title
stringlengths 14
154
| paper_url
stringlengths 42
42
| authors
sequencelengths 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
sequencelengths 384
384
|
|---|---|---|---|---|---|---|---|---|---|
Training-Free Dataset Pruning for Instance Segmentation | https://openreview.net/forum?id=rvxWEbTtRY | [
"Yalun Dai",
"Lingao Xiao",
"Ivor Tsang",
"Yang He"
] | Poster | Existing dataset pruning techniques primarily focus on classification tasks, limiting their applicability to more complex and practical tasks like instance segmentation. Instance segmentation presents three key challenges: pixel-level annotations, instance area variations, and class imbalances, which significantly complicate dataset pruning efforts. Directly adapting existing classification-based pruning methods proves ineffective due to their reliance on time-consuming model training process. To address this, we propose a novel **T**raining-**F**ree **D**ataset **P**runing (**TFDP**) method for instance segmentation. Specifically, we leverage shape and class information from image annotations to design a Shape Complexity Score (SCS), refining it into a Scale-Invariant (SI-SCS) and Class-Balanced (CB-SCS) versions to address instance area variations and class imbalances, all without requiring model training. We achieve state-of-the-art results on VOC 2012, Cityscapes, and MS COCO datasets, generalizing well across CNN and Transformer architectures. Remarkably, our approach accelerates the pruning process by an average of **1349$\times$** on COCO compared to the adapted baselines. | dataset pruning, instance segmentation | A fast and training-free dataset pruning method for instance segmentation | 989 | 2503.00828 | [
-0.022932911291718483,
-0.0007791068637743592,
0.0533859021961689,
-0.0012564886128529906,
0.05047968775033951,
-0.029885049909353256,
-0.04404296353459358,
-0.023422040045261383,
-0.020356496796011925,
-0.09444746375083923,
-0.022885337471961975,
-0.11064337939023972,
-0.023155249655246735,
0.07028578966856003,
0.005522412247955799,
-0.05225217342376709,
0.07256541401147842,
0.07961337268352509,
-0.10019794851541519,
0.018086697906255722,
-0.005674497224390507,
0.037634845823049545,
-0.02960367687046528,
0.08793100714683533,
-0.016548553481698036,
-0.011632101610302925,
0.02507670782506466,
-0.07933104038238525,
-0.014537233859300613,
-0.07158466428518295,
-0.013295755721628666,
0.011379322037100792,
0.04468882828950882,
0.06061752513051033,
0.014624632894992828,
0.041419290006160736,
0.013775241561233997,
0.009468636475503445,
-0.0355018675327301,
-0.061187099665403366,
0.007454055827111006,
-0.04462679475545883,
0.04777116701006889,
0.018388288095593452,
0.0767497643828392,
0.0677165687084198,
0.02010807767510414,
-0.08077200502157211,
0.0532318651676178,
-0.04506361111998558,
0.015425987541675568,
0.00571795180439949,
-0.08680276572704315,
0.07308348268270493,
-0.06799579411745071,
-0.015198104083538055,
0.11246338486671448,
0.042928338050842285,
0.005948624107986689,
0.027786236256361008,
-0.019551822915673256,
-0.08708547800779343,
0.03076283633708954,
0.013217009603977203,
-0.06576405465602875,
-0.07816905528306961,
0.07018275558948517,
-0.027782294899225235,
0.0379410982131958,
-0.03458365425467491,
0.022033875808119774,
0.07695559412240982,
0.0236747395247221,
0.10076821595430374,
0.04559632018208504,
0.05536352097988129,
0.00245760683901608,
0.04084545746445656,
0.03779449686408043,
-0.05607737600803375,
-0.025387216359376907,
-0.0187411867082119,
0.01225300133228302,
0.06761002540588379,
0.10442826896905899,
-0.0530746728181839,
-0.035901181399822235,
0.0238780714571476,
0.01688765548169613,
0.033441632986068726,
-0.07977184653282166,
0.012985210865736008,
