013559119775891304,
-0.1367667019367218,
0.045660678297281265,
-0.03091568686068058,
0.040911611169576645,
-0.028711333870887756,
0.030811356380581856,
-0.03261125460267067,
-0.04303530976176262,
-0.042847707867622375,
-0.076323963701725,
-0.024786530062556267,
-0.03953547030687332,
0.028336234390735626,
0.01600959710776806,
-0.0677102729678154,
0.0031518759205937386,
0.08392217010259628,
0.024241341277956963,
0.001855816226452589,
0.002373931696638465,
-0.039721403270959854,
-0.0033176036085933447,
0.09232612699270248,
0.035857222974300385,
0.06884225457906723,
-0.009385975077748299,
-0.01721121370792389,
0.05685194209218025,
0.06860309094190598,
0.018176471814513206,
-0.05162058398127556,
-0.07415524870157242,
0.01654018834233284,
-0.010347520932555199,
0.01561996154487133,
0.06546822190284729,
-0.0012382357381284237,
0.07886040210723877,
-0.04190871864557266,
0.0371692031621933,
0.01577436737716198,
0.09551545232534409,
0.0007671380299143493,
0.014417532831430435,
-0.038671065121889114,
-0.021872740238904953,
-0.06059383228421211,
0.07634168863296509,
-0.061013661324977875,
-4.2703700131596634e-8,
0.018029849976301193,
0.047080982476472855,
0.013964982703328133,
-0.0031021451577544212,
0.02160779759287834,
-0.05853449925780296,
0.002651786431670189,
-0.037984445691108704,
0.04893761873245239,
0.03761259838938713,
0.08127699047327042,
-0.05739331617951393,
0.004468540195375681,
-0.03201499208807945,
-0.014696436002850533,
0.0009756246581673622,
0.036500878632068634,
-0.026636121794581413,
0.014910662546753883,
-0.054756540805101395,
0.020609822124242783,
-0.006511523388326168,
-0.02694939449429512,
-0.03757883980870247,
0.029553553089499474,
-0.022990457713603973,
0.007237267214804888,
0.11795688420534134,
0.057730499655008316,
-0.008423457853496075,
0.008375933393836021,
-0.02857133187353611,
0.0289921797811985,
0.014319437555968761,
0.04726042598485947,
-0.010739270597696304,
0.046782538294792175,
-0.057878315448760986,
0.10568073391914368,
-0.024445341899991035,
0.04424799606204033,
0.06938457489013672,
-0.015082895755767822,
0.04408644139766693,
-0.027846992015838623,
-0.029234912246465683,
-0.09718474745750427,
-0.11447980999946594,
0.007916542701423168,
-0.09816503524780273,
-0.06332536041736603,
-0.054438699036836624,
-0.05762152746319771,
0.13226760923862457,
0.08372712880373001,
0.04723425209522247,
0.0037696969229727983,
-0.09143278002738953,
0.059256017208099365,
0.051259271800518036,
0.03007059171795845,
0.004145781509578228,
-0.05813160911202431,
-0.016492994502186775
] |
https://github.com/portal-cornell/robotouille
| 28 | 0 | 0 | 0 |
Learn-by-interact: A Data-Centric Framework For Self-Adaptive Agents in Realistic Environments
|
https://openreview.net/forum?id=3UKOzGWCVY
|
[
"Hongjin SU",
"Ruoxi Sun",
"Jinsung Yoon",
"Pengcheng Yin",
"Tao Yu",
"Sercan O Arik"
] |
Poster
|
Autonomous agents powered by large language models (LLMs) have the potential to enhance human capabilities, assisting with digital tasks from sending emails to performing data analysis. The abilities of existing LLMs at such tasks are often hindered by the lack of high-quality agent data from the corresponding environments they interact with. We propose LEARN-BY-INTERACT, a data-centric framework to adapt LLM agents to any given environments without human annotations. LEARN-BY-INTERACT synthesizes trajectories of agent-environment interactions based on documentations, and constructs instructions by summarizing or abstracting the interaction histories, a process called backward construction. We assess the quality of our synthetic data by using them in both training-based scenarios and training-free in-context learning (ICL), where we craft innovative retrieval approaches optimized for agents. Extensive experiments on SWE-bench, WebArena, OSWorld, and Spider2-V spanning across realistic coding, web, and desktop environments show the effectiveness of LEARN-BY-INTERACT in various downstream agentic tasks — baseline results are improved up to 11.1% for ICL with Claude-3.5 and 23.1% for training with Codestral-22B. We further demonstrate the critical role of backward construction, which provides up to 10.6% improvement for training. Our ablation studies demonstrate the efficiency provided by our synthesized data in ICL and the superiority of our retrieval pipeline over alternative approaches like conventional retrieval-augmented generation (RAG). We expect that LEARN-BY-INTERACT will serve as a foundation for agent data synthesis as LLMs are increasingly deployed at real-world environments.
|
Data synthesis, Agent, Adaptation
|
A data-centric framework to adapt LLM agents to any given environments without human annotations
| 11,615 | null |
[
-0.04492438584566116,
-0.07106707245111465,
-0.014559809118509293,
0.05693165957927704,
0.06613495200872421,
-0.06419362872838974,
0.010490593500435352,
0.0029165265150368214,
0.017580971121788025,
-0.03636838123202324,
-0.05481367185711861,
-0.046386539936065674,
0.06072833761572838,
0.03686261177062988,
0.05621154233813286,
-0.002619002480059862,
0.1320382058620453,
0.027222013100981712,
-0.08029821515083313,
-0.0017175090033560991,
0.023859282955527306,
0.040713176131248474,
0.09689567983150482,
0.01775127649307251,
-0.01620851829648018,
-0.036300089210271835,
-0.0268905907869339,
-0.051438916474580765,
0.03561742603778839,
0.005838755052536726,
0.0660087838768959,
0.07118877023458481,
0.0028674716595560312,
0.04223973676562309,
0.0007452073623426259,
0.04778682813048363,
-0.028504325076937675,
0.048663727939128876,
0.010593212209641933,
-0.03023543953895569,
-0.05785820633172989,
-0.015159666538238525,
0.009096331894397736,
0.010783830657601357,
0.055792465806007385,
-0.0333283431828022,
-0.021895254030823708,
-0.05224877595901489,
-0.02217893674969673,
0.058943454176187515,
-0.16984353959560394,
-0.05428736284375191,
0.02227516658604145,
0.02926127426326275,
-0.040397875010967255,
0.022829897701740265,
-0.017602602019906044,
0.048210110515356064,
-0.009344405494630337,
-0.044160593301057816,
0.007810454815626144,
-0.08126913011074066,
-0.005527069792151451,
0.02043786644935608,
-0.05886920914053917,
-0.019677305594086647,
0.000184129603439942,
0.05959057807922363,
0.018097490072250366,
-0.07450520247220993,
-0.07652800530195236,
0.026802215725183487,
0.0001668713812250644,
0.015324407257139683,
0.021030331030488014,
-0.009835369884967804,
-0.04393459111452103,
-0.005021485034376383,
0.0300834272056818,
-0.02693253941833973,
-0.017155297100543976,
0.022661855444312096,
0.03757959604263306,
0.0419665165245533,
0.08867950737476349,
0.02769056148827076,
0.08965607732534409,
0.04167274013161659,
0.05471397191286087,
0.025511234998703003,
-0.017063282430171967,
-0.05437677353620529,
0.041409824043512344,
-0.008145231753587723,
0.0544084869325161,
0.08981019258499146,
0.03133371099829674,
-0.07169116288423538,
-0.023191260173916817,
0.10016633570194244,
0.020021021366119385,
0.04975596070289612,
-0.05798109993338585,
-0.030004343017935753,
-0.05491778627038002,
0.02676820382475853,
-0.020945673808455467,
0.03603565692901611,
0.021488774567842484,
-0.07542898505926132,
0.017196450382471085,
0.053944457322359085,
0.002036125399172306,
-0.01674668863415718,
0.05775079503655434,
-0.026656771078705788,
0.04778634384274483,
-0.03203756734728813,
0.0493670329451561,
0.06686384975910187,
0.005199039354920387,
-0.0008084866567514837,
0.013802521862089634,
-0.0014912441838532686,
0.034961991012096405,
-0.061336684972047806,
0.0020760546904057264,
1.398050604427975e-33,
0.10514964163303375,
0.06780414283275604,
0.032020341604948044,
0.059580061584711075,
0.016193639487028122,
-0.021864255890250206,
0.027181297540664673,
0.04825931042432785,
-0.03452290594577789,
-0.03757217898964882,
0.056842293590307236,
0.09556014090776443,
0.004192337393760681,
0.03935561329126358,
0.025686074048280716,
0.04332892224192619,
-0.07315093278884888,
0.03490142524242401,
-0.012920394539833069,
-0.055501773953437805,
0.06784740835428238,
-0.04186565801501274,
0.027782343327999115,
-0.04946611821651459,
0.0589011088013649,
0.05189237371087074,
0.028510646894574165,
0.008851361460983753,
0.04099094122648239,
-0.0028875365387648344,
-0.07361917197704315,
-0.02265964262187481,
-0.006086159963160753,
0.07047844678163528,
-0.002797194290906191,
-0.017529379576444626,
-0.05654764547944069,
-0.03529196232557297,
0.006234278436750174,
0.04347328841686249,
-0.03717193379998207,
0.03296470642089844,
0.0657057985663414,
-0.022671161219477654,
-0.08569834381341934,
-0.1062060296535492,
0.002196732209995389,
0.0027427133172750473,
-0.02962648868560791,
-0.05632200837135315,
0.050517626106739044,
0.023417768999934196,
-0.05887287110090256,
-0.04389040172100067,
0.011201499029994011,
0.037771157920360565,
0.011838563717901707,
0.05016283690929413,
0.01462586410343647,
0.0685940682888031,
0.006928012240678072,
-0.020100899040699005,
0.042730771005153656,
0.11681224405765533,
0.06669645011425018,
-0.03366144374012947,
-0.0017987986793741584,
0.08874944597482681,
0.0691598579287529,
-0.06328365206718445,
-0.015460564754903316,
0.009163876064121723,
0.0072892894968390465,
0.01478451769798994,
0.06720840930938721,
-0.076968714594841,
0.028214210644364357,
-0.1327248215675354,
-0.004921480547636747,
0.019116153940558434,
0.06239661946892738,
-0.02810651995241642,
0.010198184289038181,
-0.03621754050254822,
-0.04361124709248543,
-0.05988764017820358,
-0.006770593114197254,
-0.11006313562393188,
-0.015533383935689926,
-0.016935326159000397,
-0.07023246586322784,
-0.06935299932956696,
-0.09845516085624695,
0.012300352565944195,
-0.0023782006464898586,
-2.2632685377404157e-33,
0.007090650498867035,
-0.0751945748925209,
-0.08469859510660172,
0.025529783219099045,
-0.013362809084355831,
-0.03636680915951729,
0.013653560541570187,
0.05232608690857887,
-0.003999468870460987,
-0.0691283643245697,
-0.054163455963134766,
-0.006408474408090115,
0.05537236109375954,
0.00001943153256434016,
-0.020803846418857574,
-0.024774666875600815,
-0.010505357757210732,
-0.022916143760085106,
-0.023707356303930283,
0.027671581134200096,
-0.0015923912869766355,
0.048705969005823135,
-0.06394001096487045,
0.03419337794184685,
-0.04301242530345917,
0.015300894156098366,
-0.007530837785452604,
0.051415834575891495,
0.046679265797138214,
0.05956582725048065,
-0.013691247440874577,
-0.02334609441459179,
-0.012230761349201202,
-0.05670924112200737,
-0.07812055945396423,
0.07794369012117386,
0.00936283078044653,
-0.05821504443883896,
-0.05443200096487999,
0.08275692909955978,
0.12392938137054443,
-0.04646475240588188,
-0.10850007086992264,
-0.026711825281381607,
-0.04627033695578575,
-0.0006395961972884834,
-0.15885280072689056,
0.017443737015128136,
0.07124119997024536,
0.04467540234327316,
0.007795829325914383,
0.011357652023434639,
-0.0432996042072773,
-0.07069652527570724,
-0.07223130762577057,
-0.06322485953569412,
0.029478680342435837,
-0.056048132479190826,
0.01413350086659193,
-0.038900226354599,
-0.11213864386081696,
-0.029829256236553192,
0.06243986636400223,
0.014310427010059357,
-0.006971368566155434,
-0.06536729633808136,
-0.07035068422555923,
0.019600018858909607,
-0.018136655911803246,
-0.01575382798910141,
0.08713831007480621,
-0.04617118090391159,
-0.06186438724398613,
-0.05303625762462616,
0.09378489851951599,
-0.011252755299210548,
-0.003516350407153368,
-0.04691363871097565,
-0.012551428750157356,
-0.05743977800011635,
-0.002341581741347909,
0.05392375960946083,
0.049693457782268524,
0.0573151633143425,
0.057427357882261276,
0.09837951511144638,
-0.041158076375722885,
0.1236349418759346,
0.03570336475968361,
0.021162955090403557,
0.03463541716337204,
0.0143048707395792,
-0.04167560487985611,
0.05220388248562813,
-0.06469452381134033,
-4.5027011452702936e-8,
-0.08765169233083725,
0.03968435898423195,
0.04253871366381645,
0.08155829459428787,
-0.026047231629490852,
0.01333096344023943,
-0.015200831927359104,
0.08062262088060379,
0.0020700402092188597,
0.009050663560628891,
0.014145306311547756,
-0.003369938349351287,
0.002378619508817792,
-0.05906806141138077,
0.02863161265850067,
0.12994712591171265,
0.03045179322361946,
-0.06303633749485016,
-0.023362882435321808,
-0.045680608600378036,
0.0797266885638237,
0.006613506935536861,
0.005356428679078817,
-0.019148901104927063,
0.04834134876728058,
-0.020789137110114098,
-0.0537458099424839,
0.09182117879390717,
0.016260456293821335,
0.011691700667142868,
0.006511988118290901,
-0.04515412822365761,
-0.01644643023610115,
-0.038861233741045,
-0.013881517574191093,
0.05931258201599121,
0.023725824430584908,
-0.12814804911613464,
-0.04072984308004379,
-0.10435790568590164,
-0.013246649876236916,
0.048538755625486374,
-0.03765586018562317,
-0.029524723067879677,
0.059170275926589966,
0.03120422549545765,
-0.06901096552610397,
-0.11967092752456665,
0.045992400497198105,
-0.0777064636349678,
-0.06033819913864136,
-0.06425189971923828,
-0.023782849311828613,
0.03687020391225815,
0.11357302963733673,
0.017827831208705902,
-0.0019187605939805508,
-0.015558220446109772,
0.12006805092096329,
0.0486663319170475,
-0.02432306483387947,
0.026074159890413284,
-0.06342730671167374,
0.009926093742251396
] | 0 | 0 | 0 | 0 |
|
Causal Representation Learning from Multimodal Biomedical Observations
|
https://openreview.net/forum?id=hjROBHstZ3
|
[
"Yuewen Sun",
"Lingjing Kong",
"Guangyi Chen",
"Loka Li",
"Gongxu Luo",
"Zijian Li",
"Yixuan Zhang",
"Yujia Zheng",
"Mengyue Yang",
"Petar Stojanov",
"Eran Segal",
"Eric P. Xing",
"Kun Zhang"
] |
Poster
|
Prevalent in biomedical applications (e.g., human phenotype research), multimodal datasets can provide valuable insights into the underlying physiological mechanisms. However, current machine learning (ML) models designed to analyze these datasets often lack interpretability and identifiability guarantees, which are essential for biomedical research. Recent advances in causal representation learning have shown promise in identifying interpretable latent causal variables with formal theoretical guarantees. Unfortunately, most current work on multimodal distributions either relies on restrictive parametric assumptions or yields only coarse identification results, limiting their applicability to biomedical research that favors a detailed understanding of the mechanisms.
In this work, we aim to develop flexible identification conditions for multimodal data and principled methods to facilitate the understanding of biomedical datasets. Theoretically, we consider a nonparametric latent distribution (c.f., parametric assumptions in previous work) that allows for causal relationships across potentially different modalities. We establish identifiability guarantees for each latent component, extending the subspace identification results from previous work. Our key theoretical contribution is the structural sparsity of causal connections between modalities, which, as we will discuss, is natural for a large collection of biomedical systems.
Empirically, we present a practical framework to instantiate our theoretical insights. We demonstrate the effectiveness of our approach through extensive experiments on both numerical and synthetic datasets. Results on a real-world human phenotype dataset are consistent with established biomedical research, validating our theoretical and methodological framework.
|
multimodal observations, identifiability, causal representation learning
| null | 11,612 |
2411.06518
|
[
-0.07858049869537354,
-0.10621372610330582,
0.03063441812992096,
0.027629435062408447,
0.11155504733324051,
0.04246813431382179,
-0.015713894739747047,
-0.05879901349544525,
0.02533491514623165,
-0.01623726636171341,
-0.02049490064382553,
-0.06893611699342728,
0.02262118272483349,
0.08981698006391525,
-0.035834260284900665,
-0.049036115407943726,
0.011853912845253944,
0.0624644011259079,
-0.1263483166694641,
0.03869156539440155,
0.03912356123328209,
0.010545607656240463,
0.028622610494494438,
0.029669594019651413,
0.028571873903274536,
-0.018220968544483185,
0.05380997434258461,
0.02397698163986206,
-0.03844587132334709,
-0.0025585945695638657,
0.03543582931160927,
0.0322740264236927,
-0.12502305209636688,
-0.03992748260498047,
-0.010710502043366432,
0.09585362672805786,
0.007730145938694477,
0.04875403642654419,
-0.0542081780731678,
-0.0204644612967968,
0.025372114032506943,
0.014100374653935432,
-0.05185435712337494,
0.048343073576688766,
0.03281991183757782,
0.0044838376343250275,
-0.005522608291357756,
-0.06821472197771072,
0.003444706555455923,
0.01010493841022253,
-0.07139920443296432,
-0.002250943798571825,
0.03411867097020149,
0.09042929857969284,
0.04365813359618187,
-0.04527510330080986,
0.014271737076342106,
-0.041597191244363785,
-0.046911049634218216,
0.0061641838401556015,
-0.01590660959482193,
-0.09632553160190582,
-0.038739126175642014,
-0.03499482572078705,
0.06192989647388458,
0.008021263405680656,
0.00234976876527071,
0.022629890590906143,
-0.05647525563836098,
0.029081104323267937,
-0.04581623151898384,
0.004238865338265896,
-0.08238399028778076,
0.04545493796467781,
-0.042158566415309906,
0.010875283740460873,
0.06623037904500961,
0.03708019107580185,
0.04582486301660538,
-0.12261103838682175,
0.0838160365819931,
0.08042313903570175,
0.008747142739593983,
-0.008560328744351864,
0.09677580744028091,
0.0022771265357732773,
-0.03761345520615578,
0.050910621881484985,
-0.03403233736753464,
0.1144225150346756,
-0.006501003168523312,
-0.000050010061386274174,
0.024659017100930214,
-0.019005082547664642,
0.05568842217326164,
-0.018032211810350418,
0.06954819709062576,
-0.04902792349457741,
0.08028431981801987,
0.09893769770860672,
-0.03845936805009842,
-0.002134085400030017,
0.0026950472965836525,
0.07906030118465424,
0.05848009139299393,
-0.09218659996986389,
0.0023019437212496996,
0.028074968606233597,
0.03875670209527016,
-0.03530142083764076,
0.029660437256097794,
0.01282061729580164,
0.00870265532284975,
0.03198738023638725,
0.005702152382582426,
-0.062388379126787186,
0.026849942281842232,
0.028644725680351257,
0.10738509148359299,
-0.05014360323548317,
-0.04148649051785469,
-0.07437864691019058,
0.028508802875876427,
-0.02351083792746067,
0.04772686958312988,
-0.04249383136630058,
-0.06609377264976501,
1.3605060495466804e-33,
0.0977674052119255,
-0.004877736326307058,
0.07236023247241974,
0.029934905469417572,
0.021497994661331177,
-0.017861098051071167,
-0.08731050044298172,
-0.026032323017716408,
0.0017744352808222175,
-0.038186803460121155,
-0.0010099143255501986,
0.008261051028966904,
-0.1043156236410141,
0.038798533380031586,
-0.01694754883646965,
0.07621786743402481,
-0.06927424669265747,
0.07250159233808517,
0.011813024990260601,
-0.018098387867212296,
0.0064969295635819435,
-0.10523883253335953,
-0.06952659785747528,
0.03609798476099968,
-0.007438968401402235,
-0.03245745599269867,
0.041802600026130676,
-0.05378308147192001,
-0.04883221164345741,
0.016357287764549255,
-0.12265262007713318,
-0.0008797726477496326,
0.09227758646011353,
-0.02996012754738331,
0.02490161545574665,
-0.06446348875761032,
-0.04379938170313835,
-0.050910744816064835,
-0.014629333280026913,
-0.003660286543890834,
0.0256219319999218,
-0.04840251803398132,
0.03358263894915581,
-0.03343334421515465,
-0.08049871772527695,
-0.04097661375999451,
0.013882753439247608,
-0.022428404539823532,
-0.0520157665014267,
-0.09162535518407822,
0.054463278502225876,
0.052893999963998795,
0.08744408190250397,
-0.15770509839057922,
-0.02613123133778572,
0.012437941506505013,
-0.061294909566640854,
0.011209163814783096,
-0.013575567863881588,
-0.020935380831360817,
-0.00004181227632216178,
0.04613455757498741,
0.02640407718718052,
0.046008817851543427,
0.027994489297270775,
0.04260743409395218,
-0.11069031804800034,
-0.02239661104977131,
0.07977929711341858,
0.009790784679353237,
-0.06593120098114014,
0.03446915000677109,
-0.06784122437238693,
-0.06228581815958023,
0.011694102548062801,
0.01761232316493988,
-0.001487384899519384,
-0.02127021737396717,
0.021577415987849236,
0.031978052109479904,
0.014265933074057102,
-0.032332006841897964,
-0.0011807888513430953,
0.018512040376663208,
-0.03786567226052284,
0.01751994527876377,
0.029635822400450706,
-0.07656125724315643,
-0.06999778747558594,
-0.08763109147548676,
-0.055978719145059586,
0.02840281091630459,
-0.05144297331571579,
0.005872261710464954,
-0.01758229359984398,
-3.171643184939862e-33,
-0.019602222368121147,
0.022584939375519753,
0.04190216585993767,
-0.06391904503107071,
0.06247678026556969,
0.02101815491914749,
0.014679995365440845,
0.006930914707481861,
0.04285697638988495,
-0.03577403724193573,
0.04838145896792412,
-0.059634480625391006,
0.03544257953763008,
-0.03756241127848625,
-0.008268149569630623,
0.09526833146810532,
-0.021320480853319168,
0.05172836035490036,
0.00968283973634243,
0.0023125139996409416,
0.030600564554333687,
-0.026717497035861015,
0.011667702347040176,
0.051615551114082336,
-0.013433489017188549,
0.028434284031391144,
-0.06381373852491379,
0.04417906329035759,
-0.0508694052696228,
0.004223338793963194,
-0.04978708177804947,
-0.01824486255645752,
-0.04945668950676918,
-0.051830556243658066,
-0.026995833963155746,
0.029545847326517105,
0.046758607029914856,
0.0005764268571510911,
-0.04178459197282791,
0.04367714002728462,
0.10927864164113998,
0.03445856273174286,
-0.0878683477640152,
0.036504730582237244,
0.04102202504873276,
-0.1028982549905777,
-0.016045715659856796,
0.0013393718982115388,
0.06966840475797653,
0.053092606365680695,
-0.01730213314294815,
0.015114977955818176,
-0.002834504935890436,
0.059486642479896545,
-0.048923034220933914,
-0.13478808104991913,
0.029005711898207664,
-0.07395310699939728,
0.07299800217151642,
-0.00339406356215477,
-0.07362473011016846,
-0.05818457528948784,
-0.05656919255852699,
0.06794173270463943,
0.042175762355327606,
-0.05978741496801376,
0.07215996086597443,
-0.07532180100679398,
-0.0030610165558755398,
0.006362662650644779,
0.0487649068236351,
0.02782684937119484,
-0.02762024663388729,
-0.013878131285309792,
0.06463666260242462,
0.026713447645306587,
0.011392067186534405,
-0.036731183528900146,
-0.07999080419540405,
-0.00950915738940239,
0.09808406233787537,
-0.02900204434990883,
0.07450442016124725,
0.005282558035105467,
0.0009483182802796364,
0.09309002012014389,
0.10996361076831818,
0.026817668229341507,
0.07570844143629074,
-0.013717340305447578,
-0.008645561523735523,
0.03104475513100624,
-0.06250050663948059,
0.0528297945857048,
-0.05914612486958504,
-3.849386231991048e-8,
0.028140394017100334,
0.0011839522048830986,
-0.008702118881046772,
-0.08925694972276688,
0.06570839136838913,
-0.006826362572610378,
-0.051667626947164536,
0.03145235404372215,
0.05029004067182541,
-0.0007525240071117878,
0.06338031589984894,
-0.011943451128900051,
0.03967234119772911,
-0.06023269519209862,
-0.014967352151870728,
0.021567517891526222,
-0.02551654726266861,
0.011156978085637093,
-0.03922922909259796,
0.018209243193268776,
0.02645678073167801,
0.0024184987414628267,
0.01605258695781231,
-0.005373642314225435,
0.11968771368265152,
-0.0686902180314064,
-0.017816301435232162,
-0.00000783037921792129,
-0.02038392424583435,
-0.020931512117385864,
-0.08354457467794418,
0.03593843802809715,
0.036035265773534775,
-0.0399504080414772,
-0.03279763087630272,
0.044882528483867645,
0.05100761353969574,
-0.05538414046168327,
-0.1263820230960846,
-0.028106046840548515,
-0.044812001287937164,
0.022224804386496544,
-0.08175326138734818,
0.02219526842236519,
0.09994787722826004,
-0.021234342828392982,
0.012766500003635883,
-0.02116936631500721,
0.00045473725185729563,
0.020617924630641937,
0.018137028440833092,
0.005990731064230204,
0.09302526712417603,
0.07657530158758163,
0.03736318275332451,
0.017227835953235626,
0.03489323705434799,
0.04084310308098793,
0.06486591696739197,
-0.029546691104769707,
0.099730484187603,
0.05868469923734665,
-0.05554581433534622,
-0.017525818198919296
] | 0 | 0 | 0 | 0 |
|
On the Learn-to-Optimize Capabilities of Transformers in In-Context Sparse Recovery
|
https://openreview.net/forum?id=NHhjczmJjo
|
[
"Renpu Liu",
"Ruida Zhou",
"Cong Shen",
"Jing Yang"
] |
Poster
|
An intriguing property of the Transformer is its ability to perform in-context learning (ICL), where the Transformer can solve different inference tasks without parameter updating based on the contextual information provided by the corresponding input-output demonstration pairs. It has been theoretically proved that ICL is enabled by the capability of Transformers to perform gradient-descent algorithms (Von Oswald et al., 2023a; Bai et al., 2024). This work takes a step further and shows that Transformers can perform learning-to-optimize (L2O) algorithms. Specifically, for the ICL sparse recovery (formulated as LASSO) tasks, we show that a K-layer Transformer can perform an L2O algorithm with a provable convergence rate linear in K. This provides a new perspective explaining the superior ICL capability of Transformers, even with only a few layers, which cannot be achieved by the standard gradient-descent algorithms. Moreover, unlike the conventional L2O algorithms that require the measurement matrix involved in training to match that in testing, the trained Transformer is able to solve sparse recovery problems generated with different measurement matrices. Besides, Transformers as an L2O algorithm can leverage structural information embedded in the training tasks to accelerate its convergence during ICL, and generalize across different lengths of demonstration pairs, where conventional L2O algorithms typically struggle or fail. Such theoretical findings are supported by our experimental results.
|
Transformer, In-context learning, Learning-to-optimize
| null | 11,604 |
2410.13981
|
[
-0.1472441554069519,
0.004687923938035965,
-0.016926677897572517,
0.029619699344038963,
0.06518080085515976,
0.005989741533994675,
0.036119792610406876,
0.0010254575172439218,
-0.02577255107462406,
-0.0031540635973215103,
-0.01817682757973671,
0.0688680112361908,
0.04959946498274803,
0.05420757085084915,
0.005561934784054756,
0.011491354554891586,
0.034271642565727234,
0.09235644340515137,
-0.1097719743847847,
-0.06385696679353714,
0.045300886034965515,
0.0203611571341753,
-0.01364363357424736,
-0.00751864118501544,
0.04254017770290375,
0.01220646221190691,
0.046650879085063934,
0.0028036560397595167,
-0.015971167013049126,
0.04098736494779587,
-0.013039626181125641,
0.008490687236189842,
-0.07239328324794769,
-0.01986776478588581,
0.005919858813285828,
0.07938428223133087,
-0.0203629732131958,
0.06510834395885468,
0.025937413796782494,
-0.025253888219594955,
0.009806733578443527,
-0.010500170290470123,
-0.04151412844657898,
-0.0027607812080532312,
0.055039867758750916,
0.02274216152727604,
0.05532781407237053,
-0.060048360377550125,
0.037912365049123764,
-0.04492795467376709,
0.029239945113658905,
-0.007814708165824413,
-0.03575749695301056,
0.06407980620861053,
-0.01974072866141796,
0.06589876115322113,
-0.006752702873200178,
0.003334254026412964,
-0.028624553233385086,
0.04311983659863472,
-0.052287813276052475,
-0.10043586045503616,
-0.04241017997264862,
0.04473879188299179,
-0.07967782020568848,
0.016378389671444893,
0.09118557721376419,
-0.000595079327467829,
-0.053005293011665344,
-0.038954079151153564,
-0.10246019065380096,
0.02484579011797905,
-0.015540503896772861,
0.07355810701847076,
0.10628253221511841,
-0.02371843345463276,
0.120050348341465,
-0.03933820128440857,
0.03555930033326149,
0.01964988000690937,
0.041416652500629425,
0.04405605047941208,
-0.035407278686761856,
-0.014770749025046825,
0.04960326850414276,
-0.0348762646317482,
-0.04040426015853882,
0.007046265061944723,
0.004789319820702076,
0.02003355883061886,
-0.021585367619991302,
-0.05110837519168854,
-0.0543862022459507,
0.011619522236287594,
0.01763494685292244,
-0.0409226194024086,
0.0456005223095417,
-0.02936464548110962,
-0.04324468970298767,
0.1043083444237709,
-0.014298807829618454,
-0.045497696846723557,
0.012448344379663467,
-0.025314930826425552,
-0.05800135061144829,
-0.05306060239672661,
0.01501113548874855,
0.057636406272649765,
0.03053201548755169,
-0.044065676629543304,
0.040491219609975815,
0.034451182931661606,
0.045432884246110916,
0.013742074370384216,
0.08273783326148987,
-0.0662565603852272,
0.0489668995141983,
-0.02519974857568741,
-0.02998081035912037,
0.07755141705274582,
-0.015721483156085014,
-0.03906342014670372,
0.005597342737019062,
0.043263763189315796,
0.00042771935113705695,
0.04469887539744377,
-0.010818284004926682,
1.844599326963765e-33,
-0.010621738620102406,
0.09432808309793472,
0.06026647612452507,
-0.02491530403494835,
-0.03844844549894333,
0.0011543546570464969,
0.063469797372818,
0.07840605825185776,
0.04766643047332764,
0.018379878252744675,
-0.01662524975836277,
0.03611080348491669,
-0.05534418672323227,
0.05070611089468002,
-0.0459127277135849,
0.019858624786138535,
-0.015413760207593441,
0.07453659921884537,
-0.007033552974462509,
-0.03981022164225578,
0.038091178983449936,
-0.04793555289506912,
0.059031013399362564,
-0.06277889013290405,
-0.025654463097453117,
-0.029871825128793716,
0.10850794613361359,
-0.05596085265278816,
-0.022063227370381355,
-0.014720953069627285,
-0.10062667727470398,
0.020338481292128563,
0.010684932582080364,
0.05107077211141586,
0.00482760788872838,
0.011542072519659996,
0.03202541545033455,
-0.035968173295259476,
-0.04580323398113251,
0.01667194440960884,
-0.033317577093839645,
0.025438982993364334,
-0.0006287893047556281,
-0.042868293821811676,
-0.08921241760253906,
-0.06494709849357605,
-0.009366682730615139,
-0.026043854653835297,
0.02156282588839531,
-0.04535968229174614,
0.00984324049204588,
-0.023234961554408073,
-0.06527045369148254,
0.0037606977857649326,
0.11497346311807632,
0.04507923871278763,
0.011516007594764233,
0.011122758500277996,
0.07926274836063385,
0.0123068168759346,
-0.0032072155736386776,
-0.026770276948809624,
-0.04493909329175949,
0.06885700672864914,
0.02761954627931118,
0.013007352128624916,
0.008292919024825096,
0.034272439777851105,
-0.015882229432463646,
-0.05197998136281967,
-0.0040906574577093124,
0.015489513985812664,
0.02688637748360634,
-0.09475430101156235,
0.14718060195446014,
-0.014375973492860794,
0.09088917076587677,
-0.011876128613948822,
-0.028572414070367813,
-0.0048145512118935585,
-0.016162004321813583,
0.012836232781410217,
0.05276120826601982,
-0.0062361364252865314,
-0.040286555886268616,
0.0590796172618866,
0.0062677739188075066,
-0.08946878463029861,
-0.02506263740360737,
-0.14214268326759338,
-0.12474854290485382,
-0.00298137916252017,
-0.06831429153680801,
-0.04896724596619606,
0.027835283428430557,
-1.3482447251126424e-33,
0.04129001870751381,
0.013376589864492416,
-0.020142029970884323,
0.00399662833660841,
0.0023801226634532213,
-0.056817010045051575,
-0.029081180691719055,
0.005301873665302992,
-0.035540878772735596,
-0.04858312010765076,
-0.01688741333782673,
-0.025181852281093597,
0.011606697924435139,
-0.01949341781437397,
0.024829896166920662,
0.04512729495763779,
-0.11389230936765671,
-0.056876491755247116,
0.012278227135539055,
0.028829187154769897,
0.03999108821153641,
0.0704071968793869,
-0.05009184777736664,
-0.022262487560510635,
-0.08830030262470245,
-0.016385186463594437,
-0.09615558385848999,
0.1413641721010208,
0.08399275690317154,
0.010075632482767105,
0.023252438753843307,
-0.0672733336687088,
-0.014270554296672344,
-0.025345198810100555,
0.009859742596745491,
0.06377188116312027,
-0.02391970343887806,
0.0341854952275753,
-0.056070320308208466,
0.10218382626771927,
0.006046129856258631,
0.002844082424417138,
-0.08026132732629776,
0.012447171844542027,
0.002918301150202751,
-0.06635145843029022,
0.04798968881368637,
-0.10036889463663101,
0.04525351896882057,
0.04391227662563324,
0.0334450900554657,
0.013905134983360767,
-0.11701284348964691,
-0.0036429755855351686,
-0.036298610270023346,
-0.03869478404521942,
0.021892838180065155,
0.001388414530083537,
0.09731796383857727,
-0.04839316010475159,
-0.0344075970351696,
-0.02172025851905346,
-0.05091293901205063,
-0.0021216939203441143,
0.02177240699529648,
0.01989663951098919,
0.0063148727640509605,
0.02362886629998684,
-0.01510945986956358,
0.11835143715143204,
-0.009722817689180374,
0.061242733150720596,
0.052722446620464325,
-0.06434570997953415,
0.06490251421928406,
-0.030844880267977715,
-0.029196275398135185,
-0.033735644072294235,
-0.04639144614338875,
-0.03573109582066536,
0.012628275901079178,
-0.07233034819364548,
-0.011497650295495987,
0.05137743055820465,
0.03387666866183281,
0.07106342911720276,
0.055615074932575226,
0.022311976179480553,
0.059054866433143616,
-0.04666081815958023,
-0.05262329801917076,
0.03786720708012581,
-0.0023415519390255213,
0.012923714704811573,
-0.02624925598502159,
-4.3251340287042694e-8,
-0.10891617834568024,
0.0842442661523819,
-0.054955512285232544,
-0.03783634677529335,
0.04294496402144432,
-0.03273671492934227,
0.038192685693502426,
0.10151726007461548,
-0.016485363245010376,
-0.05191824212670326,
0.04195236414670944,
-0.06936627626419067,
-0.040689192712306976,
-0.009539778344333172,
-0.03552723675966263,
0.07388578355312347,
0.01648460328578949,
0.0469675175845623,
0.03032272681593895,
0.026895074173808098,
-0.0027060264255851507,
0.03813815116882324,
0.018526164814829826,
-0.047141119837760925,
0.06276489794254303,
-0.09542069584131241,
-0.003345940262079239,
0.05277044326066971,
0.08622431010007858,
0.005836149211972952,
-0.14803777635097504,
-0.0038972683250904083,
0.09159800410270691,
-0.041131407022476196,
0.02524275705218315,
0.15016832947731018,
0.061697449535131454,
-0.060538772493600845,
-0.03177245333790779,
-0.06851541250944138,
-0.023249275982379913,
0.05768050253391266,
-0.023517711088061333,
0.01435857079923153,
0.01753183640539646,
-0.007253458723425865,
0.062336891889572144,
-0.08795972168445587,
0.029017511755228043,
-0.06700718402862549,
0.031184490770101547,
0.028968511149287224,
-0.023702817037701607,
0.04027222469449043,
0.08054637163877487,
0.011651741340756416,
-0.0010667323367670178,
-0.11083697527647018,
0.030814535915851593,
0.11364393681287766,
-0.061725206673145294,
0.03616805002093315,
-0.06277389824390411,
0.012126428075134754
] | 0 | 0 | 0 | 0 |
|
OvercookedV2: Rethinking Overcooked for Zero-Shot Coordination
|
https://openreview.net/forum?id=hlvLM3GX8R
|
[
"Tobias Gessler",
"Tin Dizdarevic",
"Ani Calinescu",
"Benjamin Ellis",
"Andrei Lupu",
"Jakob Nicolaus Foerster"
] |
Poster
|
AI agents hold the potential to transform everyday life by helping humans achieve their goals.
To do this successfully, agents need to be able to coordinate with novel partners without prior interaction, a setting known as zero-shot coordination (ZSC).
Overcooked has become one of the most popular benchmarks for evaluating coordination capabilities of AI agents and learning algorithms.
In this work, we investigate the origins of ZSC challenges in Overcooked.
We introduce a state augmentation mechanism which mixes states that might be encountered when paired with unknown partners into the training distribution, reducing the out-of-distribution challenge associated with ZSC.
We show that independently trained agents under this algorithm coordinate successfully in Overcooked.
Our results suggest that ZSC failure can largely be attributed to poor state coverage under self-play rather than more sophisticated coordination challenges. The Overcooked environment is therefore not suitable as a ZSC benchmark.
To address these shortcomings, we introduce OvercookedV2, a new version of the benchmark, which includes asymmetric information and stochasticity, facilitating the creation of interesting ZSC scenarios.
To validate OvercookedV2, we conduct experiments demonstrating that mere exhaustive state coverage is insufficient to coordinate well. Finally, we use OvercookedV2 to build a new range of coordination challenges, including ones that require test time protocol formation, and we demonstrate the need for new coordination algorithms that can adapt online.
We hope that OvercookedV2 will help benchmark the next generation of ZSC algorithms and advance collaboration between AI agents and humans.
|
multi-agent reinforcement learning, reinforcement learning, multi-agent systems, zero-shot coordination, overcooked, human-AI coordination
| null | 11,597 |
2503.17821
|
[
-0.05241582170128822,
-0.034124571830034256,
0.006398734170943499,
0.05053164064884186,
0.0712687149643898,
-0.01140519231557846,
0.006089965812861919,
-0.044732946902513504,
-0.013913006521761417,
0.020486190915107727,
-0.03966432809829712,
-0.07400378584861755,
0.009860907681286335,
0.03887563943862915,
-0.0016932542202994227,
-0.05346129089593887,
0.11140584945678711,
-0.004355435725301504,
-0.049821820110082626,
-0.012603445909917355,
0.005702522117644548,
0.01947093941271305,
0.03238780051469803,
-0.07154678553342819,
0.02641856111586094,
-0.033492475748062134,
-0.01599525660276413,
-0.08271130919456482,
-0.031745217740535736,
-0.01783718355000019,
0.09010575711727142,
-0.054597433656454086,
-0.010854154825210571,
-0.014672422781586647,
-0.03516588360071182,
-0.011810018680989742,
-0.007861047983169556,
-0.03661182150244713,
-0.011792628094553947,
0.0018708754796534777,
0.01619996875524521,
0.059889648109674454,
0.02792130783200264,
-0.002258356660604477,
0.018055912107229233,
0.12176144868135452,
-0.015465019270777702,
-0.09009800851345062,
0.03892350569367409,
0.0225068349391222,
-0.13822747766971588,
-0.02518904022872448,
0.0023687391076236963,
0.05595124512910843,
0.05512799695134163,
0.07353615760803223,
-0.02637062408030033,
0.045820366591215134,
-0.032370809465646744,
-0.040942154824733734,
0.006162576377391815,
-0.1307254135608673,
0.046439193189144135,
0.02513512596487999,
0.008377453312277794,
-0.01284715998917818,
0.004424292594194412,
0.06861020624637604,
-0.031529080122709274,
-0.04671010002493858,
-0.021111024543642998,
0.09571051597595215,
0.007880939170718193,
0.02586452290415764,
0.08113214373588562,
0.03329968452453613,
-0.03009490855038166,
-0.08762697875499725,
0.037268318235874176,
-0.04776820167899132,
-0.01807251200079918,
-0.04483478516340256,
-0.027369040995836258,
0.05615733936429024,
0.08635136485099792,
-0.04101899638772011,
-0.03146674856543541,
0.031624797731637955,
0.08430395275354385,
0.04411949962377548,
-0.0361190140247345,
-0.0012632732978090644,
0.011987393721938133,
-0.015140312723815441,
0.03407694399356842,
0.11761299520730972,
-0.005227156449109316,
-0.06105099245905876,
0.012948345392942429,
0.06813409924507141,
-0.08646208047866821,
-0.005169524345546961,
0.01414826326072216,
-0.025841258466243744,
0.03838416561484337,
0.024787215515971184,
0.02658475749194622,
-0.01258285902440548,
0.015679653733968735,
-0.03169197961688042,
0.008295024745166302,
0.06354714184999466,
0.0743648037314415,
-0.061595506966114044,
-0.030436713248491287,
0.08301623165607452,
-0.05145498365163803,
0.04307292401790619,
-0.04265664517879486,
-0.030643280595541,
0.02348187193274498,
-0.06709960848093033,
0.005235210992395878,
0.033117830753326416,
0.002444683574140072,
0.07105373591184616,
0.014607171528041363,
2.6632308579496837e-33,
-0.051062583923339844,
0.045931003987789154,
0.03459294140338898,
0.03042285144329071,
0.027694614604115486,
-0.037277013063430786,
0.0067511689849197865,
-0.00410498445853591,
-0.04201430082321167,
-0.026088103652000427,
0.01583247072994709,
-0.02453395538032055,
-0.03865877538919449,
0.03703612461686134,
0.06115494668483734,
-0.07709860801696777,
-0.000970425782725215,
0.04584389179944992,
0.006916587706655264,
0.01860133185982704,
0.1133943423628807,
-0.08162981271743774,
-0.007148340344429016,
-0.07008668780326843,
0.028760261833667755,
0.06938717514276505,
0.01817806251347065,
-0.06950832903385162,
0.021718956530094147,
-0.010598551481962204,
-0.05264567211270332,
0.049416590481996536,
-0.03878609463572502,
0.015794675797224045,
0.06650876998901367,
-0.03939959406852722,
-0.047906436026096344,
-0.053315721452236176,
-0.0362611785531044,
0.04162085801362991,
-0.03606506064534187,
0.03901289775967598,
-0.03933704271912575,
-0.014202361926436424,
-0.0530729815363884,
-0.07882498949766159,
0.026309754699468613,
0.014861995354294777,
-0.05165459215641022,
-0.07767657935619354,
-0.0196639746427536,
0.009901107288897038,
-0.006185024511069059,
-0.06482765078544617,
-0.015612227842211723,
0.06948406994342804,
0.049609456211328506,
-0.06633522361516953,
-0.019321244210004807,
0.024733025580644608,
0.06706350296735764,
0.042496271431446075,
-0.053189657628536224,
0.10385852307081223,
0.013773377053439617,
0.030111318454146385,
0.060993753373622894,
0.04140876233577728,
0.09733675420284271,
-0.004862629342824221,
0.01084434799849987,
0.0019083591178059578,
-0.015056618489325047,
-0.0797804594039917,
0.05562406778335571,
-0.002633869182318449,
0.08081323653459549,
-0.0767907202243805,
0.03424052894115448,
-0.038092561066150665,
-0.01407902967184782,
-0.007712924852967262,
-0.06432300806045532,
0.021571993827819824,
-0.054048940539360046,
-0.03205907344818115,
-0.08513258397579193,
-0.04425463452935219,
-0.0729409009218216,
0.07698458433151245,
-0.08488783985376358,
-0.028357356786727905,
-0.037346526980400085,
-0.026852596551179886,
-0.04220442473888397,
-1.6117039576941223e-33,
-0.03589174896478653,
-0.0003342882264405489,
-0.061445869505405426,
0.0029102088883519173,
0.007051764987409115,
-0.032006774097681046,
-0.03593411296606064,
-0.0960097536444664,
0.037537116557359695,
-0.03481733426451683,
-0.06704634428024292,
0.01798563078045845,
0.0748993307352066,
-0.012617654167115688,
0.0012188291875645518,
-0.04508295655250549,
-0.05206410214304924,
0.07920397073030472,
0.0397142618894577,
-0.04549141600728035,
0.06702406704425812,
0.03762932866811752,
-0.03279261291027069,
-0.047866083681583405,
0.054161734879016876,
0.07139753550291061,
-0.007935509085655212,
0.12543153762817383,
0.021579492837190628,
0.012422783300280571,
-0.00388735206797719,
-0.03250915929675102,
0.022645078599452972,
-0.048038288950920105,
0.01000603660941124,
0.159190371632576,
-0.008822002448141575,
-0.029582122340798378,
-0.055198267102241516,
0.04739391431212425,
0.10257842391729355,
-0.030898377299308777,
-0.16062453389167786,
0.07242461293935776,
-0.04178817570209503,
0.013986370526254177,
-0.04251256585121155,
0.04224653169512749,
0.015522917732596397,
0.0076993475668132305,
-0.0419011227786541,
0.029759179800748825,
-0.10519374907016754,
-0.05796366184949875,
-0.03501786291599274,
-0.010914001613855362,
-0.004883516579866409,
0.0020417077466845512,
0.08252234011888504,
-0.040617503225803375,
-0.023787591606378555,
0.007755680941045284,
0.00691920006647706,
0.059175267815589905,
0.03269561752676964,
0.0770341008901596,
0.020546913146972656,
0.0663076788187027,
0.062143340706825256,
-0.011406892910599709,
0.03417101129889488,
0.035593487322330475,
0.022808915004134178,
-0.034067824482917786,
-0.000926109729334712,
0.036306142807006836,
-0.07270415872335434,
-0.006174622103571892,
0.022060055285692215,
-0.009001573547720909,
-0.041597288101911545,
-0.011333351023495197,
0.01163865253329277,
0.11075717955827713,
0.12079408764839172,
0.13630902767181396,
-0.019961869344115257,
0.08847691118717194,
-0.011811435222625732,
0.008993856608867645,
-0.05413133278489113,
0.040928974747657776,
0.04291684553027153,
0.047203049063682556,
-0.1218600943684578,
-4.221060123654752e-8,
-0.09746001660823822,
-0.014602544717490673,
0.08570631593465805,
0.0070161838084459305,
0.110210120677948,
-0.013456316664814949,
-0.03872134909033775,
-0.003250179812312126,
-0.03161013498902321,
0.018829351291060448,
0.04471563175320625,
-0.0013718304689973593,
0.042811471968889236,
-0.047362878918647766,
0.0578337125480175,
0.05709194019436836,
-0.05192851647734642,
-0.013634982518851757,
-0.07148651033639908,
0.044336747378110886,
0.05935923010110855,
-0.0025357299018651247,
-0.07805036753416061,
-0.0118286507204175,
0.009685692377388477,
-0.07154877483844757,
-0.043741900473833084,
0.04112721234560013,
0.06591983884572983,
0.028825579211115837,
-0.00789474043995142,
-0.04514027014374733,
0.05639158561825752,
0.019904080778360367,
0.005217861384153366,
0.0664973333477974,
-0.0009155283914878964,
-0.04757469519972801,
0.0005415115156210959,
-0.04907545447349548,
-0.0927010178565979,
0.1011752337217331,
-0.024661801755428314,
0.03723425045609474,
-0.027845030650496483,
0.022133011370897293,
-0.0448768213391304,
-0.07071821391582489,
0.01789364218711853,
-0.04468601197004318,
-0.06230880320072174,
-0.006562367547303438,
-0.04342123866081238,
0.1087370365858078,
0.05290502309799194,
0.0462016835808754,
-0.013605793006718159,
-0.025459682568907738,
0.06401995569467545,
0.09268779307603836,
0.016737768426537514,
0.0163025613874197,
-0.09859969466924667,
0.001619531773030758
] |
https://github.com/overcookedv2/experiments
| 1 | 0 | 0 | 0 |
Sketching for Convex and Nonconvex Regularized Least Squares with Sharp Guarantees
|
https://openreview.net/forum?id=7liN6uHAQZ
|
[
"Yingzhen Yang",
"Ping Li"
] |
Poster
|
Randomized algorithms play a crucial role in efficiently solving large-scale optimization problems. In this paper, we introduce Sketching for Regularized Optimization (SRO), a fast sketching algorithm designed for least squares problems with convex or nonconvex regularization. SRO operates by first creating a sketch of the original data matrix and then solving the sketched problem. We establish minimax optimal rates for sparse signal estimation by addressing the sketched sparse convex and nonconvex learning problems. Furthermore, we propose a novel Iterative SRO algorithm, which reduces the approximation error geometrically for sketched convex regularized problems. To the best of our knowledge, this work is among the first to provide a unified theoretical framework demonstrating minimax rates for convex and nonconvex sparse learning problems via sketching. Experimental results validate the efficiency and effectiveness of both the SRO and Iterative SRO algorithms.