0.01959756389260292,
0.04075741395354271,
0.012276006862521172,
-0.021771250292658806,
-0.07106579840183258,
-0.05467955395579338,
-0.08108656108379364,
0.06289918720722198,
-0.044598571956157684,
0.037115685641765594,
0.03920365869998932,
-0.016832705587148666,
0.05177675560116768,
0.03923257440328598,
0.03734578937292099,
-0.023514650762081146,
0.08110157400369644,
-0.06283191591501236,
0.08324407786130905,
0.03924070671200752,
-0.006046304013580084,
-0.021683569997549057,
0.04060251638293266,
-0.03480659797787666,
-0.07288740575313568,
-0.06919541209936142,
-0.018180038779973984,
0.07824914902448654,
-0.05051400512456894,
0.009991352446377277,
-0.0426332987844944,
0.018358144909143448,
0.048131395131349564,
-0.001259807264432311,
-0.07732769846916199,
1.7091441741826357e-33,
0.02232005074620247,
0.031002530828118324,
0.03480895608663559,
-0.05787003040313721,
0.06872551143169403,
-0.011214576661586761,
-0.06294377148151398,
0.0661144107580185,
-0.01846816949546337,
-0.009703832678496838,
-0.06376471370458603,
-0.012026064097881317,
-0.0300982017070055,
0.004577545449137688,
0.09089336544275284,
-0.05947703495621681,
-0.013707958161830902,
0.05841962993144989,
-0.06862900406122208,
-0.029317248612642288,
-0.043119121342897415,
-0.020223135128617287,
-0.0285419300198555,
-0.03217463940382004,
0.018106333911418915,
0.0036261628847569227,
0.04536544531583786,
-0.02607368677854538,
-0.009063415229320526,
0.030322501435875893,
-0.027106797322630882,
0.041014812886714935,
0.010330489836633205,
0.06302282959222794,
-0.014452630653977394,
-0.07736600190401077,
-0.001799708348698914,
0.0038564177230000496,
0.05252541974186897,
0.024447429925203323,
0.008128947578370571,
0.02358265221118927,
0.02750120498239994,
-0.0604865588247776,
-0.01137546356767416,
-0.01345971878618002,
0.06402447074651718,
0.048819251358509064,
-0.011173597536981106,
0.04355015233159065,
0.08775848895311356,
-0.05292538180947304,
-0.016360757872462273,
-0.040568526834249496,
-0.06231570988893509,
-0.026026634499430656,
0.027589162811636925,
0.02641918510198593,
0.04705517366528511,
0.01522766798734665,
-0.04144852235913277,
0.05785340815782547,
-0.06970861554145813,
0.11288852244615555,
-0.11151983588933945,
-0.0349898599088192,
0.017455991357564926,
0.03260805830359459,
-0.05752065032720566,
0.015734871849417686,
-0.03571121767163277,
0.07081735879182816,
-0.030044233426451683,
-0.021517395973205566,
0.07757402211427689,
-0.058150891214609146,
0.06891551613807678,
-0.01384235080331564,
-0.06025140359997749,
0.05632762238383293,
-0.07349087297916412,
0.0297074057161808,
-0.006045165937393904,
-0.062379591166973114,
-0.025562288239598274,
0.018969973549246788,
0.11941789835691452,
-0.060362253338098526,
0.041998859494924545,
0.024991722777485847,
-0.07382906228303909,
-0.09921370446681976,
-0.06523352861404419,
0.026024168357253075,
-0.003337934846058488,
-1.3702081942092837e-33,
-0.013140830211341381,
0.02547895908355713,
-0.09155817329883575,
0.017503567039966583,
-0.014044989831745625,
-0.030074182897806168,
-0.022667378187179565,
0.08231973648071289,
-0.08390801399946213,
-0.12029305100440979,
-0.01688639633357525,
0.03184792771935463,
0.013356516137719154,
-0.05909736827015877,
-0.04934997856616974,
-0.00950640719383955,
-0.08378918468952179,
-0.040685877203941345,
-0.020301172509789467,
0.059137579053640366,
0.022102320566773415,
0.04556381329894066,
-0.055246587842702866,
-0.022601818665862083,
-0.12242445349693298,
0.06407208740711212,
-0.08544372767210007,
0.02071533165872097,
0.06078692525625229,
-0.01502882968634367,
-0.061633843928575516,
-0.015241794288158417,
-0.034838415682315826,
-0.07051428407430649,
-0.07521428167819977,
0.00151350861415267,
0.05863797292113304,