|
Sketching, Random Projection, Minimax Rates
|
We present novel theoretical results for the approximation error between the optimization results of the original problem and the sketched problem for regularized least square problems with sharp guarantees.
| 11,596 |
2311.01806
|
[
-0.006284938659518957,
-0.015538942068815231,
0.02023833431303501,
-0.03229077532887459,
0.06996385753154755,
0.02582208625972271,
0.009802298620343208,
-0.045503634959459305,
0.001317485817708075,
0.035991813987493515,
-0.020914718508720398,
0.08298046141862869,
0.01562877558171749,
0.09237827360630035,
-0.03493182361125946,
-0.0014259237796068192,
0.03724496439099312,
0.01621171087026596,
-0.03605439513921738,
0.015732817351818085,
0.012367963790893555,
-0.062021877616643906,
-0.008733002468943596,
-0.020598288625478745,
0.10380348563194275,
-0.0410458967089653,
0.08971980959177017,
0.026566281914711,
-0.005850484129041433,
0.0782909020781517,
-0.015803849324584007,
0.00018800263933371753,
0.0702570453286171,
-0.0698806568980217,
0.05781825631856918,
0.04881054908037186,
-0.0016018003225326538,
0.10909777134656906,
-0.0350162535905838,
0.0485454797744751,
0.012938817963004112,
0.05052730813622475,
-0.024792514741420746,
0.004548229277133942,
0.051341939717531204,
-0.09403657913208008,
0.034028310328722,
-0.09477519243955612,
0.049169279634952545,
0.02764216810464859,
0.004420863464474678,
-0.03146816045045853,
-0.08186861872673035,
-0.035139359533786774,
-0.015042250044643879,
-0.010599060915410519,
-0.00684187700971961,
-0.04997577890753746,
0.06532039493322372,
0.006443444639444351,
-0.02784498780965805,
-0.14390917122364044,
0.025912132114171982,
0.009207579307258129,
0.009048407897353172,
-0.014933260157704353,
0.027332386001944542,
0.05013345927000046,
-0.0012669923016801476,
0.05455836281180382,
-0.046139732003211975,
0.06476522237062454,
-0.00860051903873682,
0.03522329032421112,
0.10500065982341766,
-0.09570233523845673,
0.04547756910324097,
0.02444264106452465,
0.051137518137693405,
-0.11655984818935394,
-0.07582790404558182,
0.0606597438454628,
-0.011093992739915848,
0.001889238366857171,
0.05256122350692749,
0.04551870375871658,
0.026030568405985832,
-0.023232949897646904,
0.015187080949544907,
-0.017806842923164368,
-0.058520760387182236,
-0.006920107174664736,
-0.08796264976263046,
-0.0015482032904401422,
-0.05452065169811249,
-0.06603363156318665,
0.1004926934838295,
-0.033450234681367874,
-0.04040444642305374,
0.0942104235291481,
0.022500373423099518,
-0.027982104569673538,
0.025703534483909607,
0.05561463534832001,
0.006484432611614466,
-0.023751450702548027,
0.010180826298892498,
0.08777306228876114,
0.014745540916919708,
-0.05711457133293152,
0.08122069388628006,
0.026342900469899178,
0.013705584220588207,
0.057268280535936356,
-0.00033821919350884855,
-0.0961952805519104,
0.05151086673140526,
-0.06487511098384857,
0.04761577025055885,
0.05974917486310005,
-0.0833970382809639,
-0.034985460340976715,
0.040613140910863876,
0.017928490415215492,
-0.0061079999431967735,
0.040267977863550186,
-0.0032225996255874634,
4.1534259849802216e-33,
0.03444322198629379,
0.0872860699892044,
0.020068012177944183,
-0.07361959666013718,
0.01682812161743641,
0.01230600569397211,
0.013316682539880276,
-0.005224188324064016,
0.06252563744783401,
0.049938272684812546,
-0.06687664985656738,
-0.06551674753427505,
-0.02981443889439106,
0.08076383173465729,
-0.0033730873838067055,
-0.006894632708281279,
-0.029096165671944618,
0.031986039131879807,
0.03523050248622894,
-0.07087747007608414,
-0.00004410616384120658,
-0.0654754787683487,
-0.02410043589770794,
-0.03230172023177147,
-0.012846049852669239,
-0.016769761219620705,
0.02884749323129654,
-0.05999713018536568,
-0.03416578844189644,
0.023285306990146637,
-0.0065956125035882,
0.07028137147426605,
0.05883686617016792,
-0.01872764341533184,
-0.004817869048565626,
0.007417946122586727,
0.07396606355905533,
-0.03205251321196556,
-0.021568892523646355,
0.00939200259745121,
-0.03918632119894028,
0.006992849987000227,
0.0359344407916069,
-0.0841895341873169,
0.008321696892380714,
-0.12065862119197845,
0.0086133424192667,
0.05830535292625427,
-0.007828425616025925,
-0.04489343240857124,
0.03778453171253204,
-0.024303099140524864,
-0.03897569328546524,
0.009354913607239723,
-0.0003835881070699543,
-0.051222626119852066,
-0.008991331793367863,
-0.10836096107959747,
0.08362162113189697,
0.07509804517030716,
-0.003354162210598588,
0.008735018782317638,
-0.0399358756840229,
0.013093402609229088,
-0.028188325464725494,
0.0631871223449707,
0.04395623132586479,
-0.04367837309837341,
0.012353107333183289,
-0.032887134701013565,
-0.04662492126226425,
0.03358648717403412,
0.03997528553009033,
-0.10354772210121155,
0.06368459016084671,
0.0071835024282336235,
0.06100619211792946,
0.030690722167491913,
-0.030327415093779564,
-0.024727851152420044,
0.009852460585534573,
0.01071710605174303,
-0.023143289610743523,
-0.13778088986873627,
-0.05807149037718773,
-0.0035195520613342524,
0.09672912955284119,
-0.15073247253894806,
-0.0628594383597374,
0.0023674038238823414,
-0.09834899008274078,
0.06123559549450874,
-0.05719337612390518,
0.0024538999423384666,
-0.013955355621874332,
-2.0379152504282442e-33,
-0.027042381465435028,
0.0606561005115509,
0.013969192281365395,
0.004476688336580992,
0.0149652985855937,
-0.08734152466058731,
-0.034818753600120544,
-0.03299493342638016,
0.022421617060899734,
-0.08718826621770859,
0.03620542213320732,
-0.08963590115308762,
0.027226535603404045,
-0.01744954288005829,
0.041388195008039474,
0.09531662613153458,
-0.026546161621809006,
0.037988871335983276,
-0.030374599620699883,
-0.02816014178097248,
-0.03222113102674484,
0.009571477770805359,
0.002432648791000247,
0.05801898241043091,
0.012621363624930382,
-0.0024562685284763575,
-0.04344240203499794,
0.12196490913629532,
-0.01741110160946846,
-0.05209711939096451,
-0.014387632720172405,
-0.019576545804739,
-0.07908543199300766,
-0.025373894721269608,
0.0041403439827263355,
0.08505036681890488,
-0.06744503974914551,
0.04517138749361038,
-0.025150708854198456,
0.034607987850904465,
0.014556366950273514,
0.018246479332447052,
-0.05287406966090202,
0.029940759763121605,
0.00485363882035017,
-0.02545316331088543,
-0.016717730090022087,
-0.005696631968021393,
-0.015111786313354969,
0.04505728930234909,
-0.01061218325048685,
-0.05222884565591812,
-0.009444518946111202,
0.054922834038734436,
-0.06597185879945755,
-0.049636419862508774,
0.0074845170602202415,
0.0519988089799881,
0.06364616006612778,
-0.009957870468497276,
-0.03527490049600601,
-0.07244366407394409,
-0.07096686959266663,
0.1007077544927597,
0.07630777359008789,
-0.01708955317735672,
0.0074711088091135025,
-0.01448585744947195,
-0.01665312796831131,
0.11042194068431854,
-0.005767020862549543,
0.025507653132081032,
-0.012079089879989624,
0.022458605468273163,
0.012902372516691685,
0.015080403536558151,
0.03963596373796463,
0.005727027542889118,
-0.058754827827215195,
0.01826343685388565,
0.08123521506786346,
0.014592727646231651,
0.03942219167947769,
0.0070114536210894585,
0.04228019341826439,
0.10918458551168442,
0.023246224969625473,
0.05182270705699921,
0.04309498146176338,
-0.007838725112378597,
0.0005805841065011919,
-0.018680360168218613,
0.04174191877245903,
-0.08907680213451385,
0.041884880512952805,
-3.311475182954382e-8,
-0.15215735137462616,
0.018604833632707596,
-0.006430062465369701,
-0.02884012460708618,
0.08274124562740326,
-0.05228923633694649,
0.08914060145616531,
-0.04283847659826279,
-0.043670590966939926,
-0.08753781020641327,
0.06541259586811066,
-0.008872817270457745,
-0.07483092695474625,
-0.08659610897302628,
-0.06645065546035767,
0.06602728366851807,
-0.015217673033475876,
0.06300456076860428,
0.031171565875411034,
-0.020524978637695312,
-0.08617188036441803,
0.04946627840399742,
0.035457905381917953,
-0.01966087333858013,
0.023985547944903374,
-0.07206782698631287,
0.07274012267589569,
0.0349162295460701,
0.12525656819343567,
0.010296068154275417,
-0.10798970609903336,
0.005045457277446985,
0.058258093893527985,
0.03653847426176071,
-0.010311087593436241,
0.042356330901384354,
0.02348184585571289,
0.04597202688455582,
-0.09278817474842072,
-0.07254836708307266,
-0.013993103057146072,
0.03163757920265198,
0.01917060650885105,
-0.09939754009246826,
0.006972075439989567,
0.056393299251794815,
0.03629515692591667,
-0.03254979848861694,
-0.008114052005112171,
-0.05982644483447075,
0.015639174729585648,
0.010720159858465195,
0.08580002933740616,
0.055809225887060165,
0.02588268555700779,
-0.029554974287748337,
0.0240042582154274,
-0.0024618699681013823,
0.04990166798233986,
0.04587865248322487,
-0.01693388819694519,
0.07339023053646088,
-0.07576057314872742,
0.01681499183177948
] | 0 | 0 | 0 | 0 |
|
TULIP: Token-length Upgraded CLIP
|
https://openreview.net/forum?id=r9oqHOdoHf
|
[
"Ivona Najdenkoska",
"Mohammad Mahdi Derakhshani",
"Yuki M Asano",
"Nanne Van Noord",
"Marcel Worring",
"Cees G. M. Snoek"
] |
Poster
|
We address the challenge of representing long captions in vision-language models, such as CLIP. By design these models are limited by fixed, absolute positional encodings, restricting inputs to a maximum of 77 tokens and hindering performance on tasks requiring longer descriptions. Although recent work has attempted to overcome this limit, their proposed approaches struggle to model token relationships over longer distances and simply extend to a fixed new token length. Instead, we propose a generalizable method, named TULIP, able to upgrade the token length to any length for CLIP-like models. We do so by improving the architecture with relative position encodings, followed by a training procedure that (i) distills the original CLIP text encoder into an encoder with relative position encodings and (ii) enhances the model for aligning longer captions with images. By effectively encoding captions longer than the default 77 tokens, our model outperforms baselines on cross-modal tasks such as retrieval and text-to-image generation. The code repository is available at https://github.com/ivonajdenkoska/tulip
|
Vision-Language Models, CLIP, Position Encodings, Long captioning
|
We propose a generalizable method, we call TULIP, able to update the token length to any length for CLIP-like models.
| 11,595 |
2410.10034
|
[
-0.02477409690618515,
-0.0677824318408966,
0.021092060953378677,
-0.04739692062139511,
0.07540757954120636,
0.0045242891646921635,
-0.07244065403938293,
0.01922193169593811,
0.05331627279520035,
-0.06911898404359818,
0.01165462564677,
-0.0649949386715889,
-0.0013252818025648594,
0.048296306282281876,
0.0066465106792747974,
-0.02158227190375328,
0.03237910941243172,
0.14589156210422516,
-0.11457802355289459,
-0.0323554165661335,
0.0467483215034008,
-0.04860810190439224,
0.07139155268669128,
0.030550478026270866,
0.05272257700562477,
0.007233681157231331,
-0.03711015358567238,
-0.04764874279499054,
0.009189555421471596,
-0.06738130748271942,
0.005527321249246597,
-0.017945991829037666,
0.06634917110204697,
0.06510742008686066,
-0.047942887991666794,
0.08670195192098618,
-0.03938138857483864,
-0.0014601405709981918,
0.07088889181613922,
-0.1261252760887146,
-0.02067611925303936,
-0.001191532239317894,
0.07513470947742462,
0.027892975136637688,
0.06442774087190628,
0.015547076240181923,
0.012420283630490303,
-0.06074031442403793,
0.001564624602906406,
0.020868098363280296,
-0.08730080723762512,
-0.0009706891141831875,
0.012578917667269707,
0.03134980425238609,
-0.07192135602235794,
-0.012582928873598576,
0.04081860929727554,
-0.052182238548994064,
0.03694712743163109,
0.010615210048854351,
-0.008018159307539463,
-0.05616895481944084,
0.010065872222185135,
0.03849010542035103,
-0.03914424404501915,
-0.0909542441368103,
0.035120390355587006,
-0.008258051238954067,
-0.05038158595561981,
0.04166357219219208,
-0.04014629125595093,
0.027407566085457802,
-0.05750057473778725,
0.052809037268161774,
-0.025286773219704628,
0.09255775064229965,
0.08026464283466339,
-0.025257453322410583,
0.014550632797181606,
-0.09372878074645996,
0.06393487006425858,
-0.0355372317135334,
0.17709937691688538,
0.07097748667001724,
0.07039900869131088,
-0.04340527951717377,
-0.021317483857274055,
0.007015271577984095,
-0.01676688902080059,
0.026118764653801918,
-0.02528362348675728,
-0.08723843097686768,
0.11833804845809937,
-0.0018839119002223015,
0.004533829633146524,
-0.014126899652183056,
-0.01527467742562294,
-0.041825391352176666,
-0.07296473532915115,
0.08810065686702728,
0.02245941571891308,
0.06842618435621262,
0.006153385154902935,
-0.04925749450922012,
0.08798784762620926,
0.03872682526707649,
0.051947321742773056,
-0.010334073565900326,
0.007506865542382002,
-0.052827879786491394,
0.04904235899448395,
0.029558997601270676,
-0.06063034012913704,
0.01661963015794754,
0.002439055358991027,
0.014607190154492855,
-0.06488676369190216,
-0.05038658529520035,
0.03358180820941925,
0.03388470783829689,
-0.020071201026439667,
0.02655624784529209,
-0.12385600060224533,
0.02131975255906582,
-0.001104468246921897,
-0.0230370182543993,
0.028490712866187096,
3.123377386895099e-33,
0.0026139167603105307,
0.06289354711771011,
0.0792611762881279,
0.0032572129275649786,
0.09453963488340378,
-0.013334902934730053,
-0.052053991705179214,
0.06269966810941696,
-0.10281815379858017,
-0.05707947909832001,
-0.006470524240285158,
-0.06722445785999298,
-0.02478956989943981,
0.08844707906246185,
0.03726673126220703,
-0.03428633138537407,
-0.06356029212474823,
0.0858011245727539,
-0.04873640090227127,
0.047905705869197845,
-0.0560251921415329,
-0.04170387238264084,
-0.023794710636138916,
0.01542821153998375,
0.05316995829343796,
-0.0353710763156414,
0.023744333535432816,
-0.07593976706266403,
-0.053451187908649445,
0.013570599257946014,
-0.09952511638402939,
0.022242670878767967,
0.05366313084959984,
0.01213251892477274,
0.05855873227119446,
-0.04937205836176872,
0.006116196978837252,
-0.017965003848075867,
0.005816630087792873,
0.00477811973541975,
0.013209320604801178,
0.060186564922332764,
-0.05102085694670677,
-0.01126318983733654,
-0.07996528595685959,
-0.04870828241109848,
-0.0002643670595716685,
0.08108839392662048,
-0.07210519164800644,
0.021226683631539345,
0.055765558034181595,
0.060635510832071304,
-0.11549555510282516,
-0.09121832251548767,
0.021836988627910614,
-0.0912996232509613,
0.029344387352466583,
-0.003535828087478876,
0.014549782499670982,
0.0274523738771677,
-0.00286499154753983,
0.033650144934654236,
0.02459307760000229,
0.02953285165131092,
0.025656864047050476,
0.014053026214241982,
0.030363591387867928,
-0.01174156367778778,
0.04943172261118889,
0.010159196332097054,
-0.05423172935843468,
0.016842739656567574,
-0.02344062179327011,
0.01204004604369402,
-0.029064897447824478,
-0.0032588934991508722,
0.03129923716187477,
-0.13165631890296936,
-0.020186929032206535,
0.03574628382921219,
-0.007601680234074593,
-0.01813352108001709,
0.05792970582842827,
-0.05865061655640602,
-0.04409100487828255,
-0.10600059479475021,
0.023969067260622978,
-0.07875224947929382,
-0.01821286976337433,
-0.06765744090080261,
-0.0000531292098457925,
-0.029755663126707077,
-0.1065453588962555,
0.00039758605998940766,
0.011855079792439938,
-2.1063301236720553e-33,
0.051826756447553635,
0.07997491955757141,
-0.02819736860692501,
0.04173727706074715,
-0.06205173209309578,
-0.05403314530849457,
0.0017707821680232882,
0.09772861748933792,
-0.03480222448706627,
-0.03395340219140053,
-0.04021172598004341,
-0.0385163314640522,
0.011027520522475243,
-0.019080858677625656,
0.018543126061558723,
-0.016950972378253937,
0.005566791631281376,
-0.006172063760459423,
-0.03301554173231125,
0.006035120226442814,
0.03342975676059723,
0.056239012628793716,
-0.04411734640598297,
0.047887686640024185,
-0.03260127454996109,
0.10031876713037491,
-0.020507115870714188,
0.01610262505710125,
-0.004172125365585089,
-0.02447686344385147,
-0.031317088752985,
-0.01918867416679859,
-0.035651303827762604,
-0.012643194757401943,
-0.04502337798476219,
0.029082318767905235,
0.02234734408557415,
0.0045236749574542046,
0.02313937060534954,
0.08995059877634048,
0.10258794575929642,
-0.07517888396978378,
0.028780680149793625,
0.01809781976044178,
-0.03493117913603783,
-0.08450447767972946,
-0.01896171271800995,
0.028116580098867416,
-0.009494293481111526,
0.022038519382476807,
-0.052408963441848755,
0.024037430062890053,
-0.10170215368270874,
0.041943494230508804,
-0.02015434205532074,
-0.057683639228343964,
-0.11588399857282639,
-0.03984782099723816,
0.018704792484641075,
-0.07459726184606552,
-0.046097882091999054,
0.005807068198919296,
-0.0007155137718655169,
0.0005245096981525421,
0.06003330647945404,
-0.013323022052645683,
-0.04841795563697815,
0.03879981115460396,
-0.04559415206313133,
-0.027839116752147675,
0.06697879731655121,
-0.05700589343905449,
-0.014560151845216751,
0.08186355233192444,
-0.01999484933912754,
0.007972611114382744,
0.03333733603358269,
-0.017484961077570915,
-0.01172615960240364,
-0.047833606600761414,
-0.04541429504752159,
-0.038418352603912354,
0.05806107446551323,
0.08287161588668823,
0.05753809213638306,
0.0809587612748146,
0.017023606225848198,
0.05653481185436249,
0.03273991122841835,
0.06875531375408173,
-0.0630849078297615,
0.02712339162826538,
0.01257252972573042,
0.11697777360677719,
0.004769191611558199,
-3.986846408565725e-8,
-0.028050214052200317,
0.03825090825557709,
-0.07400386035442352,
0.049018025398254395,
-0.02710304968059063,
-0.0006022669840604067,
-0.03229833394289017,
0.033958714455366135,
-0.025057939812541008,
-0.016671014949679375,
0.03431794419884682,
0.06275951117277145,
-0.017784830182790756,
-0.02172391675412655,
-0.04147043451666832,
0.07334848493337631,
0.011966004967689514,
0.01881747879087925,
0.011548389680683613,
0.07803095877170563,
0.004835658241063356,
0.046248894184827805,
0.019525714218616486,
-0.04061395302414894,
-0.06345877051353455,
-0.019919615238904953,
-0.11109674721956253,
0.12668457627296448,
0.0746038481593132,
-0.06843672692775726,
0.03610331937670708,
0.0528792068362236,
0.0023360527120530605,
-0.022511741146445274,
0.025280119851231575,
-0.024098897352814674,
-0.009064757265150547,
-0.06956121325492859,
-0.00500209117308259,
0.013303522020578384,
0.06509905308485031,
-0.006220776122063398,
-0.036722149699926376,
0.03499777242541313,
0.05934205651283264,
0.03463586047291756,
0.10558629035949707,
-0.08619635552167892,
0.021265489980578423,
-0.039382316172122955,
-0.039014920592308044,
-0.05331768840551376,
-0.004067028407007456,
0.09600409865379333,
0.034053511917591095,
0.0013684581499546766,
0.007347422186285257,
0.0164188239723444,
0.08103771507740021,
0.09220513701438904,
-0.009703809395432472,
0.029167909175157547,
-0.03539486601948738,
-0.021945765241980553
] |
https://github.com/ivonajdenkoska/tulip
| 25 | 0 | 0 | 0 |
DELIFT: Data Efficient Language model Instruction Fine-Tuning
|
https://openreview.net/forum?id=Fty0wTcemV
|
[
"Ishika Agarwal",
"Krishnateja Killamsetty",
"Lucian Popa",
"Marina Danilevsky"
] |
Poster
|
Fine-tuning large language models (LLMs) is crucial for task specialization but often becomes resource-intensive due to redundant or uninformative data. Existing data selection methods typically rely either on computationally expensive gradient-based metrics or static embeddings that fail to adapt dynamically to the model’s evolving state, thus limiting their practical effectiveness. To address this,
we propose DELIFT (Data Efficient Language model Instruction Fine-Tuning), leveraging a novel, computationally efficient utility metric inspired by In-Context Learning (ICL). Our ICL-based metric measures the informational value of each data sample by quantifying its effectiveness as an in-context example in improving model predictions for other samples, reflecting its actual contribution relative to the model’s current state. Integrated with tailored submodular optimization methods, DELIFT systematically selects diverse, informative subsets optimized specifically for each fine-tuning stage: instruction tuning, task-specific adaptation, and continual fine-tuning. Experimental results across multiple datasets and model scales show DELIFT reduces fine-tuning data requirements by up to 70% without compromising performance, consistently outperforming existing methods by up to 26% in effectiveness and efficiency.
|
Data Efficient Instruction Fine-Tuning; Data Subset Selection; Submodular Functions
|
We introduce DELIFT, a novel algorithm that efficiently and systematically optimizes data selection across the three key stages of fine-tuning by using a pairwise utility metric to capture informativeness of data points.