-0.02101011760532856,
0.048723142594099045,
0.06562136113643646,
0.04074586555361748,
-0.04778209701180458,
-0.07124454528093338,
0.03501403331756592,
-0.06320860981941223,
0.039462003856897354,
0.045998405665159225,
0.016896827146410942,
-0.03429315239191055,
0.03720331937074661,
-0.08724386245012283,
0.027611199766397476,
-0.1038077175617218,
0.05125676095485687,
0.00713172135874629,
-0.03989596292376518,
-0.03430251404643059,
0.027128558605909348,
-0.015169480815529823,
-0.0077671813778579235,
-0.02833726815879345,
-0.0390075147151947,
-0.05158016085624695,
0.003006435465067625,
-0.01072576642036438,
-0.018763896077871323,
-0.018755678087472916,
0.0003998104075435549,
0.03944787383079529,
0.10382969677448273,
0.041316840797662735,
0.04763380065560341,
-0.0463465116918087,
-0.05722834914922714,
-0.12841399013996124,
0.02885441854596138,
0.028218790888786316,
-0.04338948801159859,
0.045524291694164276,
-0.0416409857571125,
0.018396534025669098,
-0.01934477873146534,
0.028566179797053337,
0.01847139559686184,
0.06537719815969467,
0.10738267749547958,
0.05323437601327896,
0.0071396660059690475,
0.05014527961611748,
0.10902110487222672,
-0.12057773023843765,
0.06119084730744362,
0.07060331106185913,
0.07646551728248596,
0.0016650496982038021,
-4.161932665169843e-8,
-0.0074508036486804485,
0.07196037471294403,
0.01588902622461319,
0.03219912201166153,
0.04591191187500954,
-0.01878693886101246,
-0.06817910820245743,
0.19021418690681458,
0.00833000335842371,
0.07296307384967804,
0.01659252494573593,
-0.03176217898726463,
-0.04272260516881943,
-0.018371380865573883,
0.012141061015427113,
0.08031654357910156,
0.03442820906639099,
-0.014111115597188473,
-0.011073055677115917,
0.022017670795321465,
-0.009664461947977543,
-0.009643875062465668,
0.004937277175486088,
-0.008307251147925854,
-0.021432975307106972,
-0.09741225093603134,
-0.01082460768520832,
0.08216828852891922,
0.04098450019955635,
-0.033733662217855453,
-0.03915323689579964,
0.008531108498573303,
-0.01146598532795906,
0.03689175099134445,
0.0293335672467947,
0.021188469603657722,
-0.05198170989751816,
0.0729510635137558,
-0.025934286415576935,
-0.04148920252919197,
0.07129716873168945,
-0.0020278652664273977,
0.0052784462459385395,
0.013556932099163532,
0.020560339093208313,
0.02638702280819416,
0.07025492191314697,
-0.032757412642240524,
0.03532211855053902,
0.011003444902598858,
0.016662338748574257,
0.026375843212008476,
-0.048432644456624985,
0.06644608080387115,
-0.024688243865966797,
0.07509426772594452,
0.024607032537460327,
-0.07570497691631317,
0.044260550290346146,
0.09877458214759827,
-0.005518786609172821,
-0.043770965188741684,
-0.10011042654514313,
-0.03260055556893349
] |
Towards Understanding Text Hallucination of Diffusion Models via Local Generation Bias | https://openreview.net/forum?id=SKW10XJlAI | [
"Rui Lu",
"Runzhe Wang",
"Kaifeng Lyu",
"Xitai Jiang",
"Gao Huang",
"Mengdi Wang"
] | Poster | Score-based diffusion models have achieved incredible performance in generating realistic images, audio, and video data. While these models produce high-quality samples with impressive details, they often introduce unrealistic artifacts, such as distorted fingers or hallucinated texts with no meaning. This paper focuses on textual hallucinations, where diffusion models correctly generate individual symbols but assemble them in a nonsensical manner. Through experimental probing, we consistently observe that such phenomenon is attributed it to the network's local generation bias. Denoising networks tend to produce outputs that rely heavily on highly correlated local regions, particularly when different dimensions of the data distribution are nearly pairwise independent. This behavior leads to a generation process that decomposes the global distribution into separate, independent distributions for each symbol, ultimately failing to capture the global structure, including underlying grammar. Intriguingly, this bias persists across various denoising network architectures including MLP and transformers which have the structure to model global dependency. These findings also provide insights into understanding other types of hallucinations, extending beyond text, as a result of implicit biases in the denoising models. Additionally, we theoretically analyze the training dynamics for a specific case involving a two-layer MLP learning parity points on a hypercube, offering an explanation of its underlying mechanism. | Diffusion model, Deep learning theory, generative model, Hallucination | Experimentally and theoretically investigate the mechanism of text artifacts for diffusion models, finding it is related to neural network's bias of sparse dependency for output in input. | 987 | 2503.03595 | [
0.0208215843886137,
-0.10858577489852905,
0.03953292965888977,
0.007435305509716272,
0.06312774866819382,
-0.029952196404337883,
-0.09129349887371063,
-0.11721674352884293,
0.12621206045150757,
-0.09861353784799576,
-0.030170265585184097,
0.048748474568128586,
0.0456460565328598,
0.01232256181538105,
-0.03892522305250168,
-0.030922044068574905,
0.014212138019502163,
0.018435025587677956,
-0.09504161030054092,
-0.08339312672615051,
0.04448447749018669,
0.034093212336301804,
-0.02789326384663582,
-0.021358273923397064,
0.10961952805519104,
0.0631626769900322,
-0.030539387837052345,
-0.040079522877931595,
0.024070341140031815,
-0.0472104474902153,
0.06727728247642517,
0.08282391726970673,
-0.04351864382624626,
-0.004495091270655394,
-0.016451910138130188,
0.04685479775071144,
-0.04275231063365936,
0.05219651386141777,
0.07110848277807236,
-0.010028397664427757,
0.0660334900021553,
0.07915569096803665,
0.05611115321516991,
0.030243275687098503,
0.030280470848083496,
-0.014150941744446754,
-0.044027749449014664,
0.028920041397213936,
-0.06830509752035141,
-0.00906965509057045,
0.005748806521296501,
0.02575775608420372,
0.009575074538588524,
0.05546550080180168,
-0.03475796431303024,
0.042306430637836456,
0.07258222252130508,
0.08770492672920227,
0.005824029911309481,
-0.0058222562074661255,
-0.0579807311296463,
-0.03268631547689438,
-0.02743145450949669,
0.01503897923976183,
0.024185754358768463,
0.03651364892721176,
0.056367289274930954,
0.06583675742149353,
-0.05441000685095787,
0.05008387938141823,
0.029541699215769768,
0.005361431743949652,
-0.0979449450969696,
0.0843009427189827,
0.03336635231971741,
0.025449424982070923,
-0.013675806112587452,
-0.11070549488067627,
0.03275126591324806,
-0.09345043450593948,
0.08294195681810379,
0.022969475015997887,
0.12663011252880096,
-0.01655174233019352,
0.038212355226278305,
-0.048023756593465805,
0.025251619517803192,
-0.020053623244166374,
-0.052687887102365494,
0.014728606678545475,
-0.11008166521787643,
-0.08497501909732819,
-0.0349414125084877,
-0.003414723090827465,
0.015076879411935806,
0.055131081491708755,
-0.015271736308932304,
-0.004308189265429974,
0.011156857945024967,
0.09126531332731247,
0.0004723801393993199,
-0.04081067815423012,
0.0043497649021446705,
-0.03046719916164875,
0.044869594275951385,
-0.061774611473083496,
0.08973927050828934,
0.03595893830060959,
0.013438720256090164,
0.001311587868258357,
-0.016337811946868896,
0.05269942432641983,
-0.05577465519309044,
-0.010347370058298111,
0.0402677059173584,
-0.06167063117027283,
0.049578603357076645,
-0.027515031397342682,
0.007796410005539656,
0.06155472248792648,
-0.01913405768573284,
-0.021857453510165215,
-0.0533030666410923,