| 11,590 |
2411.04425
|
[
-0.0210589487105608,
-0.09800039231777191,
0.030002975836396217,
0.0542043000459671,
0.10184689611196518,
-0.0439639613032341,
0.04004095867276192,
-0.004193740431219339,
0.022199934348464012,
-0.05760161578655243,
-0.07317229360342026,
-0.09292630851268768,
0.02345959097146988,
0.0199115090072155,
0.0489179864525795,
-0.01233653910458088,
0.11106182634830475,
0.040935467928647995,
-0.1380533128976822,
-0.08642109483480453,
0.08315277844667435,
0.045247700065374374,
0.07713726907968521,
0.028261573985219002,
0.019298259168863297,
0.016422707587480545,
0.021452927961945534,
-0.012799873948097229,
0.012939602136611938,
0.0032747439108788967,
0.026856904849410057,
0.05033297836780548,
0.003089302219450474,
0.00039051362546160817,
-0.036056675016880035,
0.11144920438528061,
-0.07890794426202774,
0.016247399151325226,
0.02409842610359192,
-0.014716346748173237,
-0.028074340894818306,
0.018257243558764458,
0.03159739077091217,
0.038294658064842224,
0.04259603098034859,
-0.03372219204902649,
-0.03446196764707565,
-0.06393736600875854,
-0.050545673817396164,
0.0746603012084961,
-0.07229672372341156,
0.01033489964902401,
-0.012118508107960224,
0.043001845479011536,
0.014296160079538822,
0.02238878794014454,
0.034495532512664795,
0.030314791947603226,
-0.04158218577504158,
-0.026778127998113632,
-0.07868692278862,
-0.16032254695892334,
-0.03454338014125824,
-0.03529627248644829,
-0.03545422479510307,
-0.007004374172538519,
-0.010750007815659046,
0.009456686675548553,
-0.023213665932416916,
0.013189342804253101,
-0.07867950946092606,
0.0713067427277565,
-0.006048873066902161,
0.08023630827665329,
-0.00917193666100502,
0.05353010445833206,
0.0877278745174408,
-0.05475535988807678,
0.06310755014419556,
-0.07860895991325378,
0.028063958510756493,
0.0578438937664032,
0.0023504106793552637,
-0.015738358721137047,
0.17942211031913757,
-0.02415439300239086,
0.07450070232152939,
-0.03573845699429512,
0.0653156116604805,
0.026453152298927307,
-0.00540293799713254,
-0.017961643636226654,
0.010377263650298119,
-0.0030826248694211245,
0.025632550939917564,
0.03514315187931061,
0.03378851339221001,
-0.08375654369592667,
-0.0738697424530983,
0.05525655299425125,
0.027354490011930466,
0.09544704109430313,
-0.019282370805740356,
0.016527632251381874,
-0.07608257234096527,
-0.06176161766052246,
0.004748140461742878,
0.11062033474445343,
-0.051182955503463745,
-0.07194636017084122,
0.0231277197599411,
0.048211291432380676,
0.030872348695993423,
0.038705144077539444,
0.028304755687713623,
0.001964292721822858,
0.020767616108059883,
-0.06584447622299194,
0.05846508592367172,
0.17114517092704773,
-0.0733008086681366,
0.0013183843111619353,
-0.009048818610608578,
-0.007381749339401722,
-0.08309244364500046,
-0.039935849606990814,
-0.08379426598548889,
3.0166294591999807e-33,
0.008803441189229488,
0.015117490664124489,
0.0069560399278998375,
0.05376569554209709,
0.03044203668832779,
-0.0351053848862648,
-0.011137079447507858,
0.06469151377677917,
-0.06104014441370964,
-0.0007662083371542394,
0.024624362587928772,
0.0036524103488773108,
-0.0462789386510849,
0.09126213937997818,
0.06716887652873993,
0.0022907608654350042,
-0.08438141644001007,
0.07216943055391312,
-0.07047852873802185,
0.011397456750273705,
0.08812739700078964,
0.0073431795462965965,
-0.00730046397075057,
-0.0754128023982048,
0.059225086122751236,
0.015017542988061905,
0.09121561050415039,
-0.020276052877306938,
-0.05898991599678993,
0.01911887153983116,
-0.08970099687576294,
-0.031204765662550926,
-0.02003086730837822,
0.05779389292001724,
0.019774459302425385,
-0.033597227185964584,
-0.02324465662240982,
-0.0017441717209294438,
0.014469570480287075,
0.003060542978346348,
0.05029600113630295,
0.002204928547143936,
0.0709490031003952,
-0.01248426828533411,
-0.11412401497364044,
-0.11212767660617828,
-0.043672434985637665,
-0.0315328910946846,
0.009417442604899406,
-0.05222680792212486,
0.086981400847435,
0.015574648045003414,
-0.0038758572190999985,
-0.04101666435599327,
-0.017339494079351425,
-0.007697365712374449,
0.034347597509622574,
0.030558917671442032,
0.04721703380346298,
0.06777297705411911,
-0.022756105288863182,
0.010154949501156807,
-0.030633820220828056,
0.0994013100862503,
0.0514051578938961,
0.046323928982019424,
-0.025606220588088036,
0.09132359176874161,
0.0328119695186615,
-0.07627642899751663,
-0.005550719331949949,
-0.0016467849491164088,
0.02426059916615486,
-0.0152019914239645,
0.038804180920124054,
-0.006324430927634239,
0.04364898428320885,
-0.05687601864337921,
-0.004555266350507736,
0.0004897401668131351,
0.08660025894641876,
0.04385913908481598,
0.02347533032298088,
-0.03465067967772484,
-0.04612446576356888,
-0.0785926952958107,
0.03499976918101311,
-0.069706030189991,
-0.04254194721579552,
-0.05834420025348663,
-0.04145896062254906,
-0.006562335416674614,
-0.029290731996297836,
-0.03445308655500412,
-0.04444464296102524,
-2.9929421460049577e-33,
0.0294779185205698,
-0.04886353388428688,
0.024886179715394974,
0.026500031352043152,
0.00921747274696827,
-0.034191373735666275,
0.004219760186970234,
0.023962557315826416,
-0.009015906602144241,
-0.09916536509990692,
-0.020149150863289833,
-0.05875420570373535,
0.030524441972374916,
0.017308836802840233,
0.017723295837640762,
0.03603162616491318,
-0.04819290712475777,
0.018493598327040672,
-0.0014646955532953143,
0.04463450238108635,
0.02326611988246441,
0.04123586788773537,
-0.04824534431099892,
0.034959908574819565,
-0.0329943411052227,
0.0017716530710458755,
-0.08791171759366989,
0.029618872329592705,
-0.0006636318867094815,
0.001673673978075385,
-0.11012540757656097,
0.02680923603475094,
0.030690329149365425,
-0.029678815975785255,
-0.027656879276037216,
0.03543051332235336,
0.061250656843185425,
-0.04484729468822479,
-0.07954829186201096,
0.1213407963514328,
0.060229744762182236,
0.027019768953323364,
-0.08347591757774353,
0.05427239090204239,
0.0166830662637949,
0.0047779749147593975,
-0.0747220516204834,
-0.06067512556910515,
0.03684097155928612,
0.02382684126496315,
-0.04276131093502045,
0.006812899839133024,
-0.04448303207755089,
-0.0010334085673093796,
-0.03991175442934036,
-0.10820620507001877,
-0.01828489452600479,
-0.11463694274425507,
0.0544583760201931,
-0.0319347009062767,
-0.04683008790016174,
0.031998101621866226,
0.035707831382751465,
-0.068813756108284,
0.048640768975019455,
-0.05815701559185982,
-0.01327483355998993,
-0.0810365304350853,
0.01901623047888279,
-0.01010999083518982,
-0.015705419704318047,
-0.0283511970192194,
0.010950537398457527,
-0.05847298353910446,
-0.03603098541498184,
-0.019104154780507088,
0.034261103719472885,
-0.06255228072404861,
-0.049270208925008774,
0.028959032148122787,
0.027694297954440117,
-0.0028131152503192425,
-0.022426610812544823,
-0.001155026606284082,
-0.029895687475800514,
0.08730417490005493,
0.034727007150650024,
0.13156193494796753,
0.003080762689933181,
-0.011286016553640366,
-0.0633639320731163,
-0.02426447719335556,
0.011002992279827595,
0.056974541395902634,
-0.057228561490774155,
-4.464857283892343e-8,
-0.011961501091718674,
0.05475842207670212,
-0.03899986296892166,
0.09237278252840042,
-0.005648551508784294,
-0.0006812528008595109,
-0.06799117475748062,
0.09938312321901321,
0.07223046571016312,
-0.033119622617959976,
0.028421973809599876,
0.0023623902816325426,
-0.061094824224710464,
-0.02354753576219082,
0.005715487990528345,
0.047440722584724426,
0.04135458916425705,
0.04841839522123337,
0.03296831622719765,
-0.0019556547049432993,
0.03880570828914642,
0.059113647788763046,
0.043683674186468124,
0.007347972132265568,
0.06226037070155144,
-0.0747462585568428,
-0.03984019160270691,
0.07252484560012817,
0.09478095173835754,
-0.02068677358329296,
-0.047793880105018616,
-0.026135528460144997,
0.016826994717121124,
-0.05940736457705498,
-0.04014706611633301,
0.06882836669683456,
-0.00113632099237293,
-0.03983139619231224,
-0.007022262550890446,
0.006757366470992565,
0.0483766533434391,
0.10443248599767685,
-0.0682247057557106,
0.05347965657711029,
0.030282029882073402,
-0.02070355787873268,
-0.05019392818212509,
-0.04888126626610756,
0.017344819381833076,
-0.03591496869921684,
-0.0007968881400302052,
0.007947838865220547,
-0.03558214381337166,
0.0354117788374424,
0.02820943295955658,
0.049658432602882385,
-0.06488870084285736,
-0.02188817597925663,
0.025218162685632706,
0.03563182055950165,
0.009752009995281696,
-0.004226548131555319,
-0.032963234931230545,
-0.01339504774659872
] |
https://github.com/agarwalishika/delift
| 8 | 0 | 0 | 0 |
U-shaped and Inverted-U Scaling behind Emergent Abilities of Large Language Models
|
https://openreview.net/forum?id=jjfve2gIXe
|
[
"Tung-Yu Wu",
"Melody Lo"
] |
Poster
|
Large language models (LLMs) have been shown to exhibit *emergent abilities* in some downstream tasks, where model performance stagnates at first and then improves sharply and unpredictably with scale beyond a threshold. In this work, we investigate the phenomenon by grouping questions based on difficulty level and provide a possible explanation for emergent abilities. Specifically, we observe U-shaped scaling for hard questions and inverted-U scaling followed by steady improvement for easy questions. The two scaling patterns initially offset each other, causing stagnant overall performance. The performance starts to soar when the scaling pattern of easy questions reverts from inverse to standard scaling, leading to emergent abilities. Based on this finding, we propose a simple yet effective pipeline, called *Slice-and-Sandwich*, to predict the emergence threshold and model performance beyond the threshold. Our code is publicly available at https://github.com/tony10101105/ExpEmergence.
|
large language models, emergent abilities, scaling laws
|
The emergent abilities of LLMs can be decomposed into complementary non-trivial scaling trends in easy and hard samples, which sheds light on deeper understanding and prediction of emergent abilities.
| 11,589 | null |
[
-0.006306299939751625,
-0.07774019986391068,
0.01768619380891323,
0.05584372952580452,
0.006005951669067144,
-0.02011185884475708,
-0.051690615713596344,
0.039507560431957245,
0.026044633239507675,
-0.03499101102352142,
-0.05155528709292412,
-0.07816290855407715,
-0.010727069340646267,
0.03000708483159542,
-0.00009846521425060928,
-0.00516734691336751,
0.01971529982984066,
-0.006918573286384344,
-0.08635209500789642,
-0.08317597210407257,
0.00637753214687109,
0.0212194062769413,
0.007046719081699848,
-0.02844838798046112,
0.08180195838212967,
-0.024185672402381897,
-0.05671204999089241,
-0.004857032559812069,
0.10439978539943695,
-0.004819301888346672,
0.045571766793727875,
0.014077366329729557,
0.045207250863313675,
0.04053270444273949,
-0.028931261971592903,
0.07320315390825272,
0.011721678078174591,
-0.008063637651503086,
0.0017233811086043715,
-0.07071886956691742,
-0.01922658644616604,
-0.02760440483689308,
0.03844781219959259,
-0.007333990652114153,
0.08314482867717743,
-0.08406339585781097,
-0.0493796207010746,
0.07743728160858154,
-0.05741984769701958,
-0.035801734775304794,
-0.053039152175188065,
-0.01700541563332081,
0.008063958957791328,
0.07592625170946121,
-0.027228238061070442,
0.008386674337089062,
0.034561675041913986,
0.058268044143915176,
-0.01316186785697937,
-0.053050797432661057,
-0.12495949119329453,
-0.12453830242156982,
-0.004001873545348644,
0.030413426458835602,
0.046144355088472366,
0.011718963272869587,
0.050690390169620514,
-0.015025203116238117,
-0.03510001674294472,
0.09039516001939774,
-0.01902516931295395,
0.05661989748477936,
-0.039337679743766785,
0.08778099715709686,
0.08065128326416016,
0.014216362498700619,
-0.03448377549648285,
-0.0352749302983284,
0.030718835070729256,
0.029819468036293983,
-0.004103565122932196,
-0.020979752764105797,
-0.012876288965344429,
0.03967514634132385,
0.07470547407865524,
0.03914256766438484,
0.04073631763458252,
0.005484112538397312,
0.001184268039651215,
-0.01225521881133318,
-0.04018257558345795,
-0.02042854018509388,
0.025401802733540535,
0.04837699607014656,
0.025302931666374207,
-0.014068777672946453,
0.028606189414858818,
-0.11355582624673843,
-0.033682696521282196,
0.03339497745037079,
0.0496533066034317,
0.08408498764038086,
0.05642000213265419,
-0.002423313446342945,
0.0022341134026646614,
0.00027627969393506646,
0.013232806697487831,
0.015170701779425144,
-0.0036160768941044807,
-0.06145518273115158,
-0.03380998969078064,
0.015694158151745796,
0.041482891887426376,
0.027926955372095108,
-0.004750074818730354,
-0.07920002937316895,
0.06650258600711823,
-0.048757873475551605,
-0.026143651455640793,
0.05578525736927986,
0.03754429891705513,
0.07865096628665924,
-0.032137591391801834,
-0.022216999903321266,
0.008193359710276127,
0.032386839389801025,
-0.05944773927330971,
7.347933912576019e-34,
0.07573450356721878,
0.02431584522128105,
0.02624494768679142,
0.08290427178144455,
0.04097984731197357,
0.033452123403549194,
-0.005073754116892815,
0.00721233244985342,
-0.057012882083654404,
-0.03806040808558464,
-0.034110162407159805,
0.06307418644428253,
-0.005271803122013807,
0.07392489165067673,
0.07555542141199112,
-0.022752979770302773,
0.016209011897444725,
0.029842188581824303,
-0.07421724498271942,
-0.07008511573076248,
0.011910272762179375,
0.038508884608745575,
0.05826765298843384,
-0.08727169036865234,
0.027662301436066628,
0.012589486315846443,
0.02815777063369751,
-0.08832008391618729,
-0.10455512255430222,
0.010536535643041134,
-0.09566347301006317,
-0.11391881108283997,
-0.1128700003027916,
0.04611470177769661,
-0.017540572211146355,
-0.024125682190060616,
0.08250351250171661,
-0.043995440006256104,
0.08903956413269043,
-0.04359576106071472,
-0.017881441861391068,
0.06002860888838768,
0.012295330874621868,
-0.009936935268342495,
-0.0525909885764122,
-0.013764453120529652,
0.015242911875247955,
-0.012554297223687172,
-0.03893719241023064,
-0.06665407121181488,
0.06610524654388428,
0.01073524821549654,
0.005138344597071409,
-0.052513238042593,
0.04673153907060623,
0.032290469855070114,
-0.02476789802312851,
0.07645314186811447,
-0.010902069509029388,
0.048706285655498505,
-0.008878529071807861,
0.009337636642158031,
-0.039671849459409714,
0.07715717703104019,
0.03579559177160263,
-0.052276890724897385,
-0.002673541195690632,
0.056989189237356186,
0.07061978429555893,
0.0004969020956195891,
0.03302130103111267,
-0.018807504326105118,
-0.03639132156968117,
-0.037617363035678864,
0.038028642535209656,
0.029627112671732903,
0.10082124918699265,
-0.10136502981185913,
0.020893175154924393,
-0.003802479710429907,
-0.05589791387319565,
-0.027117831632494926,
0.07458595186471939,
-0.1088661178946495,
-0.04291372373700142,
-0.05471087992191315,
0.07293292880058289,
-0.05539281666278839,
-0.04232543706893921,
-0.005167264025658369,
-0.06591171026229858,
-0.02680441364645958,
0.07228246331214905,
-0.017490772530436516,
0.031753670424222946,
-2.5011878596628774e-33,
-0.047124579548835754,
-0.11141173541545868,
-0.06402651220560074,
0.06486576050519943,
-0.0199267715215683,
-0.07824870944023132,
-0.008420872502028942,
0.04973027482628822,
-0.08327636122703552,
-0.052072037011384964,
-0.08156149834394455,
0.015698684379458427,
0.049056511372327805,
-0.0012525261845439672,
0.08438004553318024,
-0.06524065136909485,
0.03471427038311958,
0.05079881474375725,
0.12248744070529938,
0.07990172505378723,
0.04280082881450653,
0.07940453290939331,
-0.06918048858642578,
0.050348926335573196,
0.0006970438407734036,
0.008248798549175262,
-0.04329678416252136,
-0.03931674361228943,
-0.02355164848268032,
-0.03495482727885246,
-0.07081621140241623,
0.0033277946058660746,
0.00406492967158556,
0.037198036909103394,
0.01044683437794447,
0.08983350545167923,
-0.05846158787608147,
-0.05092440918087959,
-0.021928494796156883,
0.0580497682094574,
-0.013681327924132347,
-0.004561326932162046,
-0.0625237300992012,
-0.02454107254743576,
-0.039467036724090576,
-0.034304726868867874,
-0.03587542101740837,
0.0032421695068478584,
-0.06734129041433334,
0.008925758302211761,
-0.006221671123057604,
0.06501185894012451,
-0.03317287191748619,
0.03246510028839111,
-0.07669726759195328,
0.02110561728477478,
-0.005023002624511719,
-0.07402072846889496,
0.0019431117689236999,
0.04259851947426796,
-0.05924782156944275,
0.019348567351698875,
0.02249975875020027,
-0.08715185523033142,
0.029558828100562096,
-0.049043137580156326,
-0.024199409410357475,
0.005219858139753342,
-0.036789048463106155,
-0.06065741553902626,
0.028054147958755493,
-0.018137428909540176,
-0.008332646451890469,
-0.02027837187051773,
-0.038350000977516174,
0.00976584292948246,
-0.03968857601284981,
-0.05540962517261505,
0.011785352602601051,
-0.06424643099308014,
-0.021414348855614662,
0.04066218063235283,
-0.028963729739189148,
0.06294184923171997,
-0.009970343671739101,
-0.021011412143707275,
0.05036630854010582,
0.1324102133512497,
0.03609176725149155,
0.08634981513023376,
0.024784373119473457,
-0.03299369663000107,
-0.0359746478497982,
0.027867279946804047,
-0.0850929319858551,
-4.12513827541261e-8,
-0.05695159360766411,
0.05220138281583786,
-0.018085891380906105,
0.12375736236572266,
0.043992962688207626,
-0.01161891222000122,
-0.08263899385929108,
0.07744171470403671,
-0.015347146429121494,
-0.022296447306871414,
0.012323945760726929,
0.0002638032892718911,
-0.010934379883110523,
0.011811981908977032,
0.07080266624689102,
0.09339963644742966,
-0.007342671975493431,
0.07470203936100006,
-0.037801600992679596,
-0.05183664709329605,
0.023222608491778374,
0.07983395457267761,
-0.04825936257839203,
0.003240518970414996,
-0.055432356894016266,
-0.03470294922590256,
-0.06968539953231812,
0.058286700397729874,
0.03054145909845829,
-0.020583877339959145,
-0.013875821605324745,
-0.003223577281460166,
-0.06085078418254852,
0.05353566259145737,
0.06630749255418777,
0.1043296605348587,
-0.1300538331270218,
0.026374123990535736,
-0.0323270820081234,
-0.02451418898999691,
0.0843374952673912,
0.020241810008883476,
0.033753279596567154,
0.047423120588064194,
-0.04290705546736717,
-0.01936909928917885,
-0.11179838329553604,
-0.03791026771068573,
0.04444221779704094,
-0.11642695963382721,
0.01780284196138382,
-0.05520882457494736,
-0.03797281160950661,
0.016505399718880653,
0.08805876225233078,
0.08631421625614166,
-0.010828239843249321,
-0.01382747944444418,
-0.014469361864030361,
0.022148167714476585,
0.09990816563367844,
0.0497501902282238,
-0.013388230465352535,
0.033521197736263275
] | 0 | 0 | 0 | 0 |
|
Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models
|
https://openreview.net/forum?id=tU074jg2vS
|
[
"Pit Neitemeier",
"Björn Deiseroth",
"Constantin Eichenberg",
"Lukas Balles"
] |
Poster
|
Tokenization is a fundamental step in natural language processing, breaking text into units that computational models can process. While learned subword tokenizers have become the de-facto standard, they present challenges such as large vocabularies, limited adaptability to new domains or languages, and sensitivity to spelling errors and variations. To overcome these limitations, we investigate a hierarchical architecture for autoregressive language modelling that combines character-level and word-level processing. It employs a lightweight character-level encoder to convert character sequences into word embeddings, which are then processed by a word-level backbone model and decoded back into characters via a compact character-level decoder. This method retains the sequence compression benefits of word-level tokenization without relying on a rigid, predefined vocabulary. We demonstrate, at scales up to 7 billion parameters, that hierarchical transformers match the downstream task performance of subword-tokenizer-based models while exhibiting significantly greater robustness to input perturbations. Additionally, during continued pretraining on an out-of-domain language, our model trains almost twice as fast, achieves superior performance on the target language, and retains more of its previously learned knowledge. Hierarchical transformers pave the way for NLP systems that are more robust, flexible, and generalizable across languages and domains.
|
transformer, autoregressive, generative, language modelling, tokenizer-free, byte-level, hierarchical
| null | 11,586 | null |
[
-0.0724848210811615,
-0.130788192152977,
0.08092959970235825,
0.039538681507110596,
-0.027964863926172256,
-0.024736423045396805,
-0.07797148078680038,
0.04681447893381119,
0.030695617198944092,
-0.051209308207035065,
-0.019404621794819832,
-0.049912869930267334,
0.05473042279481888,
0.049772027879953384,
-0.034717850387096405,
0.017649061977863312,
0.07429581135511398,
0.07334127277135849,
-0.12558263540267944,
-0.0839705616235733,
0.0644744336605072,
0.01723402738571167,
0.0019856812432408333,
0.0035109524615108967,
0.09206940233707428,
0.05230285972356796,
-0.07100988179445267,
-0.03133873641490936,
0.02362559735774994,
-0.027453625574707985,
0.026473473757505417,
-0.01872764341533184,
0.011420848779380322,
0.03144867718219757,
0.0036418980453163385,
0.0656256377696991,
-0.020349862053990364,
-0.038736727088689804,
0.028686579316854477,
-0.06281383335590363,
0.0005475819343701005,
-0.051901303231716156,
0.0005651703686453402,
0.06511455029249191,
0.07118472456932068,
-0.032300785183906555,
0.018919529393315315,
-0.03737768530845642,
-0.0055485675111413,
0.03042319230735302,
-0.022166475653648376,
-0.02090783230960369,
0.007988180965185165,
0.08081871271133423,
-0.06329306960105896,
0.022902820259332657,
-0.0015895419055595994,
0.02666771598160267,
-0.02772529609501362,
-0.034945376217365265,
-0.08843310922384262,
-0.06952159851789474,
-0.024307169020175934,
-0.03468957915902138,
-0.095368891954422,
-0.07221435755491257,
0.027920180931687355,
0.07688380032777786,
0.025742847472429276,
0.08771664649248123,
0.00043601199286058545,
0.058165453374385834,
-0.09012928605079651,
0.04211229830980301,
0.026063254103064537,
0.030481476336717606,
0.04841070994734764,
-0.07323251664638519,
0.11568067222833633,
-0.08777069300413132,
0.047308798879384995,
0.01988428831100464,
0.041752446442842484,
0.006396039389073849,
0.11435457319021225,
-0.01526100654155016,
0.05073503777384758,
0.0207667239010334,
-0.023644262924790382,
0.013508501462638378,
-0.04272053763270378,
-0.13979600369930267,
0.12621396780014038,
-0.026609860360622406,
-0.04341829940676689,
0.0076898010447621346,
0.00802406296133995,
-0.03554050624370575,
-0.026107335463166237,
0.06487637013196945,
-0.011223165318369865,
0.08174533396959305,
0.0697924941778183,
-0.019825398921966553,
-0.0822245180606842,
-0.03408258035778999,
0.002681863261386752,
0.07667011767625809,
-0.0006271883612498641,
-0.019474321976304054,
-0.027901548892259598,
0.05524159595370293,
0.012451103888452053,
0.003109993413090706,
0.05274275690317154,
0.009977408684790134,
-0.06465195119380951,
-0.0766061395406723,
0.04003525897860527,
0.06937839090824127,
-0.06278891116380692,
0.042817745357751846,
-0.0336453802883625,
0.06268080323934555,
-0.0037060610484331846,
-0.060973480343818665,
-0.011413073167204857,
-1.4341800661489712e-33,
0.015081786550581455,
0.08698511123657227,
-0.02989715337753296,
0.0442291796207428,
0.018112920224666595,
-0.045073483139276505,
-0.021685849875211716,
0.08413439244031906,
-0.04903199151158333,
0.027090787887573242,
-0.0012173044960945845,
0.02566937729716301,
-0.016466382890939713,
0.06531467288732529,
0.05644692853093147,
0.005051746964454651,
-0.05424027517437935,
0.004500285722315311,
-0.034084610641002655,
-0.027952050790190697,
0.013352464884519577,
0.03596341982483864,
-0.027392977848649025,
-0.0411578044295311,
-0.0581863671541214,
-0.009905713610351086,
0.07383813709020615,
-0.06900206208229065,
-0.07087355852127075,
0.030620871111750603,
-0.11333240568637848,
-0.056705277413129807,
0.04951055720448494,
0.022173697128891945,
-0.004020150750875473,
-0.10215587913990021,
0.07390394061803818,
-0.06418076902627945,
-0.01795377768576145,
-0.06363782286643982,
-0.030073855072259903,
0.029778696596622467,
0.029697434976696968,
-0.02002212032675743,
-0.03563554584980011,
-0.053637221455574036,
0.02228838950395584,
0.018668893724679947,
-0.0953516811132431,
-0.024794772267341614,
0.10312677174806595,
0.04716772586107254,
-0.025324640795588493,
0.02706889435648918,
0.03594021126627922,
0.025390122085809708,
0.0212782621383667,
-0.006448655854910612,
0.00734718469902873,
0.03172899782657623,
-0.0016245734877884388,
0.029519369825720787,
0.046178288757801056,
0.08032143115997314,
0.08670280873775482,
-0.04440728947520256,
0.01959289237856865,
0.07017450779676437,
-0.02560858055949211,
0.033479321748018265,
-0.045826636254787445,
-0.004872564692050219,
-0.03865167498588562,
-0.02580735646188259,
0.04392951726913452,
0.014477778226137161,
0.057939477264881134,
-0.0870736762881279,
-0.061709776520729065,
0.05052967742085457,
-0.05332368612289429,
-0.05083716660737991,
0.030416933819651604,
-0.05675286427140236,
-0.025797775015234947,
-0.09062141925096512,
0.09005767852067947,
-0.0998137965798378,
-0.0027244682423770428,
-0.01816748082637787,
-0.05171018838882446,
-0.07020048052072525,
0.009457756765186787,
-0.01329889427870512,
0.048738040030002594,
1.2078699448023296e-33,
0.03278018534183502,
-0.0621553398668766,
-0.05781986191868782,
0.027647173032164574,
-0.06113921478390694,
-0.10646952688694,
0.023268353193998337,
0.06655825674533844,
-0.06709190458059311,
-0.05731801688671112,
-0.0931423157453537,
0.0012041450245305896,
0.07200087606906891,
-0.002067269990220666,
0.09367761760950089,
-0.04934627190232277,
-0.007282653823494911,
0.04722480848431587,
0.06477318704128265,
0.10172290354967117,
-0.027700411155819893,
0.059532199054956436,
-0.06676661223173141,
0.05964398384094238,
0.027962855994701385,
0.053419262170791626,
-0.06259836256504059,
0.07905861735343933,
0.01496104709804058,
0.023677172139286995,
-0.04491390287876129,
0.06116859242320061,
0.02976987324655056,
-0.029182663187384605,
-0.11334463953971863,
0.047907084226608276,
-0.008770911023020744,
-0.008890623226761818,
0.0035213169176131487,
-0.03227689117193222,
0.08394695818424225,
-0.022998636588454247,
-0.03599099442362785,
0.011675061658024788,
-0.018901394680142403,
-0.030821355059742928,
-0.09651873260736465,
-0.0004176155780442059,
0.03530826419591904,
0.0018183542415499687,
-0.009683188982307911,
0.04038699343800545,
-0.057707447558641434,
0.06654322147369385,
0.012454913929104805,
-0.07262880355119705,
-0.056213922798633575,
-0.07863661646842957,
0.013263143599033356,
-0.01961454190313816,
0.01054514478892088,
-0.03509952872991562,
0.056852854788303375,
-0.0710805132985115,
0.006193977780640125,
-0.07075797766447067,
-0.031079396605491638,
0.017157943919301033,
0.030252140015363693,
-0.0009059797739610076,
0.010798174887895584,
-0.03299221396446228,
0.04191267490386963,
0.054867934435606,
-0.011067816987633705,
-0.060558021068573,
0.01436275988817215,
-0.05101115256547928,
-0.03058602474629879,
-0.054091643542051315,
-0.03020702861249447,
0.02052978239953518,
0.004583976697176695,
0.04942898079752922,
0.04173307865858078,
0.029712235555052757,
-0.012124653905630112,
0.08701382577419281,
0.03287067264318466,
0.04119650274515152,
0.0396772138774395,
0.028247511014342308,
0.018852107226848602,
0.08770158886909485,
-0.030180135741829872,
-3.5201704662313205e-8,
-0.02379627339541912,
0.05394146963953972,
0.0058121406473219395,
0.020021842792630196,
0.032421715557575226,
-0.07037237286567688,
-0.07424445450305939,
0.022974461317062378,
0.012315930798649788,
0.006986010819673538,
0.03425607830286026,
-0.03076501563191414,
-0.04650600254535675,
-0.05499542877078056,
-0.028216183185577393,
0.12580060958862305,
0.0644868016242981,
0.027034392580389977,
-0.009023791179060936,
-0.01961764320731163,
-0.010191977955400944,
0.1317344754934311,
-0.014221122488379478,
-0.020342769101262093,
-0.010987904854118824,
-0.025181956589221954,
-0.031885452568531036,
0.05721075460314751,
0.0889536440372467,
-0.029298631474375725,
0.002744191326200962,
0.04967990145087242,
-0.02512810193002224,
-0.08925603330135345,
-0.0290688406676054,
0.11558530479669571,
0.08908526599407196,
-0.01212757732719183,
-0.03582766279578209,
-0.05312883481383324,
0.12172519415616989,
0.04109477251768112,
-0.0458049401640892,
0.015356718562543392,
0.022155674174427986,
-0.02214772067964077,
-0.031897127628326416,
-0.0315144807100296,
0.0898161381483078,
-0.02211172878742218,
0.03843306750059128,
0.019316527992486954,
0.002360400976613164,
0.0010535281617194414,
0.019389808177947998,
0.0497591532766819,
-0.01683720201253891,
0.0005932118510827422,
0.030317209661006927,
0.025364669039845467,
0.029866842553019524,
-0.013736180029809475,
0.009419629350304604,
0.023564498871564865
] | 0 | 0 | 0 | 0 |
|
Locally Connected Echo State Networks for Time Series Forecasting
|
https://openreview.net/forum?id=KeRwLLwZaw
|
[
"Filip Matzner",
"František Mráz"
] |
Poster
|
Echo State Networks (ESNs) are a class of recurrent neural networks in which only a small readout regression layer is trained, while the weights of the recurrent network, termed the reservoir, are randomly assigned and remain fixed. Our work introduces the Locally Connected ESN (LCESN), a novel ESN variant with a locally connected reservoir, forced memory, and a weight adaptation strategy. LCESN significantly reduces the asymptotic time and space complexities compared to the conventional ESN, enabling substantially larger networks. LCESN also improves the memory properties of ESNs without affecting network stability. We evaluate LCESN's performance on the NARMA10 benchmark task and compare it to state-of-the-art models on nine real-world datasets. Despite the simplicity of our model and its one-shot training approach, LCESN achieves competitive results, even surpassing several state-of-the-art models. LCESN introduces a fresh approach to real-world time series forecasting and demonstrates that large, well-tuned random recurrent networks can rival complex gradient-trained feedforward models. We provide our GPU-based implementation of LCESN as an open-source library.
|
Time Series Analysis, Time Series Forecasting, TSF, Recurrent Neural Networks, RNN, Regression, Echo State Networks, ESN
|
Improved locally connected ESN method comparable with state-of-the-art on real-world time series datasets.
| 11,567 | null |
[
-0.12494050711393356,
-0.11282695829868317,
0.012472173199057579,
0.07541002333164215,
0.05565870553255081,
0.0790562629699707,
-0.07353707402944565,
-0.04744739830493927,
-0.04732802137732506,
-0.12465229630470276,
-0.12291545420885086,
0.020626654848456383,
-0.04443629831075668,
0.021240198984742165,
-0.147378072142601,
-0.06846155971288681,
0.057081256061792374,
-0.0012362576089799404,
-0.036762844771146774,
-0.054421450942754745,
0.021332064643502235,
0.06414790451526642,
0.03911570832133293,
0.04521205276250839,
0.10808847844600677,
-0.016596665605902672,
-0.018834179267287254,
-0.05121716111898422,
-0.04173462092876434,
-0.005047198385000229,
0.035275816917419434,
-0.04060690104961395,
-0.049643125385046005,
-0.03518565371632576,
-0.02204163186252117,
-0.0012046443298459053,
-0.08967050164937973,
-0.016971321776509285,
-0.027370596304535866,
-0.07984733581542969,
0.11313500255346298,
0.015385841019451618,
0.009428275749087334,
0.012268091551959515,
0.04710822552442551,
-0.010343149304389954,
0.011071491986513138,
-0.07533624023199081,
0.02987854741513729,
0.051824912428855896,
-0.027881436049938202,
0.04076743870973587,
-0.015151332132518291,
0.08917621523141861,
-0.004539686720818281,
0.08583591133356094,
0.0005393893807195127,
0.0707937479019165,
-0.03706589713692665,
-0.036174386739730835,
0.0006254129111766815,
-0.10541551560163498,
0.02702840231359005,
0.0018232733709737659,
-0.044100020080804825,
0.031770262867212296,
0.05583501234650612,
0.13221001625061035,
-0.008568032644689083,
-0.01712002046406269,
0.04401423782110214,
0.09494198113679886,
-0.04492151737213135,
-0.0019184710690751672,
0.008397053927183151,
0.06367869675159454,
0.08535695821046829,
-0.04842274636030197,
0.06667076051235199,
0.012979340739548206,
0.0073999324813485146,
0.060587845742702484,
0.10580122470855713,
-0.10354725271463394,
0.08880580961704254,
0.011823901906609535,
0.013296606950461864,
0.022408634424209595,
0.009154172614216805,
-0.0040681553073227406,
0.01640397310256958,
-0.028557881712913513,
0.03242377191781998,
-0.025569146499037743,
-0.022251956164836884,
0.03597143664956093,
0.019510269165039062,
-0.013071248307824135,
-0.0845351293683052,
0.10493066161870956,
-0.011090653017163277,
0.045181646943092346,
0.022646617144346237,
0.03327665850520134,
0.04786797985434532,
-0.022694336250424385,
0.0718829408288002,
0.020418336614966393,
0.00007961031951708719,
-0.05706467851996422,
0.04356518015265465,
0.021570784971117973,
0.06178086996078491,
-0.01411402877420187,
0.033001333475112915,
-0.08649879693984985,
-0.05086858570575714,
-0.07086198031902313,
-0.0047953468747437,
0.0991922989487648,
-0.013024452142417431,
-0.039156101644039154,
-0.07116135954856873,
-0.022856563329696655,
0.052079301327466965,
0.047519750893116,
-0.10378459841012955,
4.404235245307923e-34,
-0.038350045680999756,
0.007898945361375809,
0.04701598361134529,
-0.03272518888115883,
0.06156446412205696,
-0.03472397103905678,
-0.011766362935304642,
0.08717159926891327,
-0.028001300990581512,
-0.024722112342715263,
-0.057095497846603394,
0.044909439980983734,
-0.04157588630914688,
0.039885349571704865,
0.05680263414978981,
-0.026760125532746315,
-0.02560700662434101,
0.013797934167087078,
0.05306203290820122,
-0.032070670276880264,
0.019141405820846558,
0.012168684042990208,
0.036224350333213806,
-0.057787686586380005,
-0.0058266217820346355,
0.02966400608420372,
0.05844726413488388,
-0.014652712270617485,
-0.06725619733333588,
0.03545127436518669,
-0.018864167854189873,
0.004943687468767166,
0.004368307534605265,
0.0026964189019054174,
0.049058761447668076,
-0.10854372382164001,
-0.005532525945454836,
-0.027258630841970444,
0.017266154289245605,
-0.009153594262897968,
0.003671269165351987,
0.025933310389518738,
-0.04363490641117096,
0.03243774548172951,
-0.038031402975320816,
-0.10655198246240616,
0.004297378472983837,
0.009771937504410744,
-0.045733626931905746,
-0.07206391543149948,
-0.019685454666614532,
-0.007810637820512056,
-0.053648464381694794,
-0.04449649155139923,
0.015633495524525642,
-0.026029476895928383,
0.05438714101910591,
0.006657819729298353,
0.026604048907756805,
0.04142793267965317,
0.013088311068713665,
0.049081623554229736,
0.015187857672572136,
0.02447410114109516,
0.03852642700076103,
0.09264498203992844,
-0.0736781656742096,
0.0718902200460434,
0.045079704374074936,
-0.012075493112206459,
0.03518252074718475,
0.0288372952491045,
0.01360931247472763,
0.043720368295907974,
0.10762160271406174,
0.04202546179294586,
0.04288515821099281,
-0.09656932950019836,
-0.06425091624259949,
0.10309960693120956,
0.06201670691370964,
0.009901727549731731,
-0.03010367788374424,
0.01662316359579563,
-0.031089097261428833,
-0.06323112547397614,
0.03258722275495529,
-0.024970632046461105,
0.04116516932845116,
0.006913420744240284,
-0.0786934569478035,
-0.04249965772032738,
0.08103103935718536,
-0.042958829551935196,
-0.06675554811954498,
-1.176954822197125e-35,
0.025800304487347603,
0.06271283328533173,
-0.06096744164824486,
0.04509902745485306,
0.002873766701668501,
-0.02558727189898491,
-0.004924210254102945,
0.03244442865252495,
0.01146641280502081,
0.00034962789504788816,
0.0016641971888020635,
-0.014147520996630192,
0.07868944853544235,
0.021207286044955254,
0.07456102967262268,
-0.03953304514288902,
0.005999524611979723,
-0.0714876651763916,
-0.03397490456700325,
-0.04906127229332924,
-0.021416373550891876,
0.021135471761226654,
-0.04191690310835838,
0.0870497077703476,
-0.034119460731744766,
-0.03597911074757576,
-0.017096471041440964,
0.06538884341716766,
-0.03499842807650566,
-0.006324921268969774,
-0.1199803501367569,
-0.04482674598693848,
-0.013790199533104897,
0.03155127540230751,
-0.08734768629074097,
0.14328090846538544,
0.11200428009033203,
-0.05614437535405159,
-0.04314812272787094,
0.04136224463582039,
0.11925610899925232,
-0.007093119900673628,
-0.009983799420297146,
-0.018312519416213036,
-0.014025500975549221,
-0.009642758406698704,
-0.086153045296669,
0.04455555975437164,
-0.023222336545586586,
0.035544972866773605,
-0.0009429118945263326,
0.004702448379248381,
-0.07516033202409744,
0.027056077495217323,
-0.005546923261135817,
-0.060316864401102066,
-0.026029663160443306,
-0.04496374353766441,
-0.015582025051116943,
0.02920670062303543,
-0.021237395703792572,
-0.03406523913145065,
-0.01844906620681286,
-0.02745932713150978,
-0.02429543249309063,
-0.013192623853683472,
0.032361701130867004,
-0.005873099900782108,
0.07931296527385712,
0.006870730314403772,
0.02472379431128502,
-0.022563036531209946,
0.037263959646224976,
0.025588158518075943,
-0.05745656415820122,
0.017904922366142273,
-0.026726998388767242,
-0.04723912104964256,
-0.031024718657135963,
-0.03942933306097984,
0.027605818584561348,
-0.006715558469295502,
0.03724127635359764,
-0.010767372325062752,
0.15031097829341888,
0.10134829580783844,
0.0495472177863121,
0.006880233995616436,
0.05517877638339996,
0.0038148434832692146,
-0.027053050696849823,
0.050718650221824646,
-0.05406826734542847,
0.026403609663248062,
-0.020586632192134857,
-3.852560226391688e-8,
-0.03876188024878502,
0.0061176312156021595,
0.06590495258569717,
0.0015584357315674424,
0.024914216250181198,
-0.06352536380290985,
0.05985584482550621,
-0.041771359741687775,
0.018218588083982468,
0.02601347491145134,
0.12266101688146591,
-0.09790864586830139,
-0.05813320726156235,
-0.04930723458528519,
0.015899723395705223,
0.024916205555200577,
-0.022838452830910683,
0.029507534578442574,
-0.012763050384819508,
-0.0575750470161438,
0.03911822289228439,
0.09734462946653366,
-0.015975449234247208,
0.02531915344297886,
0.054667577147483826,
-0.08375512808561325,
-0.028956657275557518,
0.06345873326063156,
0.004772016312927008,
-0.001920721959322691,
-0.0030432939529418945,
-0.023774703964591026,
-0.007979895919561386,
-0.021301690489053726,
-0.029054177924990654,
-0.01187894307076931,
0.05826803296804428,
0.07567346841096878,
0.023158961907029152,
-0.038616642355918884,
0.024932052940130234,
0.02105848118662834,
-0.038128647953271866,
0.01017494685947895,
-0.023883072659373283,
0.011782399378716946,
0.04387686401605606,
-0.11250802129507065,
0.01878918521106243,
-0.051706116646528244,
0.09379969537258148,
-0.016034716740250587,
0.026397455483675003,
0.017883548513054848,
0.09769000113010406,
-0.01630636304616928,
0.02794705331325531,
-0.08105874061584473,
-0.03630352392792702,
0.05899614095687866,
-0.028556350618600845,
0.03422820568084717,
-0.02011054754257202,
-0.010898178443312645
] | 0 | 0 | 0 | 0 |
|
Scaling Laws for Adversarial Attacks on Language Model Activations and Tokens
|
https://openreview.net/forum?id=YzxMu1asQi
|
[
"Stanislav Fort"
] |
Poster
|
We explore a class of adversarial attacks targeting the activations of language models to derive upper-bound scaling laws on their attack susceptibility. By manipulating a relatively small subset of model activations, $a$, we demonstrate the ability to control the exact prediction of a significant number (in some cases up to 1000) of subsequent tokens $t$. We empirically verify a scaling law where the maximum number of target tokens predicted, $t_\mathrm{max}$, depends linearly on the number of tokens $a$ whose activations the attacker controls as $t_\mathrm{max} = \kappa a$. We find that the number of bits the attacker controls on the input to exert a single bit of control on the output (a property we call \textit{attack resistance $\chi$}) is remarkably stable between $\approx 16$ and $\approx 25$ over orders of magnitude of model sizes and between model families. Compared to attacks directly on input tokens, attacks on activations are predictably much stronger, however, we identify a surprising regularity where one bit of input steered either via activations or via tokens is able to exert a surprisingly similar amount of control over the model predictions. This gives support for the hypothesis that adversarial attacks are a consequence of dimensionality mismatch between the input and output spaces. A practical implication of the ease of attacking language model activations instead of tokens is for multi-modal and selected retrieval models. By using language models as a controllable test-bed to study adversarial attacks, we explored input-output dimension regimes that are inaccessible in computer vision and greatly extended the empirical support for the dimensionality theory of adversarial attacks.