-0.010888349264860153,
-0.006151137407869101,
-0.04058253392577171,
-0.07003813982009888,
8.554192695433753e-34,
0.00900366809219122,
-0.025912653654813766,
-0.0210929773747921,
0.03347263112664223,
0.12076633423566818,
-0.026895815506577492,
-0.07531391829252243,
-0.055355314165353775,
0.056055549532175064,
-0.059643249958753586,
-0.056196119636297226,
0.03825755789875984,
-0.05236009135842323,
0.03387238085269928,
0.010082462802529335,
0.016708645969629288,
-0.04477299004793167,
0.016449350863695145,
-0.03422844782471657,
0.028879867866635323,
-0.017027344554662704,
0.028088727965950966,
-0.006993802264332771,
-0.020682442933321,
-0.08277340233325958,
0.025848977267742157,
0.024320922791957855,
-0.07192176580429077,
-0.021714111790060997,
0.0006042615859769285,
-0.11600441485643387,
0.03784652799367905,
0.024065330624580383,
0.006875712890177965,
0.02965603582561016,
0.019946275278925896,
0.028154641389846802,
-0.017583047971129417,
0.0026470962911844254,
-0.016295436769723892,
0.00749063678085804,
0.03545480594038963,
-0.044516559690237045,
-0.0430443212389946,
-0.06812798231840134,
0.044707294553518295,
0.003906833473592997,
-0.02767864242196083,
-0.07053864002227783,
-0.03283529728651047,
0.03358784690499306,
0.035260386765003204,
0.04655676335096359,
0.01328850444406271,
-0.015364599414169788,
0.0213798638433218,
0.08214158564805984,
-0.09901562333106995,
0.08148034662008286,
-0.02065884880721569,
0.09722309559583664,
0.05475204437971115,
-0.004460427910089493,
0.06781813502311707,
0.0742090493440628,
0.004190237261354923,
-0.017122674733400345,
0.0015623687068000436,
0.03551553562283516,
-0.07745031267404556,
-0.026707066223025322,
0.01664179377257824,
-0.051076244562864304,
-0.018606726080179214,
-0.0054201846942305565,
0.010664042085409164,
-0.0069428193382918835,
0.018486397340893745,
-0.018679331988096237,
0.038920238614082336,
-0.043954141438007355,
-0.08134379237890244,
0.01660432666540146,
-0.12745332717895508,
-0.04640334099531174,
0.014940669760107994,
-0.017646005377173424,
-0.11980576813220978,
-0.05162614583969116,
-0.07654022425413132,
-0.021428130567073822,
-0.011046629399061203,
0.00994404498487711,
-0.043695010244846344,
0.013871471397578716,
-2.6013015911155423e-33,
-0.07697643339633942,
0.04675491899251938,
0.039151184260845184,
0.039716001600027084,
-0.06400769203901291,
0.008885127492249012,
0.002934419084340334,
0.03625546395778656,
0.050168950110673904,
-0.03490275889635086,
-0.04002182185649872,
-0.028826938942074776,
0.0034760478883981705,
-0.05013201758265495,
0.01975763775408268,
-0.12628571689128876,
0.09161423146724701,
0.024577658623456955,
-0.04567967727780342,
0.016250599175691605,
0.03677184879779816,
0.07794459164142609,
-0.1370270550251007,
0.005992378573864698,
0.030204538255929947,
0.06305808573961258,
-0.0257836002856493,
0.06274023652076721,
0.04326062276959419,
-0.028858792036771774,
-0.0218192208558321,
0.08431629091501236,
-0.013740246184170246,
-0.043310366570949554,
-0.0747978538274765,
0.07629021257162094,
0.0022809719666838646,
0.011580436490476131,
-0.08731163293123245,
-0.000873328885063529,
0.03172704577445984,
0.009973269887268543,
-0.07384596765041351,
0.04967696964740753,
0.01691593788564205,
-0.07930530607700348,
-0.08774611353874207,
0.0025913866702467203,
-0.005208334419876337,
0.05990348383784294,
0.05645629018545151,
0.06756320595741272,
-0.056576747447252274,
0.026633452624082565,
-0.1338309943675995,
-0.05797731131315231,
-0.11705842614173889,
0.027376364916563034,
0.11490027606487274,
0.039527103304862976,
-0.04569573327898979,
-0.002156273229047656,
-0.05611561983823776,
-0.09383291751146317,
-0.08267275243997574,
-0.05810373276472092,
-0.02146041952073574,
0.034426528960466385,
0.056729722768068314,