|
adversarial attacks, language models, scaling laws, activation steering
|
Manipulating just 1 token’s activations in a language model can precisely dictate the subsequent generation of 100s of tokens and we demonstrate a remarkably stable, linear scaling with attack length of this control across model sizes and families
| 11,563 | null |
[
-0.0641142949461937,
-0.08557967096567154,
0.030932804569602013,
0.052540671080350876,
0.06376093626022339,
0.05041678622364998,
0.026686225086450577,
0.036834754049777985,
0.08012886345386505,
0.03903265669941902,
-0.024242084473371506,
-0.057078588753938675,
0.06757272034883499,
-0.0017997972900047898,
-0.009552005678415298,
0.010411862283945084,
0.021996792405843735,
0.06597734242677689,
-0.12747250497341156,
-0.04060888662934303,
-0.01756206713616848,
0.0020217366982251406,
-0.034795429557561874,
-0.015854714438319206,
0.026211166754364967,
-0.06344939768314362,
-0.03038373775780201,
-0.04714323952794075,
-0.0463315024971962,
-0.04026763141155243,
-0.027187421917915344,
-0.014069703407585621,
0.013747386634349823,
0.03676043823361397,
-0.09736859053373337,
-0.02694004587829113,
-0.012502439320087433,
-0.06736090779304504,
0.09848015010356903,
-0.051911838352680206,
0.008167076855897903,
-0.004437094088643789,
0.030578255653381348,
-0.010660168714821339,
-0.017325567081570625,
-0.021729523316025734,
0.04529464244842529,
0.01968660019338131,
-0.06273314356803894,
-0.024894269183278084,
-0.03547695279121399,
0.0016310167266055942,
0.03438020497560501,
0.011458848603069782,
-0.06620336323976517,
-0.10077351331710815,
0.048450957983732224,
0.06884684413671494,
0.06197439879179001,
0.04306826740503311,
-0.07036745548248291,
-0.04782116413116455,
-0.03265731781721115,
0.002210794948041439,
-0.017039917409420013,
-0.03540002927184105,
0.03952182084321976,
0.022125741466879845,
-0.016089148819446564,
0.13921546936035156,
0.03633705899119377,
0.022759797051548958,
-0.07999791949987411,
0.006180615164339542,
0.12997053563594818,
0.025018898770213127,
-0.0012229798594489694,
-0.01303697656840086,
0.024673476815223694,
-0.059250347316265106,
0.004924583248794079,
-0.036153897643089294,
0.052038129419088364,
-0.027662502601742744,
0.06286492943763733,
-0.017794538289308548,
-0.002866957103833556,
-0.02083079144358635,
0.16479210555553436,
0.04853762686252594,
-0.03469837084412575,
-0.004850388970226049,
0.03846099227666855,
0.07348091900348663,
0.12193557620048523,
-0.010874983854591846,
-0.06291767209768295,
-0.06407899409532547,
-0.045879021286964417,
0.08622109889984131,
-0.028227051720023155,
0.024275170639157295,
-0.018747126683592796,
0.0033134380355477333,
0.09505631029605865,
0.00605788454413414,
0.038809485733509064,
-0.0550786554813385,
0.029790153726935387,
-0.02535354159772396,
0.08142097294330597,
0.05025983974337578,
0.015212150290608406,
0.007032585330307484,
0.023557035252451897,
-0.053784433752298355,
0.011879493482410908,
0.005062349606305361,
-0.037792958319187164,
0.028375785797834396,
0.03816278278827667,
0.015907907858490944,
-0.041803259402513504,
-0.008115928620100021,
-0.05877656489610672,
0.026589980348944664,
-0.05855073407292366,
9.903008407743758e-35,
0.05413723364472389,
0.050115589052438736,
-0.0345376618206501,
0.0011721417540684342,
0.06862019002437592,
-0.027494488283991814,
0.00026182099827565253,
-0.007641966920346022,
-0.016678858548402786,
0.016996361315250397,
-0.10224534571170807,
0.02326592244207859,
-0.00012445666652638465,
0.08980575948953629,
0.09887808561325073,
0.08004499971866608,
0.0489056296646595,
0.018042705953121185,
-0.018458167091012,
-0.058990560472011566,
0.06114635989069939,
-0.0008274181163869798,
-0.03338814899325371,
-0.06513678282499313,
-0.0069579388946294785,
-0.03182743117213249,
-0.014566103927791119,
-0.022524060681462288,
-0.07102210819721222,
0.046205073595047,
-0.11807511746883392,
-0.018256189301609993,
0.030992761254310608,
0.04442892596125603,
0.01340063288807869,
-0.005776641890406609,
0.029443372040987015,
-0.036914050579071045,
-0.00512572331354022,
-0.06006605178117752,
-0.014430106617510319,
0.060081224888563156,
-0.025531131774187088,
-0.00620757881551981,
-0.013368479907512665,
-0.05385960265994072,
-0.02306385710835457,
-0.012749862857162952,
-0.10961654782295227,
-0.1106429472565651,
0.012052358128130436,
0.0013026074739173055,
-0.009754957631230354,
0.015407558530569077,
0.020295919850468636,
-0.0592997707426548,
-0.029877061024308205,
0.001674771774560213,
0.04063257947564125,
0.131044402718544,
-0.02452397346496582,
-0.008595887571573257,
0.034126002341508865,
0.055425968021154404,
-0.04655067250132561,
0.029268300160765648,
-0.07648385316133499,
-0.017143506556749344,
0.0439569465816021,
-0.019667671993374825,
-0.003719802014529705,
0.01650785282254219,
-0.007994826883077621,
-0.06843851506710052,
-0.04255666211247444,
-0.010905829258263111,
0.10158497095108032,
-0.07388746738433838,
-0.01156557071954012,
0.08516297489404678,
-0.006343838758766651,
-0.06566093116998672,
0.06648960709571838,
-0.014142709784209728,
-0.08672676235437393,
-0.025269202888011932,
-0.03871582821011543,
-0.0740458071231842,
-0.03470509499311447,
-0.055641040205955505,
-0.030900361016392708,
-0.059041641652584076,
0.029333220794796944,
-0.035394392907619476,
-0.014724647626280785,
-2.2325425847778596e-33,
-0.07384601980447769,
-0.026648443192243576,
-0.03203677386045456,
0.05401218682527542,
-0.10841548442840576,
-0.05638928711414337,
-0.037976160645484924,
0.10383030772209167,
-0.10173982381820679,
-0.03564975783228874,
-0.06971475481987,
-0.004082458559423685,
0.05908766761422157,
-0.04636511206626892,
0.06575600802898407,
-0.12247117608785629,
0.02953164651989937,
0.06045862287282944,
0.05469301715493202,
-0.006537537556141615,
0.06397344172000885,
0.06970960646867752,
-0.05359961465001106,
0.08816366642713547,
0.037884749472141266,
-0.013589267618954182,
-0.1223999559879303,
0.08805723488330841,
0.018627317622303963,
0.009704015217721462,
-0.07691177725791931,
-0.02340267039835453,
0.0126013970002532,
0.04735962674021721,
-0.015103875659406185,
0.07510165870189667,
0.07924327999353409,
0.06337754428386688,
-0.02767198160290718,
0.10733912140130997,
0.06366893649101257,
-0.05475451797246933,
-0.018296627327799797,
0.030091650784015656,
0.017954034730792046,
-0.015152671374380589,
0.05042465403676033,
0.019547412171959877,
0.031337641179561615,
-0.10554851591587067,
0.09088782966136932,
-0.01211254857480526,
0.044764455407857895,
0.07584153860807419,
-0.053978461772203445,
-0.04835604131221771,
0.009845536202192307,
-0.013396726921200752,
0.0331883579492569,
0.014462188817560673,
-0.05832139402627945,
-0.010053185746073723,
-0.015550997108221054,
0.02452169917523861,
-0.018675856292247772,
-0.006246011704206467,
-0.05893766134977341,
0.0367138646543026,
0.08924735337495804,
-0.06096973642706871,
0.042276784777641296,
-0.004188003484159708,
-0.003899088827893138,
0.07398734986782074,
-0.05408083647489548,
-0.044130850583314896,
-0.01072380505502224,
-0.1336345076560974,
-0.017199676483869553,
0.03434403985738754,
-0.036900706589221954,
0.04086337611079216,
0.04155135899782181,
-0.029418259859085083,
0.07856864482164383,
0.004996196366846561,
0.061508819460868835,
0.03820827975869179,
-0.006056291051208973,
0.06344614923000336,
-0.027051378041505814,
0.010727676562964916,
-0.02817547135055065,
0.039897218346595764,
-0.0035705002956092358,
-4.138352949212276e-8,
0.0029857433401048183,
0.05014028772711754,
0.03420543670654297,
0.05443493649363518,
0.0491110198199749,
0.04175431653857231,
-0.04357397183775902,
-0.02176596038043499,
0.015330523252487183,
-0.009843949228525162,
0.06781139969825745,
0.04019995033740997,
-0.034670718014240265,
-0.032338377088308334,
-0.05869913473725319,
0.06373143196105957,
-0.04513568431138992,
-0.014402755536139011,
-0.02910877950489521,
-0.05977780744433403,
-0.00534550566226244,
0.0415475107729435,
0.028588000684976578,
-0.07406695187091827,
-0.012885760515928268,
-0.03440716490149498,
-0.0664428099989891,
0.07485898584127426,
-0.0152326924726367,
0.02707603946328163,
-0.08359017968177795,
-0.020335722714662552,
-0.03325708582997322,
-0.00023561694251839072,
-0.013571918942034245,
0.12921644747257233,
-0.04023014381527901,
-0.056213706731796265,
0.015675533562898636,
-0.033789705485105515,
-0.029374876990914345,
-0.017073940485715866,
-0.00902088638395071,
0.0038436350878328085,
-0.0077357846312224865,
0.006922553293406963,
-0.04992123693227768,
-0.06614845246076584,
0.07133190333843231,
-0.014802603982388973,
0.0694216787815094,
0.03629913553595543,
-0.004173973109573126,
0.10644572973251343,
-0.01628107577562332,
0.010967203415930271,
-0.07136538624763489,
-0.045674361288547516,
-0.025354577228426933,
0.06724154204130173,
0.08256236463785172,
0.028444437310099602,
-0.05604703724384308,
-0.036136195063591
] | 0 | 0 | 0 | 0 |
|
BEEM: Boosting Performance of Early Exit DNNs using Multi-Exit Classifiers as Experts
|
https://openreview.net/forum?id=EzrZX9bd4G
|
[
"Divya Jyoti Bajpai",
"Manjesh Kumar Hanawal"
] |
Poster
|
Early Exit (EE) techniques have emerged as a means to reduce inference latency in Deep Neural Networks (DNNs). The latency improvement and accuracy in these techniques crucially depend on the criteria used to make exit decisions. We propose a new decision criterion BEEM where exit classifiers are treated as experts and aggregate their confidence scores. The confidence scores are aggregated only if neighbouring experts are consistent in prediction as the samples pass through them, thus capturing their ensemble effect. A sample exits when the aggregated confidence value exceeds a threshold. The threshold is set using the error rates of the intermediate exits aiming to surpass the performance of conventional DNN inference. Experimental results on the COCO dataset for Image captioning and GLUE datasets for various language tasks demonstrate that our method enhances the performance of state-of-the-art EE methods, achieving improvements in speed-up by a factor $1.5\times$ to $2.1\times$. When compared to the final layer, its accuracy is comparable in harder Image Captioning and improves in the easier language tasks. The source code is available at https://github.com/Div290/BEEM1/tree/main.
|
Early Exits; Expert-based exiting
|
An exiting criteria for early exits with theoretical analysis
| 11,559 |
2502.00745
|
[
0.008233646862208843,
-0.07578174024820328,
0.0783725157380104,
0.04778129979968071,
0.08461906760931015,
0.0387737937271595,
0.01025926973670721,
0.01477313693612814,
0.02476673386991024,
-0.06754868477582932,
-0.07887712866067886,
-0.1296122819185257,
0.008197926916182041,
-0.006182592827826738,
-0.06443873792886734,
-0.018994957208633423,
0.10618194937705994,
0.09934279322624207,
-0.10239719599485397,
-0.04808790981769562,
0.0031534412410110235,
-0.007788217626512051,
0.07537341117858887,
0.001046400866471231,
0.07838217169046402,
-0.012430176138877869,
-0.06234080716967583,
-0.006601634901016951,
0.031038640066981316,
-0.046520136296749115,
-0.03581461310386658,
-0.055301766842603683,
0.0022498290054500103,
0.021014109253883362,
-0.038097430020570755,
0.039952486753463745,
-0.01894061081111431,
0.03260178118944168,
0.04317764565348625,
-0.042436402291059494,
0.01415383443236351,
0.02968347817659378,
0.05714716762304306,
0.0586954690515995,
0.10363718122243881,
0.08358369022607803,
-0.009891461580991745,
-0.04994582384824753,
0.030813338235020638,
0.02406044863164425,
-0.018604042008519173,
0.011025584302842617,
0.013835345394909382,
0.09066618233919144,
-0.06939545273780823,
0.02809075452387333,
0.10464093089103699,
0.012143081985414028,
-0.007971145212650299,
0.023224104195833206,
-0.021635646000504494,
-0.09791572391986847,
0.006125543732196093,
-0.03165372833609581,
0.07153357565402985,
-0.03416505828499794,
0.0368911512196064,
-0.009469538927078247,
0.00528767891228199,
-0.005127920303493738,
-0.016026116907596588,
0.04437587037682533,
-0.05686367303133011,
0.07294056564569473,
0.025075605139136314,
0.025935335084795952,
0.03990036994218826,
-0.05217842757701874,
0.037302933633327484,
-0.06000295653939247,
0.025604480877518654,
-0.00047421487397514284,
0.03738599643111229,
0.02195512317121029,
0.11580194532871246,
-0.003856835188344121,
-0.033139701932668686,
0.009447496384382248,
-0.043258968740701675,
-0.037795837968587875,
-0.060514211654663086,
-0.06146061047911644,
0.03881014883518219,
0.031224677339196205,
0.017872417345643044,
0.037197332829236984,
-0.060627128928899765,
-0.0685337483882904,
-0.008170416578650475,
0.070429727435112,
-0.03503916785120964,
-0.02774919755756855,
0.00422350037842989,
-0.09515231102705002,
0.06763982027769089,
-0.03332971781492233,
0.09139586240053177,
0.007947319187223911,
0.06949052214622498,
0.002219732850790024,
0.021550090983510017,
0.0045008533634245396,
-0.03998455032706261,
0.022141462191939354,
0.06487137824296951,
0.03072121925652027,
-0.04026170074939728,
-0.0028919975738972425,
0.04194154962897301,
0.024876905605196953,
-0.06272396445274353,
0.08356021344661713,
-0.004996189847588539,
-0.011067754589021206,
0.03881465271115303,
-0.040872570127248764,
-0.04337629675865173,
3.9437537923080194e-33,
0.055176012217998505,
-0.06733153015375137,
0.04541916027665138,
-0.008506379090249538,
0.060601115226745605,
-0.006451083347201347,
-0.03743314743041992,
0.044627342373132706,
-0.04585094377398491,
-0.03440789133310318,
-0.07053421437740326,
-0.11895820498466492,
-0.025870652869343758,
0.015378246083855629,
0.05262613296508789,
0.07164587825536728,
-0.014457757584750652,
-0.01307759340852499,
-0.030362138524651527,
0.0513281412422657,
0.05482105538249016,
-0.0513206347823143,
0.027673790231347084,
0.0052896104753017426,
0.009980435483157635,
0.025873037055134773,
0.01870482973754406,
-0.02465588040649891,
0.03034873865544796,
-0.00421347888186574,
-0.11675313115119934,
-0.09104057401418686,
0.023570992052555084,
0.002708354964852333,
0.0757574588060379,
-0.07034014165401459,
-0.00010757453856058419,
0.031887806951999664,
0.016279906034469604,
-0.01653764769434929,
-0.054051823914051056,
-0.00719836913049221,
-0.060375384986400604,
0.015717003494501114,
-0.04673637077212334,
-0.06057574972510338,
0.0004179900570306927,
0.009498328901827335,
-0.03191797807812691,
0.01046393346041441,
0.024961961433291435,
-0.03223474696278572,
-0.06625155359506607,
-0.02672697603702545,
0.013768022879958153,
-0.034078143537044525,
0.03195202350616455,
0.023380562663078308,
0.024320701137185097,
-0.036691173911094666,
-0.04966350272297859,
0.0786946639418602,
-0.052800826728343964,
0.07602816820144653,
0.003002636134624481,
0.03519860655069351,
-0.016979960724711418,
0.07087960094213486,
-0.0009478168794885278,
-0.08561836928129196,
-0.041732460260391235,
-0.028984710574150085,
-0.010606979951262474,
-0.0405401736497879,
0.012182199396193027,
0.05212407186627388,
0.036265723407268524,
-0.06452181190252304,
0.053130194544792175,
0.043768156319856644,
0.023125842213630676,
-0.004075314849615097,
-0.08988029509782791,
-0.08092430233955383,
-0.031627148389816284,
0.0122910737991333,
0.058738164603710175,
-0.08067601174116135,
0.01806540973484516,
0.049774277955293655,
-0.0930539071559906,
-0.03823921084403992,
-0.06844549626111984,
0.03929009288549423,
-0.08673713356256485,
-2.248642816622294e-33,
0.024084635078907013,
0.10786387324333191,
-0.0401684045791626,
0.08863752335309982,
-0.05478906258940697,
-0.01108984462916851,
0.034913480281829834,
0.0018486197805032134,
0.001709154574200511,
-0.04091491550207138,
0.008553577587008476,
-0.03084077313542366,
0.04883011803030968,
-0.025055773556232452,
0.003014156362041831,
-0.0033604451455175877,
-0.03442809730768204,
0.02558393031358719,
0.043483179062604904,
0.05212727561593056,
0.02562151849269867,
0.08297241479158401,
-0.10953094065189362,
0.02964257448911667,
-0.07816489040851593,
0.0007101669325493276,
-0.04963470995426178,
0.0611630417406559,
-0.06381425261497498,
-0.07863812148571014,
0.02578778937458992,
-0.0665285512804985,
-0.001712112920358777,
0.047846127301454544,
0.05319115146994591,
0.041871942579746246,
0.00737595371901989,
-0.11743012070655823,
-0.05008089542388916,
0.04496059939265251,
0.08704862743616104,
-0.0389152392745018,
-0.014359134249389172,
-0.034681789577007294,
-0.0554325133562088,
-0.003091210965067148,
-0.08982779085636139,
0.0357733853161335,
-0.007842827588319778,
0.08347538858652115,
0.0029194147791713476,
-0.06186536327004433,
-0.07665544748306274,
0.047568146139383316,
-0.03351268917322159,
-0.05768387019634247,
-0.002661312697455287,
0.0359727181494236,
0.07955881208181381,
0.044514670968055725,
-0.03293268010020256,
-0.01564468815922737,
0.05012057349085808,
-0.04623863473534584,
0.030860161408782005,
-0.023404086008667946,
-0.06528311967849731,
0.09788209199905396,
0.021770581603050232,
0.031225398182868958,
0.05454869568347931,
-0.016769031062722206,
0.05904014781117439,
0.018813643604516983,
-0.07085948437452316,
0.040745098143815994,
0.04144075885415077,
-0.043030716478824615,
0.0008304507355205715,
-0.06441868096590042,
-0.044156257063150406,
-0.05341147258877754,
0.021041279658675194,
0.10448683053255081,
0.13580124080181122,
0.0959252193570137,
-0.0027594072744250298,
0.0375865213572979,
0.06780774146318436,
0.080296590924263,
-0.03331645950675011,
0.02409476786851883,
0.04359838739037514,
0.028248030692338943,
-0.06230032816529274,
-4.647166207405462e-8,
-0.07164071500301361,
0.06347616761922836,
0.02526572160422802,
0.04938080161809921,
0.03966968506574631,
-0.09850388020277023,
-0.009236115030944347,
0.10195436328649521,
-0.0011055039940401912,
0.010534970089793205,
0.14948932826519012,
-0.0028634383343160152,
-0.04030119255185127,
-0.022987736389040947,
-0.029716448858380318,
0.010858109220862389,
0.08277098834514618,
0.07437369972467422,
0.04467650502920151,
-0.06905511021614075,
0.0036453346256166697,
-0.03354869782924652,
0.03934550657868385,
-0.102602019906044,
-0.016494054347276688,
-0.06543074548244476,
0.02496684342622757,
0.09472806751728058,
-0.0347219854593277,
-0.02773556113243103,
0.00395599752664566,
0.011829298920929432,
0.035500336438417435,
0.007225198205560446,
0.053171027451753616,
0.09186768531799316,
-0.004330225754529238,
0.015325143933296204,
0.029411226511001587,
0.02447255700826645,
-0.004008058458566666,
0.03568115457892418,
-0.050893865525722504,
0.011676009744405746,
0.008216208778321743,
-0.022074101492762566,
0.04905695095658302,
-0.0523291677236557,
0.03159373253583908,
-0.011992794461548328,
0.018652938306331635,
-0.08085677772760391,
-0.030070042237639427,
0.12305787205696106,
0.06408658623695374,
0.003639815840870142,
-0.03573453798890114,
-0.04803602769970894,
-0.03909189999103546,
0.09676115959882736,
-0.03157486021518707,
-0.04061534255743027,
-0.04189993441104889,
0.0006295585772022605
] | 0 | 0 | 0 | 0 |
|
Real-time design of architectural structures with differentiable mechanics and neural networks
|
https://openreview.net/forum?id=Tpjq66xwTq
|
[
"Rafael Pastrana",
"Eder Medina",
"Isabel M. de Oliveira",
"Sigrid Adriaenssens",
"Ryan P Adams"
] |
Poster
|
Designing mechanically efficient geometry for architectural structures like shells, towers, and bridges, is an expensive iterative process.
Existing techniques for solving such inverse problems rely on traditional optimization methods, which are slow and computationally expensive, limiting iteration speed and design exploration.
Neural networks would seem to offer a solution via data-driven amortized optimization, but they often require extensive fine-tuning and cannot ensure that important design criteria, such as mechanical integrity, are met.
In this work, we combine neural networks with a differentiable mechanics simulator to develop a model that accelerates the solution of shape approximation problems for architectural structures represented as bar systems.
This model explicitly guarantees compliance with mechanical constraints while generating designs that closely match target geometries.
We validate our approach in two tasks, the design of masonry shells and cable-net towers.
Our model achieves better accuracy and generalization than fully neural alternatives, and comparable accuracy to direct optimization but in real time, enabling fast and reliable design exploration.
We further demonstrate its advantages by integrating it into 3D modeling software and fabricating a physical prototype.
Our work opens up new opportunities for accelerated mechanical design enhanced by neural networks for the built environment.
|
Differentiable physics, mechanical design, physics-in-the-loop neural networks, inverse problems, architectural structures
|
We couple neural networks with a differentiable mechanics simulator to accelerate the solution of shape-matching problems for mechanical design.
| 11,555 |
2409.02606
|
[
-0.07169921696186066,
0.004997112788259983,
0.06552983820438385,
0.0014274325221776962,
-0.04590301588177681,
-0.09000984579324722,
-0.07347551733255386,
-0.023626402020454407,
-0.033651284873485565,
-0.004968834109604359,
-0.07621254771947861,
-0.016672836616635323,
-0.02018965594470501,
0.06500571966171265,
-0.08471273630857468,
0.017258791252970695,
0.024408061057329178,
0.08111780881881714,
-0.05844143405556679,
0.045221343636512756,
-0.003629678161814809,
0.025181809440255165,
0.002188545186072588,
-0.04645496606826782,
0.04063055291771889,
0.011362438090145588,
0.011092090979218483,
0.037176940590143204,
-0.023705830797553062,
-0.0031293013598769903,
0.09409008175134659,
-0.0897546038031578,
-0.06877946108579636,
0.004319414030760527,
-0.026961809024214745,
0.0876426175236702,
0.014387384988367558,
0.025356082245707512,
-0.031570129096508026,
-0.08861093968153,
-0.05230613052845001,
0.05360979959368706,
0.03152995929121971,
0.04207413271069527,
0.13854843378067017,
0.024829048663377762,
-0.011821169406175613,
-0.08693405240774155,
-0.0019443577621132135,
0.048609793186187744,
-0.05354796722531319,
-0.003684266936033964,
0.01358382124453783,
-0.04073585942387581,
0.03512416407465935,
0.08305510878562927,
0.018552266061306,
-0.025832461193203926,
0.001421416993252933,
-0.015070798806846142,
0.08693529665470123,
-0.039856918156147,
-0.028008967638015747,
-0.02940240316092968,
-0.042108204215765,
0.055952224880456924,
0.04932330176234245,
0.029055211693048477,
0.026846108958125114,
0.01951782777905464,
0.080911785364151,
0.005617126356810331,
0.0006288900622166693,
-0.004906786605715752,
0.04274242743849754,
0.011139431037008762,
0.04976249858736992,
0.02591894008219242,
-0.031234754249453545,
-0.03115260787308216,
-0.04499168321490288,
0.007451250683516264,
-0.023706089705228806,
-0.014113660901784897,
0.0025326109025627375,
0.010179735720157623,
-0.06656001508235931,
0.1509450376033783,
0.0038666401524096727,
0.01165238581597805,
0.03301042690873146,
0.04864177107810974,
-0.018907541409134865,
-0.06137072667479515,
0.04762222245335579,
-0.03833003714680672,
-0.01156532857567072,
-0.06838510930538177,
-0.0763443186879158,
0.06816867738962173,
-0.05428943410515785,
0.031426332890987396,
0.06342054158449173,
0.0018915246473625302,
0.03969539329409599,
0.04971940815448761,
0.051186077296733856,
0.029332222416996956,
-0.0077655501663684845,
-0.053297460079193115,
0.036267172545194626,
-0.02892296202480793,
0.002743376651778817,
-0.005996685940772295,
-0.0034829103387892246,
-0.04419384151697159,
-0.020647037774324417,
-0.10141680389642715,
0.08856860548257828,
0.1088794469833374,
0.02635694108903408,
-0.014434895478188992,
-0.006596692837774754,
0.0372115783393383,
0.00008453705231659114,
-0.06295289099216461,
-0.1293906420469284,
1.8740263586686624e-34,
-0.054468072950839996,
0.027011718600988388,
0.018036993220448494,
-0.00015669989807065576,
0.10994177311658859,
-0.055142179131507874,
0.0992206260561943,
0.08672849088907242,
0.06125751510262489,
-0.007639471907168627,
-0.07465619593858719,
-0.04928090050816536,
-0.09198112040758133,
0.07195313274860382,
0.10575295239686966,
-0.03223599120974541,
-0.019749799743294716,
-0.014192027971148491,
-0.020328963175415993,
0.022264810279011726,
0.006344183813780546,
-0.11750011891126633,
0.023253073915839195,
0.011803075671195984,
0.008637445978820324,
-0.026738906279206276,
0.10376682132482529,
0.07388786971569061,
-0.09713223576545715,
-0.01518820971250534,
-0.03203016519546509,
0.019228054210543633,
0.033337488770484924,
0.07241007685661316,
0.030325857922434807,
0.007768305949866772,
0.05595492944121361,
-0.007305238861590624,
0.016802040860056877,
0.03857671841979027,
0.004449345637112856,
-0.00009012693772092462,
0.03463761508464813,
0.0807717964053154,
-0.05921972170472145,
-0.012168267741799355,
0.021116791293025017,
0.04618126526474953,
-0.030971450731158257,
-0.05218588933348656,
0.008917432278394699,
0.08381403982639313,
0.010136394761502743,
-0.024434484541416168,
-0.026645777747035027,
-0.08555422723293304,
0.012824933044612408,
0.022651689127087593,
0.03982766717672348,
0.06560014933347702,
-0.027171676978468895,
0.011390241794288158,
0.018475744873285294,
0.04895009100437164,
-0.07789992541074753,
0.010615862905979156,
0.005316100548952818,
0.08999993652105331,
0.039703141897916794,
-0.03640010952949524,
-0.04536449909210205,
-0.042979028075933456,
-0.01864824816584587,
-0.03585382178425789,
-0.017455339431762695,
0.07243222743272781,
0.016021816059947014,
-0.005857151001691818,
-0.014574971981346607,
0.008275614120066166,
-0.012551266700029373,
0.0932302325963974,
0.023040207102894783,
0.015261958353221416,
-0.005397197790443897,
-0.0224117673933506,
0.14307253062725067,
-0.023867139592766762,
-0.03437606617808342,
-0.07360295951366425,
-0.09444176405668259,
-0.0897059217095375,
0.03716383874416351,
-0.030538106337189674,
0.011453519575297832,
-6.381598577673403e-34,
-0.01784050278365612,
-0.015303337946534157,
-0.04799838364124298,
-0.006142135709524155,
-0.005882109981030226,
-0.016961894929409027,
-0.1255825161933899,
-0.07202549278736115,
-0.019017791375517845,
-0.02070690505206585,
0.010136831551790237,
-0.004584162496030331,
0.08214287459850311,
-0.0005316428723745048,
0.005267668515443802,
0.04402482509613037,
-0.052531227469444275,
-0.10098505765199661,
-0.006566241383552551,
0.004529600031673908,
0.041385166347026825,
0.10619024932384491,
-0.08734830468893051,
0.007238228805363178,
-0.07120254635810852,
-0.020494369789958,
-0.10158040374517441,
-0.0000441246411355678,
0.04170013964176178,
0.06231241673231125,
-0.07525445520877838,
-0.024162553250789642,
0.04242701455950737,
0.060911569744348526,
0.0214877650141716,
0.05310826003551483,
0.050535377115011215,
0.012154688127338886,
0.0852789655327797,
-0.12232223898172379,
0.0598573200404644,
-0.03775406628847122,
-0.05220712348818779,
0.014330334030091763,
0.0013585821725428104,
-0.012567034922540188,
-0.014288698323071003,
-0.051577962934970856,
-0.09331481158733368,
0.034694649279117584,
0.057144053280353546,
0.0727875679731369,
0.005937278736382723,
-0.12586486339569092,
0.008887706324458122,
-0.061312075704336166,
-0.028248703107237816,
0.003538558492437005,
0.15911076962947845,
0.022288408130407333,
-0.04397142305970192,
-0.09347386658191681,
0.047882236540317535,
-0.025605153292417526,
0.005206135101616383,
0.038325145840644836,
0.003743233159184456,
-0.033776044845581055,
-0.03159717842936516,
0.021562157198786736,
-0.03945218771696091,
0.017836997285485268,
0.027276473119854927,
0.03860519081354141,
-0.05153811350464821,
-0.0957399532198906,
0.03608984127640724,
0.012731743045151234,
-0.030041325837373734,
0.03676064684987068,
0.039977386593818665,
-0.0552351288497448,
0.0352739617228508,
0.08218266069889069,
0.021383017301559448,
-0.00308169424533844,
-0.03329414129257202,
0.09321556985378265,
0.004502885043621063,
0.04163090139627457,
0.050499945878982544,
0.05882401764392853,
-0.023859333246946335,
-0.023611675947904587,
-0.050821200013160706,
-4.0152755786948546e-8,
-0.05974423885345459,
0.03302746266126633,
0.009083620272576809,
-0.011513927020132542,
-0.012879230082035065,
0.013193204998970032,
0.0711105614900589,
-0.04443266615271568,
-0.0232076533138752,
-0.10829892009496689,
0.08279585093259811,
-0.03716638684272766,
0.00230093183927238,
-0.0137984873726964,
-0.05988467484712601,
0.04068918153643608,
-0.04035906866192818,
0.005720529705286026,
0.000948331318795681,
-0.017813581973314285,
-0.028295783326029778,
0.05235603451728821,
0.053476013243198395,
0.0013037617318332195,
0.016789529472589493,
-0.06498566269874573,
-0.1454770714044571,
-0.00042938411934301257,
0.14967340230941772,
0.033844929188489914,
0.005122395697981119,
-0.02162882871925831,
0.06546639651060104,
-0.01929963193833828,
0.022556284442543983,
0.02649649791419506,
0.06268902122974396,
-0.016458293423056602,
-0.043997637927532196,
-0.05230269208550453,
-0.007784545421600342,
-0.006933100055903196,
-0.04479460418224335,
-0.0095160361379385,
0.04191578924655914,
-0.05209538713097572,
-0.027843527495861053,
-0.05338526517152786,
0.05885039642453194,
0.03808310255408287,
0.027225693687796593,
0.04534266144037247,
-0.02565651945769787,
-0.018443619832396507,
0.018018575385212898,
0.04318483546376228,
0.013807659037411213,
-0.04397410526871681,
0.06176391988992691,
0.025037167593836784,
0.006183518562465906,
0.028053043410182,
-0.023156413808465004,
-0.007058772724121809
] | 0 | 0 | 0 | 0 |
|
DAMO: Decoding by Accumulating Activations Momentum for Mitigating Hallucinations in Vision-Language Models
|
https://openreview.net/forum?id=JUr0YOMvZA
|
[
"Kaishen Wang",
"Hengrui Gu",
"Meijun Gao",
"Kaixiong Zhou"
] |
Poster
|
Large Vision-Language Models (VLMs) exhibit significant potential in multimodal tasks but often struggle with hallucinations—responses that are plausible yet visually ungrounded. In this work, we investigate the layer-wise prediction tendencies of VLMs and conduct an in-depth analysis of their decoding mechanism. We observe that VLMs tend to ``overthink'' during the final stages of decoding, making significant prediction shifts in the last few layers often favoring incorrect results, which leads to a surge in hallucinative outputs. Leveraging this localized pattern, we propose a novel decoding strategy inspired by the momentum analogy used in gradient descent-based optimizers. Our method enforces decoding consistency across layers in an adaptive manner during forward passes—an under-explored approach in existing works. This strategy significantly improves the reliability and performance of VLMs in various multimodal tasks, while introducing only negligible efficiency overhead.
|
Vision-Language Models (VLMs), Hallucinations, Decoding Method, Momentum Techniques
|
To address the hallucination problem in Vision-Language Models (VLMs), we propose a novel decoding method inspired by momentum techniques.
| 11,551 | null |
[
0.015382646583020687,
-0.08538604527711868,
0.03089175745844841,
-0.00692637637257576,
0.07562548667192459,
-0.0003058479633182287,
-0.03629831597208977,
-0.07417219877243042,
0.057687338441610336,
-0.0655575692653656,
-0.07858014106750488,
-0.06482745707035065,
-0.007274919655174017,
0.029580768197774887,
-0.013939118944108486,
-0.031555429100990295,
0.03876691684126854,
0.15462340414524078,
-0.09410033375024796,
-0.05408218875527382,
0.07478383183479309,
-0.007110085338354111,
0.007772689685225487,
-0.011524059809744358,
0.10619434714317322,
0.07815820723772049,
-0.03605073690414429,
-0.037433452904224396,
0.029765408486127853,
-0.02779397927224636,
0.0322754792869091,
0.03415604308247566,
-0.011415495537221432,
0.014865368604660034,
-0.040688108652830124,
0.06920893490314484,
-0.0742463767528534,
0.013950410299003124,
-0.017373502254486084,
-0.04501091688871384,
-0.04471416771411896,
0.05306888371706009,
-0.0018416859675198793,
0.007269867230206728,
0.03167961165308952,
0.03834874555468559,
0.02544037066400051,
0.0005308359395712614,
0.006950328126549721,
-0.07273908704519272,
-0.033392369747161865,
0.009190988726913929,
0.02550811879336834,
0.007529131602495909,
-0.07050644606351852,
0.0628289133310318,
0.0850304588675499,
0.029512926936149597,
-0.04399741068482399,
-0.01955251768231392,
-0.07535149902105331,
-0.0663725957274437,
-0.0001718536950647831,
0.0026746548246592283,
-0.016146443784236908,
0.04183703660964966,
0.03579307347536087,
0.021408390253782272,
0.004311487544327974,
0.05024215579032898,
0.006280567497014999,
0.02635372057557106,
-0.03162958845496178,
0.008561327122151852,
0.09215972572565079,
-0.004703623242676258,
-0.002436497015878558,
-0.10952016711235046,
0.030970895662903786,
-0.08376384526491165,
0.03485063835978508,
0.0037951546255499125,
0.02839815616607666,
-0.03033524565398693,
0.1307372748851776,
-0.0039053054060786963,
-0.05969047546386719,
0.013569176197052002,
0.007372918538749218,
0.03194054961204529,
-0.10264888405799866,
-0.18062394857406616,
-0.03472139686346054,
-0.026574144139885902,
0.040448352694511414,
-0.002743405755609274,
-0.015031685121357441,
-0.05656970664858818,
-0.013141423463821411,
0.028243422508239746,
0.008168620057404041,
-0.01548459567129612,
-0.023423489183187485,
-0.08418095856904984,
-0.0008511951309628785,
-0.014037424698472023,
0.11223495751619339,
0.09399410337209702,
-0.005586068145930767,
-0.008228827267885208,
0.035601869225502014,
0.04527796804904938,
0.02736620418727398,
0.020812394097447395,
-0.0044067613780498505,
0.004848358687013388,
-0.007632690481841564,
-0.0157662034034729,
0.02202877402305603,
0.027172356843948364,
-0.01703466847538948,
0.001816906500607729,
-0.0225367471575737,
-0.0036383052356541157,
0.014257282018661499,
-0.07243027538061142,
-0.06450952589511871,
7.304309296836089e-34,
0.0028956090100109577,
0.022540200501680374,
0.011918557807803154,
0.04638805240392685,
0.06745554506778717,
0.010532958433032036,
-0.01637781597673893,
-0.05538047477602959,
0.04965700954198837,
-0.009550431743264198,
-0.05460858717560768,
-0.0557042695581913,
-0.05724652484059334,
0.08462056517601013,
-0.02431206777691841,
-0.011446323245763779,
-0.06193060055375099,
0.027359316125512123,
-0.08228102326393127,
0.023454071953892708,
0.06003190949559212,
0.00037172785960137844,
-0.06643400341272354,
-0.018313346430659294,
-0.049421269446611404,
0.05122160166501999,
0.021984050050377846,
-0.09408322721719742,
-0.059646524488925934,
0.0326681062579155,
-0.13977277278900146,
0.06338104605674744,
-0.02796591818332672,
0.002371203387156129,
0.053534843027591705,
0.02097189985215664,
0.05232368782162666,
-0.010458720847964287,
0.025534508749842644,
-0.004069504328072071,
-0.02302691340446472,
0.07251797616481781,
-0.040987174957990646,
-0.05401688069105148,
-0.09913923591375351,
-0.025726092979311943,
0.000015826975868549198,
0.05628068000078201,
-0.08868075162172318,
-0.03340890631079674,
0.044198181480169296,
0.007467735558748245,
-0.01656433381140232,
0.004988977685570717,
0.02334209531545639,
-0.06063277646899223,
0.05249188467860222,
-0.02388094924390316,
0.06327130645513535,
-0.008382195606827736,
0.09613432735204697,
0.09076476842164993,
0.0003111300466116518,
0.07719264179468155,
-0.018102305009961128,
0.009598340839147568,
-0.02631792053580284,
0.029408320784568787,
-0.0073438710533082485,
-0.061110835522413254,
-0.06951913982629776,
0.04907842352986336,
-0.027849217876791954,
-0.0312836579978466,
0.04117973893880844,
0.022721322253346443,
0.02223031036555767,
-0.022145144641399384,
0.01207783818244934,
0.05422777682542801,
0.020222369581460953,
0.019188933074474335,
-0.008582529611885548,
-0.007997849956154823,
-0.020730553194880486,
0.017313281074166298,
0.021225688979029655,
-0.11764001101255417,
-0.11908111721277237,
-0.07775942236185074,
-0.02163972519338131,
-0.06236710771918297,
0.06510436534881592,
-0.06358882039785385,
-0.05257558077573776,
-2.6598665563834318e-33,
-0.010611047968268394,
0.022298820316791534,
0.08025187999010086,
0.019750675186514854,
-0.053546324372291565,
-0.05032164603471756,
0.05088477581739426,
-0.03664165362715721,
0.02051347866654396,
-0.06298662722110748,
-0.03730008751153946,
0.016689876094460487,
-0.012801403179764748,
0.023531915619969368,
0.09880539774894714,
-0.08042188733816147,
0.05036839097738266,
0.010741481557488441,
0.02058759145438671,
0.018356874585151672,
-0.005604259204119444,
0.08109207451343536,
-0.10308621823787689,
-0.02713763900101185,
-0.036270998418331146,
0.05200522020459175,
-0.07362958788871765,
0.06680545955896378,
0.0737476721405983,
-0.0479632131755352,
-0.009722582064568996,
0.0758754312992096,
-0.011882364749908447,
-0.026802413165569305,
-0.03905370086431503,
0.12535980343818665,
-0.039700791239738464,
-0.026600617915391922,
-0.10251407325267792,
0.05581284314393997,
0.027761412784457207,
-0.02818801812827587,
-0.03209840506315231,
0.02925482951104641,
0.014719380997121334,
-0.06993844360113144,
-0.048442136496305466,
-0.023998888209462166,
-0.028282811865210533,
0.044679827988147736,
0.00630264263600111,
0.07341373711824417,
-0.10516295582056046,
-0.016784219071269035,
-0.08030296862125397,
-0.03312315791845322,
-0.1019081324338913,
0.018034882843494415,
0.11548087745904922,
-0.020048830658197403,
-0.028611648827791214,
0.02028650976717472,
-0.01888224110007286,
-0.012364188209176064,
-0.012100771069526672,
0.040357738733291626,
-0.0007267020409926772,
0.0019542055670171976,
0.07292909175157547,
0.031734030693769455,
0.04928283020853996,
-0.017661022022366524,
0.008456401526927948,
0.10938942432403564,
0.03796306997537613,
-0.04005555808544159,
-0.012511763721704483,
-0.0051115890964865685,
0.007207398768514395,
-0.012306904420256615,
-0.015735631808638573,
-0.04178449884057045,
-0.020757857710123062,
0.0655258521437645,
0.09380161017179489,
0.028445759788155556,
-0.0626811683177948,
0.059093914926052094,
0.003121740184724331,
-0.01982366479933262,
-0.020372971892356873,
-0.01527896523475647,
0.04645444080233574,
0.05533719062805176,
-0.03084135241806507,
-4.386214058627047e-8,
-0.08891347795724869,
0.014298729598522186,
-0.007889335043728352,
-0.001067484263330698,
-0.005420652683824301,
-0.08598052710294724,
-0.007990431040525436,
0.0374302864074707,
0.0038790905382484198,
-0.11763180047273636,
0.11468404531478882,
-0.059495117515325546,
0.013864601962268353,
-0.07938717305660248,
-0.016079237684607506,
0.09990262985229492,
0.00809448305517435,
0.0626436173915863,
-0.013607443310320377,
-0.024982035160064697,
0.04197445884346962,
0.03963541239500046,
-0.055475473403930664,
-0.015889303758740425,
0.010480193421244621,
0.025732651352882385,
-0.04277120903134346,
0.053098514676094055,
0.04397546499967575,
-0.05510587617754936,
0.07528071105480194,
0.08077172189950943,
0.027376560494303703,
0.008651075884699821,
-0.03811898082494736,
0.06516803056001663,
0.0005459858221001923,
0.023040875792503357,
0.02450963854789734,
-0.036849379539489746,
0.012967949733138084,
-0.033069342374801636,
0.005705272778868675,
0.0518583320081234,
-0.005058565642684698,
0.023009629920125008,
0.026800423860549927,
-0.07762888818979263,
0.006939233280718327,
-0.049476560205221176,
-0.004410224501043558,
0.12770240008831024,
-0.018568942323327065,
0.1248222291469574,
0.07630796730518341,
-0.005587542895227671,
0.06523514539003372,
-0.06513693928718567,
0.0038360957987606525,
0.04928520321846008,
0.0051375520415604115,
0.05008389428257942,
-0.10639388114213943,
0.026505958288908005
] | 0 | 0 | 0 | 0 |
|
Talking Turns: Benchmarking Audio Foundation Models on Turn-Taking Dynamics
|
https://openreview.net/forum?id=2e4ECh0ikn
|
[
"Siddhant Arora",
"Zhiyun Lu",
"Chung-Cheng Chiu",
"Ruoming Pang",
"Shinji Watanabe"
] |
Poster
|
The recent wave of audio foundation models (FMs) could provide new capabilities for conversational modeling. However, there have been limited efforts to evaluate these audio FMs comprehensively on their ability to have natural and interactive conversations. To engage in meaningful conversation with the end user, we would want the FMs to additionally perform a fluent succession of turns without too much overlapping speech or long stretches of silence. Inspired by this, we ask whether the recently proposed audio FMs can understand, predict, and perform turn-taking events? To answer this, we propose a novel evaluation protocol that can assess spoken dialog system's turn-taking capabilities using a supervised model as a judge that has been trained to predict turn-taking events in human-human conversations. Using this protocol, we present the first comprehensive user study that evaluates existing spoken dialogue systems on their ability to perform turn-taking events and reveal many interesting insights, such as they sometimes do not understand when to speak up, can interrupt too aggressively and rarely backchannel. We further evaluate multiple open-source and proprietary audio FMs accessible through APIs on carefully curated test benchmarks from Switchboard to measure their ability to understand and predict turn-taking events and identify significant room for improvement. We will open source our evaluation platform to promote the development of advanced conversational AI systems.
|
Turn-taking, Conversation AI, Audio Foundation Models, Evaluation Metric, Evaluation Benchmark
|
We propose a novel evaluation protocol using a turn-taking judge model to automatically assess spoken dialog systems, providing valuable insights into their turn-taking capabilities.
| 11,546 |
2503.01174
|
[
-0.01949799805879593,
-0.1084432527422905,
0.04895799234509468,
0.02127857692539692,
0.06454510986804962,
-0.036524105817079544,
0.027762651443481445,
-0.015236110426485538,
0.03606537729501724,
-0.07479816675186157,
-0.14143148064613342,
-0.0610215999186039,
-0.11016753315925598,
0.06722099334001541,
0.022168032824993134,
-0.08481651544570923,
0.06605587899684906,
-0.008005164563655853,
-0.01882365345954895,
0.05432295426726341,
0.024154748767614365,
0.04174061864614487,
0.048626724630594254,
0.02147703804075718,
-0.05503256991505623,
0.008252750150859356,
-0.035455696284770966,
-0.056309133768081665,
-0.0019569469150155783,
-0.017485151067376137,
0.005116594489663839,
0.013762857764959335,
0.033764731138944626,
0.029733112081885338,
-0.17315427958965302,
-0.018773898482322693,
0.05562301352620125,
0.002713000401854515,
0.004989310167729855,
-0.02381271682679653,
-0.05216526985168457,
0.015066421590745449,
0.0557468980550766,
0.0008447187137790024,
0.012113742530345917,
-0.06074847653508186,
-0.07377982884645462,
-0.020576095208525658,
-0.021750357002019882,
0.07194565236568451,
-0.07959567755460739,
-0.04313763231039047,
0.061235275119543076,
0.09256386011838913,
-0.03888016194105148,
0.1253652125597,
0.05961594358086586,
0.05984418839216232,
0.04011869430541992,
-0.0005460020620375872,
-0.08016332983970642,
-0.1410173624753952,
-0.05492774769663811,
-0.04134552925825119,
-0.07695852965116501,
0.06303997337818146,
0.007024332415312529,
-0.02987295761704445,
-0.014518754556775093,
0.015400983393192291,
-0.07756184041500092,
0.05042440816760063,
-0.008689159527420998,
0.047263406217098236,
0.05102413520216942,
0.006536020431667566,
0.009987222962081432,
0.0008167813648469746,
0.014748153276741505,
-0.008471241220831871,
0.016362495720386505,
-0.07213523238897324,
0.05543655529618263,
-0.045814767479896545,
0.03974130377173424,
-0.05300116911530495,
-0.013491218909621239,
0.04744276776909828,
-0.0068520437926054,
0.04911012575030327,
-0.13026544451713562,
-0.012450198642909527,
0.04228328913450241,
-0.0353471040725708,
0.06829246133565903,
0.09796847403049469,
-0.07309755682945251,
-0.11586534231901169,
0.017094163224101067,
0.03509926795959473,
-0.019609225913882256,
0.04210386797785759,
0.0028613305184990168,
-0.08624287694692612,
0.04348016530275345,
-0.0033277305774390697,
-0.005221397615969181,
-0.0011730652768164873,
-0.03515992313623428,
0.01587934046983719,
0.010161863639950752,
0.006115121301263571,
0.05016187205910683,
-0.02915632165968418,
0.13216565549373627,
0.023747563362121582,
-0.02696675807237625,
-0.024350594729185104,
-0.042794663459062576,
0.010313251055777073,
0.0436105839908123,
-0.005811607930809259,
0.004454478155821562,
-0.007034614682197571,
0.02200143225491047,
-0.00748045789077878,
-0.05321277305483818,
1.4460015444741605e-34,
0.0284773837774992,
0.06237397715449333,
0.0589817576110363,
0.13036227226257324,
-0.002726169303059578,
-0.042379606515169144,
-0.026456303894519806,
0.03510778024792671,
-0.010068978182971478,
-0.02170587331056595,
0.02099531888961792,
0.053957827389240265,
-0.022828267887234688,
-0.03276276960968971,
0.039404258131980896,
0.022303998470306396,
-0.05762274190783501,
0.043672818690538406,
-0.03614060953259468,
-0.06289048492908478,
0.07273370772600174,
0.016864679753780365,
0.029402950778603554,
0.05560643598437309,
0.11794959008693695,
0.0020037645008414984,
0.03036361001431942,
-0.09900803118944168,
-0.0012925202026963234,
-0.009539946913719177,
-0.08984512835741043,
-0.061933599412441254,
-0.010727623477578163,
0.01661393977701664,
0.06363484263420105,
0.010672315955162048,
-0.05283099785447121,
-0.016999615356326103,
-0.02159375138580799,
-0.0012933152029290795,
-0.14440181851387024,
-0.045454807579517365,
-0.02884337492287159,
-0.028364645317196846,
-0.07848650217056274,
-0.03764692693948746,
-0.06014993041753769,
0.04287618771195412,
-0.045473869889974594,
0.0004401849291753024,
-0.030135180801153183,
0.08891784399747849,
-0.02207544445991516,
0.054627884179353714,
-0.06714510172605515,
-0.03679342940449715,
0.07952085137367249,
-0.023044584318995476,
-0.0042809355072677135,
0.025860225781798363,
0.0681430771946907,
0.0690702497959137,
-0.03497384488582611,
-0.0016912136925384402,
-0.0006504910998046398,
-0.0027795187197625637,
-0.054847411811351776,
0.042995233088731766,
0.01124914176762104,
-0.011968287639319897,
-0.016020575538277626,
0.0001312562671955675,
-0.018792061135172844,
-0.005159446038305759,
-0.01842207834124565,
0.06653010100126266,
0.058786869049072266,
-0.039982255548238754,
0.08110370486974716,
0.050043463706970215,
0.07022020220756531,
-0.027073616161942482,
-0.024177484214305878,
-0.061650510877370834,
0.033287663012742996,
-0.08769090473651886,
-0.017674976959824562,
-0.0957549437880516,
-0.10904153436422348,
0.013283955864608288,
-0.05764607712626457,
0.054145827889442444,
-0.019494932144880295,
0.10602646321058273,
-0.024739347398281097,
-4.637800462686542e-34,
0.020823216065764427,
0.008252774365246296,
-0.03790952265262604,
0.12068041414022446,
0.012881290167570114,
0.00865963939577341,
0.04153154790401459,
0.057835061103105545,
0.0174066461622715,
-0.03285578638315201,
-0.009419219568371773,
-0.02881651557981968,
0.04709059000015259,
0.0006714447517879307,
0.07400091737508774,
-0.08486268669366837,
0.017448531463742256,
-0.022018227726221085,
0.016994010657072067,
0.10683498531579971,
0.021641789004206657,
0.06455616652965546,
-0.05909625068306923,
0.037424683570861816,
-0.11732729524374008,
0.006974813062697649,
-0.009598911739885807,
0.04029972851276398,
0.026361048221588135,
-0.042097024619579315,
-0.05782993137836456,
-0.010022052563726902,
-0.006071804091334343,
-0.014316470362246037,
-0.000010845310498552863,
0.04171259328722954,
0.020248079672455788,
-0.007065278012305498,
-0.0008533147047273815,
0.01419583149254322,
0.08162551373243332,
-0.00927435327321291,
-0.03104506991803646,
0.01253320463001728,
-0.030464503914117813,
-0.030653707683086395,
-0.08322935551404953,
0.10433531552553177,
-0.06130259111523628,
-0.036312151700258255,
0.022586312144994736,
0.055636320263147354,
0.024518972262740135,
0.02729598432779312,
-0.0515594556927681,
-0.07816709578037262,
-0.011437210254371166,
-0.039877261966466904,
0.08036584407091141,
0.050993964076042175,
-0.0542181171476841,
0.03181060776114464,
0.029718168079853058,
-0.08141623437404633,
0.019882507622241974,
-0.0633857399225235,
-0.004550463054329157,
-0.019197380170226097,
0.042098067700862885,
0.01022652443498373,
0.057613879442214966,
-0.034903962165117264,
0.013237083330750465,
-0.018011286854743958,
-0.0012923125177621841,
0.020178696140646935,
-0.012012172490358353,
-0.12883587181568146,
-0.04563816264271736,
-0.03877728432416916,
0.017507940530776978,
0.005066885147243738,
0.07003040611743927,
0.0005847134161740541,
0.0676584392786026,
0.10659176111221313,
0.021357456222176552,
0.07256080210208893,
0.023042088374495506,
0.07043585181236267,
-0.011710356920957565,
0.025339998304843903,
-0.01907683163881302,
0.06723053008317947,
-0.008577991276979446,
-3.40860601966142e-8,
-0.1126539334654808,
0.012564519420266151,
0.04997586831450462,
0.0299880001693964,
-0.00940142385661602,
-0.031857628375291824,
0.05144738778471947,
0.0022572162561118603,
-0.020688459277153015,
-0.018918201327323914,
0.07088690996170044,
0.013224280439317226,
0.02259652502834797,
0.05025205761194229,
-0.014692277647554874,
0.05896040424704552,
-0.057541776448488235,
0.0846080556511879,
-0.032203223556280136,
-0.041234903037548065,
0.09146015346050262,
-0.01769128628075123,
0.022035809233784676,
0.010000525042414665,
0.003920566290616989,
-0.04701673611998558,
0.00823126919567585,
0.07162168622016907,
-0.017208240926265717,
-0.0346982441842556,
-0.0370030552148819,
0.027193205431103706,
-0.029103076085448265,
-0.01502124685794115,
-0.024614812806248665,
0.030528755858540535,
0.053662996739149094,
-0.011692178435623646,
0.05330679938197136,
-0.010956959798932076,
-0.016636010259389877,
0.041186049580574036,
-0.12976856529712677,
0.05220881476998329,
-0.054900169372558594,
-0.0038303984329104424,
-0.024539876729249954,
-0.05619240924715996,
0.0598796084523201,
-0.026930104941129684,
-0.00998486578464508,
-0.04951617121696472,
0.03484341502189636,
0.06486282497644424,
0.057636845856904984,
0.0920497477054596,
-0.01935860514640808,
0.02743411809206009,
0.017009403556585312,
0.08709809184074402,
0.00027224249788559973,
0.0071145957335829735,
-0.022966958582401276,
0.05223885178565979
] | 0 | 0 | 0 | 0 |
|
Mind the GAP: Glimpse-based Active Perception improves generalization and sample efficiency of visual reasoning
|
https://openreview.net/forum?id=iXCeQ2m6vT
|
[
"Oleh Kolner",
"Thomas Ortner",
"Stanisław Woźniak",
"Angeliki Pantazi"
] |
Poster
|
Human capabilities in understanding visual relations are far superior to those of AI systems, especially for previously unseen objects. For example, while AI systems struggle to determine whether two such objects are visually the same or different, humans can do so with ease. Active vision theories postulate that the learning of visual relations is grounded in actions that we take to fixate objects and their parts by moving our eyes. In particular, the low-dimensional spatial information about the corresponding eye movements is hypothesized to facilitate the representation of relations between different image parts. Inspired by these theories, we develop a system equipped with a novel Glimpse-based Active Perception (GAP) that sequentially glimpses at the most salient regions of the input image and processes them at high resolution. Importantly, our system leverages the locations stemming from the glimpsing actions, along with the visual content around them, to represent relations between different parts of the image. The results suggest that the GAP is essential for extracting visual relations that go beyond the immediate visual content. Our approach reaches state-of-the-art performance on several visual reasoning tasks being more sample-efficient, and generalizing better to out-of-distribution visual inputs than prior models.
|
visual reasoning, active vision, out-of-distribution, generalization, sample efficiency, relational features, brain-inspired, neuro-inspired
| null | 11,544 |
2409.20213
|
[
-0.010102944448590279,
-0.0827041044831276,
0.06646204739809036,
0.036255430430173874,
0.12844420969486237,
0.007107734680175781,
0.021220145747065544,
-0.013708371669054031,
0.09676366299390793,
-0.051558755338191986,
-0.039563484489917755,
-0.012485985644161701,
-0.0211569145321846,
0.0167559701949358,
0.07269221544265747,
-0.004961532074958086,
0.08090455085039139,
0.03234000504016876,
-0.09044314175844193,
-0.019231745973229408,
0.06174611300230026,
-0.08359495550394058,
-0.003383334493264556,
-0.029258668422698975,
0.01215208787471056,
0.031197603791952133,
0.042856525629758835,
-0.03984225168824196,
-0.0058787609450519085,
-0.04667593911290169,
0.03985116258263588,
-0.009419475682079792,
-0.028296275064349174,
-0.0018699020147323608,
-0.03997744992375374,
0.1179114505648613,
-0.019443564116954803,
0.08138018101453781,
-0.033279985189437866,
-0.08337520062923431,
-0.0030370221938937902,
0.0247624684125185,
0.020827479660511017,
0.05928485840559006,
0.061036836355924606,
0.11204297095537186,
0.048116255551576614,
0.017352109774947166,
-0.007285740226507187,
-0.058167893439531326,
-0.18912217020988464,
-0.028341123834252357,
-0.02291540801525116,
0.003925634548068047,
0.02370903268456459,
0.10259265452623367,
0.0015545779606327415,
-0.036344852298498154,
0.01494027953594923,
0.00439791101962328,
0.03323481231927872,
-0.04410870373249054,
0.033498626202344894,
-0.07705933600664139,
-0.038037169724702835,
-0.04087899252772331,
0.06933093816041946,
-0.0643186867237091,
-0.012508283369243145,
0.032680582255125046,
-0.004963184706866741,
0.11163219809532166,
-0.05994739383459091,
0.042130377143621445,
-0.0358562134206295,
0.05391485616564751,
0.02449653297662735,
-0.07955478131771088,
0.024690626189112663,
-0.07107440382242203,
0.04106609895825386,
0.030944060534238815,
-0.024086374789476395,
0.05554172396659851,
0.07236599177122116,
0.006257082335650921,
-0.0471089743077755,
-0.021670738235116005,
-0.026195276528596878,
-0.00604996969923377,
-0.05309008061885834,
-0.13144047558307648,
0.0274882260710001,
-0.0189498420804739,
0.08512575179338455,
-0.02653505653142929,
0.02925366722047329,
-0.03745732083916664,
0.0006942864274606109,
0.04239693284034729,
0.02469930239021778,
0.048462193459272385,
-0.002036020392552018,
-0.026318248361349106,
0.03714824840426445,
0.02961667627096176,
0.025293394923210144,
0.014050116762518883,
0.04286543279886246,
-0.003318686503916979,
0.03413234278559685,
-0.033903975039720535,
-0.028083937242627144,
-0.0038662520237267017,
-0.02988438308238983,
0.0017590208444744349,
-0.025297177955508232,
-0.039089303463697433,
0.0742882788181305,
0.000983335659839213,
0.06798365712165833,
0.005563822574913502,
-0.022936860099434853,
0.0475514680147171,
0.08498793840408325,
-0.06778435409069061,
-0.029385199770331383,
1.2157124206677335e-33,
0.03479636833071709,
-0.022677980363368988,
0.09408455342054367,
0.01732446998357773,
0.05347033217549324,
-0.04750217869877815,
-0.04432631656527519,
0.009783831425011158,
0.011025791056454182,
0.10599646717309952,
-0.035086553543806076,
-0.006099401973187923,
-0.0047066546976566315,
0.08027634024620056,
0.000021744986952398904,
-0.07783597707748413,
-0.02082904800772667,
0.03518553078174591,
-0.08888747543096542,
0.005031480919569731,
-0.11482032388448715,
-0.007971812970936298,
-0.0034033115953207016,
0.02722485549747944,
-0.03270852565765381,
0.0037087388336658478,
0.01588936522603035,
-0.0689496248960495,
0.017510581761598587,
-0.01975635439157486,
-0.045689038932323456,
0.06318497657775879,
-0.03841685876250267,
0.029165145009756088,
-0.002017476363107562,
-0.09130344539880753,
-0.006142046302556992,
-0.07781463861465454,
0.07305008918046951,
0.009413189254701138,
-0.038934994488954544,
0.11520229279994965,
0.017458181828260422,
-0.012708015739917755,
-0.05624759569764137,
0.02106521464884281,
-0.025072386488318443,
0.11444048583507538,
-0.06675370782613754,
0.009671368636190891,
0.08326928317546844,
0.013117754831910133,
0.011403616517782211,
-0.09706372022628784,
-0.06335115432739258,
-0.0531960092484951,
0.009215052239596844,
0.03457432985305786,
0.02621973305940628,
-0.020651832222938538,
-0.035369399935007095,
0.006059891544282436,
0.01003488153219223,
0.088045634329319,
-0.0014798054471611977,
0.06704498082399368,
-0.035520534962415695,
0.0652247741818428,
0.053236525505781174,
0.04570065066218376,
-0.005598217248916626,
0.03634723275899887,
0.006132980342954397,
-0.0841134563088417,
0.01818690076470375,
0.04606270045042038,
0.04770399630069733,
-0.06704770028591156,
0.06926322728395462,
-0.0026775975711643696,
-0.009109235368669033,
-0.07125379145145416,
-0.029974833130836487,
-0.0030976044945418835,
-0.05364881455898285,
-0.019444534555077553,
0.02616824023425579,
-0.05728719383478165,
-0.06124413013458252,
-0.011721258983016014,
-0.037542399019002914,
-0.035338275134563446,
-0.04426802694797516,
-0.042078807950019836,
-0.03118397668004036,
-9.175604008671532e-34,
0.08110411465167999,
-0.01893855445086956,
0.03522435203194618,
-0.055731236934661865,
-0.029496371746063232,
-0.023132001981139183,
0.034445006400346756,
-0.009051963686943054,
-0.051581501960754395,
-0.03417176008224487,
-0.0409410186111927,
-0.03620831295847893,
-0.034366581588983536,
0.0081257876008749,
-0.05746285617351532,
-0.02571483701467514,
0.0625850111246109,
0.006036626640707254,
-0.020758187398314476,
0.08325257152318954,
0.02312852069735527,
0.03458830341696739,
-0.0522150844335556,
-0.05788527429103851,
-0.02051796205341816,
0.08324988931417465,
0.0463191382586956,
0.04636140167713165,
-0.08741974085569382,
-0.024181179702281952,
-0.005731523036956787,
-0.038704462349414825,
-0.03159436583518982,
-0.07084815204143524,
-0.05224815383553505,
0.06426239758729935,
-0.019896531477570534,
-0.08372116833925247,
-0.03908124566078186,
0.03304086625576019,
0.03584914654493332,
-0.08302967250347137,
0.011493297293782234,
0.021468788385391235,
-0.06134376302361488,
0.015415456146001816,
-0.034664783626794815,
0.0639108270406723,
-0.10206269472837448,
0.09740737080574036,
-0.04862383008003235,
0.0270923413336277,
-0.07020217180252075,
-0.03212061524391174,
-0.05699148401618004,
0.018678834661841393,
-0.059587039053440094,
0.028662962839007378,
0.03376467153429985,
-0.010769298300147057,
0.02363339252769947,
0.023592613637447357,
-0.06894081830978394,
-0.014931155368685722,
0.03607650473713875,
0.039193909615278244,
-0.020923428237438202,
-0.027148418128490448,
0.047605425119400024,
-0.054066251963377,
0.046654779464006424,
-0.05690806359052658,
-0.007987652905285358,
0.010282931849360466,
0.08299464732408524,
0.05174841359257698,
0.04054811969399452,
0.014026816003024578,
0.021156620234251022,
0.001824862090870738,
-0.04049651324748993,
-0.0967731773853302,
0.0035231399815529585,
0.08391838520765305,
0.03340097516775131,
0.1261957585811615,
-0.004417844116687775,
0.018955593928694725,
0.08898875117301941,
-0.015572267584502697,
-0.11661581695079803,
0.013637538067996502,
0.002852401463314891,
0.056144773960113525,
-0.04438173398375511,
-3.9587881417446624e-8,
-0.08133391290903091,
0.04638303443789482,
0.030111119151115417,
-0.024638207629323006,
0.042082592844963074,
-0.019981632009148598,
-0.029908696189522743,
0.15571662783622742,
-0.09270180016756058,
0.0016935436287894845,
0.07251308858394623,
-0.029597699642181396,
0.0028667503502219915,
-0.048702288419008255,
0.08440159261226654,
-0.01503883395344019,
0.049778468906879425,
-0.0025733429938554764,
-0.050060439854860306,
0.03894205018877983,
0.04979706555604935,
0.004022050648927689,
-0.01318424567580223,
0.004090491216629744,
0.014362625777721405,
-0.08519668877124786,
-0.07329057902097702,
0.04944417253136635,
0.017488570883870125,
-0.027283497154712677,
0.07304416596889496,
0.050756096839904785,
0.07014112919569016,
0.0019887296948581934,
0.07021130621433258,
0.050435882061719894,
-0.024416755884885788,
-0.0006789377657696605,
0.010558721609413624,
-0.0797697901725769,
-0.013097263872623444,
0.024355841800570488,
0.04126596450805664,
0.10890905559062958,
-0.011187262833118439,
0.00878835842013359,
0.03681672737002373,
-0.12423171103000641,
-0.0016780277946963906,
-0.010746808722615242,
-0.048984598368406296,
-0.03461216390132904,
0.009001475758850574,
0.11123854666948318,
-0.02092736028134823,
-0.025431983172893524,
0.027254847809672356,
-0.0015704588731750846,
-0.005105136893689632,
0.043326735496520996,
0.051231298595666885,
0.056281160563230515,
-0.09877043217420578,
0.027173563838005066
] |
https://github.com/ibm/glimpse-based-active-perception
| 1 | 0 | 0 | 0 |
MMTEB: Massive Multilingual Text Embedding Benchmark
|
https://openreview.net/forum?id=zl3pfz4VCV
|
[
"Kenneth Enevoldsen",
"Isaac Chung",
"Imene Kerboua",
"Márton Kardos",
"Ashwin Mathur",
"David Stap",
"Jay Gala",
"Wissam Siblini",
"Dominik Krzemiński",
"Genta Indra Winata",
"Saba Sturua",
"Saiteja Utpala",
"Mathieu Ciancone",
"Marion Schaeffer",
"Diganta Misra",
"Shreeya Dhakal",
"Jonathan Rystrøm",
"Roman Solomatin",
"Ömer Veysel Çağatan",
"Akash Kundu",
"et al. (62 additional authors not shown)"
] |
Poster
|
Text embeddings are typically evaluated on a narrow set of tasks, limited in terms of languages, domains, and task types. To circumvent this limitation and to provide a more comprehensive evaluation, we introduce the Massive Multilingual Text Embedding Benchmark (MMTEB) -- a large-scale community-driven initiative expanding MTEB to over 500 quality-controlled evaluation tasks across 1,000+ languages. MMTEB includes a wide range of challenging novel tasks such as instruction following, long-document retrieval, and code retrieval, and represents the largest multilingual collection of evaluation tasks for embedding models to date. We use this collection to construct multiple highly multilingual benchmarks. We evaluate a representative set of models on these benchmarks.
Our findings indicate that, while LLM-based models can achieve state-of-the-art performance on a subset of languages, the best-performing publicly available model across languages is the notably smaller, multilingual-e5-large-instruct.
Massive benchmarks often impose high computational demands, limiting accessibility, particularly for low-resource communities. To address this, we downsample tasks based on inter-task correlation (i.e., selecting only a diverse set of tasks) while preserving relative rankings.
We further optimize tasks such as retrieval by sampling hard negatives, creating smaller but effective splits. These optimizations allow us to introduce benchmarks at a significantly lower computational cost. For instance, we introduce a new zero-shot English benchmark that maintains a similar ordering at a fraction of the cost.
|
natural language processing, benchmark, sentence embeddings, multilingual
|
We introduce the Massive Multilingual Text Embedding Benchmark (MMTEB) including 500+ tasks across 1,000+ languages, greatly expanding multilingual evaluation for embeddings.
| 11,538 |
2502.13595
|
[
-0.0660136416554451,
-0.07898189127445221,
0.021893782541155815,
0.026054805144667625,
0.055734384804964066,
-0.009942895732820034,
-0.07696248590946198,
0.032285530120134354,
0.01842627301812172,
-0.04684888571500778,
-0.055183734744787216,
-0.0632425844669342,
0.0519302561879158,
0.04979827627539635,
-0.028218213468790054,
-0.0004536157939583063,
0.11388827115297318,
0.05379410833120346,
-0.054737359285354614,
-0.05621466040611267,
0.03750143200159073,
0.05388198420405388,
0.06010916829109192,
-0.062108591198921204,
0.08770976215600967,
-0.011805790476500988,
-0.06631515175104141,
-0.0369257889688015,
-0.0031758672557771206,
-0.04083405062556267,
0.04621412977576256,
0.02920270338654518,
-0.044898614287376404,
0.066195547580719,
-0.004710776265710592,
0.0864429771900177,
-0.10297568142414093,
-0.03998270258307457,
-0.008456249721348286,
-0.009655439294874668,
-0.004316949285566807,
0.037838585674762726,
0.01854139007627964,
0.03571000322699547,
0.09138647466897964,
0.00550396554172039,
-0.01726406067609787,
-0.05161110684275627,
-0.048187896609306335,
0.020461339503526688,
-0.07496629655361176,
-0.03689880669116974,
-0.014449525624513626,
-0.0031635540071874857,
-0.01965903490781784,
-0.05295254662632942,
0.04735477641224861,
-0.020635979250073433,
-0.001723195891827345,
-0.08531775325536728,
-0.027926480397582054,
-0.09267888218164444,
0.05896526947617531,
0.009471047669649124,
0.0036641524638980627,
-0.042966362088918686,
0.026544610038399696,
0.0530971959233284,
-0.03967060148715973,
0.07107776403427124,
-0.04246010258793831,
0.03599295765161514,
-0.05275387689471245,
0.09347935020923615,
-0.009016687050461769,
0.06125914677977562,
-0.016618626192212105,
-0.03883206099271774,
0.026868566870689392,
-0.09289373457431793,
0.039052966982126236,
0.006716657429933548,
0.03704354539513588,
-0.018592946231365204,
0.1269867867231369,
-0.035180870443582535,
0.09113144129514694,
0.005531653296202421,
0.014474891126155853,
0.026934457942843437,
-0.007174172904342413,
-0.016108492389321327,
0.08802162110805511,
0.035722922533750534,
0.010159512981772423,
0.004265960771590471,
0.07931534200906754,
0.003241891274228692,
-0.05174757167696953,
0.08868096023797989,
0.010244852863252163,
0.08861465752124786,
0.04800117388367653,
-0.053892120718955994,
-0.053546544164419174,
-0.004674124531447887,
0.08784687519073486,
0.05744689702987671,
-0.03805241733789444,
-0.1030297726392746,
-0.0029752706177532673,
0.05549652874469757,
0.006661094259470701,
-0.01337342243641615,
0.06071148440241814,
-0.03912388160824776,
0.09026253968477249,
-0.03381507843732834,
0.06742268800735474,
0.05462375283241272,
-0.05360220745205879,
0.07541295886039734,
0.0227385051548481,
-0.07620591670274734,
-0.037535831332206726,
0.04151008278131485,
-0.012510375119745731,
4.4741211395291925e-34,
0.0701153576374054,
0.026576485484838486,
-0.04032992199063301,
0.028285255655646324,
0.02757658250629902,
-0.06425963342189789,
0.009206762537360191,
0.06747370958328247,
-0.1455346792936325,
-0.04437927156686783,
-0.0007531683659180999,
0.07815450429916382,
-0.00033844495192170143,
0.08822907507419586,
0.0004070025752298534,
-0.01890120841562748,
-0.001020226045511663,
0.02436658926308155,
0.011342757381498814,
0.028729364275932312,
0.031347230076789856,
-0.021086810156702995,
0.013603956438601017,
-0.021354926750063896,
0.0005853953189216554,
0.020435933023691177,
0.11654170602560043,
-0.06103100627660751,
-0.05067349970340729,
-0.001175848301500082,
-0.07040097564458847,
-0.08193375915288925,
0.01663116365671158,
0.0014515909133479,
0.03730202838778496,
-0.05133412778377533,
-0.013147220946848392,
-0.0071412366814911366,
0.004890547599643469,
-0.04317565634846687,
-0.04430205002427101,
-0.0202037300914526,
-0.04682295769453049,
-0.014108859933912754,
-0.044189199805259705,
-0.0820799395442009,
-0.04424303397536278,
-0.007181523833423853,
-0.01796412654221058,
-0.07427345961332321,
0.07690701633691788,
-0.011661773547530174,
0.0003618118353188038,
0.05635395273566246,
0.04143761470913887,
0.01122219581156969,
0.027921929955482483,
0.03827029466629028,
0.049404241144657135,
0.023503372445702553,
-0.026316717267036438,
0.02747560478746891,
0.006916017737239599,
0.06128467246890068,
0.052624307572841644,
0.024645164608955383,
0.023568594828248024,
0.08051541447639465,
0.0491400882601738,
-0.027806531637907028,
-0.004138495307415724,
0.0011197557905688882,
0.04253361374139786,
-0.046867869794368744,
-0.0033468434121459723,
0.03172652795910835,
-0.012566695921123028,
-0.15159304440021515,
0.014067317359149456,
-0.003258798271417618,
0.02494317851960659,
-0.07931371033191681,
0.036507897078990936,
-0.1066015437245369,
-0.018452875316143036,
-0.10291237384080887,
0.01779862679541111,
-0.09035572409629822,
0.0658474713563919,
0.027432793751358986,
0.012261293828487396,
-0.024832047522068024,
-0.038336850702762604,
-0.06565266102552414,
0.07550586014986038,
1.2336078527910643e-33,
0.026418879628181458,
-0.0660584345459938,
-0.02466408535838127,
0.06411222368478775,
-0.015858586877584457,
-0.03158607706427574,
-0.022510146722197533,
0.013434117659926414,
-0.06060492619872093,
-0.025927100330591202,
-0.05531222000718117,
-0.07097432762384415,
0.056897860020399094,
-0.00535542843863368,
0.041801486164331436,
0.004831656347960234,
0.019937407225370407,
0.018961556255817413,
0.0067622605711221695,
0.08386633545160294,
-0.033385809510946274,
0.058089479804039,
-0.025051487609744072,
0.07583165168762207,
0.02199581265449524,
0.11625680327415466,
-0.06347917765378952,
-0.03677469864487648,
-0.09133502095937729,
-0.10289135575294495,
-0.022303210571408272,
-0.0016271908534690738,
-0.03536023944616318,
-0.04746371507644653,
-0.055002931505441666,
0.055642783641815186,
0.04674191027879715,
-0.0023039127700030804,
-0.041182637214660645,
0.046089354902505875,
0.09294306486845016,
-0.0012819658732041717,
-0.002219465794041753,
-0.01231352798640728,
0.010493789799511433,
0.028089508414268494,
-0.12050963193178177,
0.020313810557127,
0.027695517987012863,
0.0016106576658785343,
-0.04071532189846039,
-0.03391479700803757,
-0.029679277911782265,
0.028455765917897224,
0.05894546955823898,
-0.13066129386425018,
-0.07680437713861465,
-0.08438155055046082,
-0.02274113893508911,
0.023792652413249016,
-0.014407754875719547,
-0.02640225738286972,
0.039642252027988434,
-0.020049648359417915,
0.0206014271825552,
-0.0736808180809021,
-0.03135979548096657,
0.007988478057086468,
-0.054492220282554626,
0.009094629436731339,
0.03870810940861702,
-0.0656275823712349,
-0.06510073691606522,
-0.0009730914607644081,
-0.023433877155184746,
0.044307779520750046,
0.0490119494497776,
0.04180433601140976,
0.010609653778374195,
-0.05609459429979324,
0.06474416702985764,
-0.01307786162942648,
0.013860250823199749,
-0.0558866411447525,
0.05614502355456352,
0.03825852647423744,
-0.0190441831946373,
0.13051648437976837,
0.009462585672736168,
0.028208572417497635,
-0.002786500845104456,
0.02320695109665394,
-0.0003315335779916495,
0.030920658260583878,
0.055433403700590134,
-3.4342679811061316e-8,
-0.04841975122690201,
0.04329342767596245,
-0.05287744104862213,
0.020720424130558968,
-0.06648659706115723,
-0.01616651564836502,
-0.03342605009675026,
0.08987421542406082,
0.02457490749657154,
-0.002690300578251481,
0.06549911946058273,
-0.0357537642121315,
-0.030036548152565956,
-0.019427508115768433,
0.03431033715605736,
0.0746651217341423,
-0.009758012369275093,
0.011497685685753822,
0.03962124511599541,
0.01308688148856163,
0.03596430644392967,
0.0866730660200119,
0.049685437232255936,
-0.0017821502406150103,
-0.013819050043821335,
-0.021197091788053513,
-0.018941499292850494,
0.08284308761358261,
0.09006894379854202,
-0.08256281912326813,
0.01505796704441309,
0.03819761797785759,
-0.07896890491247177,
-0.08833339810371399,
0.003300632117316127,
0.08476419001817703,
-0.0009934429544955492,
0.015013080090284348,
0.04550408571958542,
0.027044640854001045,
0.08690862357616425,
0.06143417954444885,
-0.042205460369586945,
0.020633000880479813,
0.037821706384420395,
-0.012447177432477474,
-0.0761091336607933,
-0.01953175477683544,
0.015320587903261185,
-0.08469963073730469,
-0.05886894091963768,
-0.12385649234056473,
-0.03440739959478378,
0.07991235703229904,
0.007910724729299545,
0.0880139023065567,
-0.08560960739850998,
0.0090343551710248,
0.08283894509077072,
0.04531589895486832,
0.007722226902842522,
-0.05171584710478783,
-0.00011491774785099551,
0.028357619419693947
] |
https://github.com/embeddings-benchmark/mteb
| 2,446 | 0 | 0 | 0 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.