0.07402300089597702,
0.009753335267305374,
0.003478693775832653,
-0.02264358103275299,
0.04695745185017586,
-0.026173051446676254,
-0.0337555930018425,
-0.01777138002216816,
-0.0221025962382555,
-0.030510181561112404,
-0.0564073771238327,
0.009045572020113468,
0.01106115523725748,
-0.03542909026145935,
0.08016345649957657,
0.11463053524494171,
-0.028080953285098076,
-0.042515676468610764,
0.010056977160274982,
-0.031695373356342316,
0.036619044840335846,
0.02483450435101986,
-0.020196029916405678,
-0.059690818190574646,
0.049753494560718536,
-0.07744775712490082,
-4.4150013422949996e-8,
-0.06992774456739426,
0.02615305222570896,
-0.08457407355308533,
0.05086074769496918,
0.036928240209817886,
-0.07652914524078369,
0.0408177524805069,
0.02490212954580784,
0.020040392875671387,
-0.07945724576711655,
0.04778232425451279,
-0.05000356212258339,
-0.06300653517246246,
-0.0919189378619194,
-0.008100259117782116,
0.09855126589536667,
0.014700797386467457,
0.061886075884103775,
-0.012980378232896328,
-0.008829738013446331,
0.06723546236753464,
0.05396297946572304,
-0.00299454340711236,
-0.021597614511847496,
0.02633550949394703,
0.02160060405731201,
-0.022692400962114334,
0.01652735471725464,
0.009045304730534554,
-0.009585004299879074,
0.05490916222333908,
0.05162360891699791,
0.001289804233238101,
-0.04052416607737541,
-0.009353505447506905,
0.11999987810850143,
0.049308568239212036,
-0.004246200900524855,
-0.014578709378838539,
-0.017750050872564316,
0.01676221936941147,
-0.02990495041012764,
0.0194198377430439,
0.01359968539327383,
-0.016730351373553276,
-0.019432619214057922,
0.04849044978618622,
-0.00890542846173048,
0.06005590781569481,
-0.014091682620346546,
0.025951506569981575,
0.007230781950056553,
-0.03431662917137146,
0.018978223204612732,
0.07487531006336212,
-0.03564021736383438,
0.07674438506364822,
0.0628441646695137,
0.00657676113769412,
0.05699247121810913,
0.05052800849080086,
0.015426335856318474,
0.012287506833672523,
-0.005866043735295534
] |
VideoShield: Regulating Diffusion-based Video Generation Models via Watermarking | https://openreview.net/forum?id=uzz3qAYy0D | [
"Runyi Hu",
"Jie Zhang",
"Yiming Li",
"Jiwei Li",
"Qing Guo",
"Han Qiu",
"Tianwei Zhang"
] | Poster | Artificial Intelligence Generated Content (AIGC) has advanced significantly, particularly with the development of video generation models such as text-to-video (T2V) models and image-to-video (I2V) models. However, like other AIGC types, video generation requires robust content control. A common approach is to embed watermarks, but most research has focused on images, with limited attention given to videos. Traditional methods, which embed watermarks frame-by-frame in a post-processing manner, often degrade video quality. In this paper, we propose VideoShield, a novel watermarking framework specifically designed for popular diffusion-based video generation models. Unlike post-processing methods, VideoShield embeds watermarks directly during video generation, eliminating the need for additional training. To ensure video integrity, we introduce a tamper localization feature that can detect changes both temporally (across frames) and spatially (within individual frames). Our method maps watermark bits to template bits, which are then used to generate watermarked noise during the denoising process. Using DDIM Inversion, we can reverse the video to its original watermarked noise, enabling straightforward watermark extraction. Additionally, template bits allow precise detection for potential spatial and temporal modification. Extensive experiments across various video models (both T2V and I2V models) demonstrate that our method effectively extracts watermarks and detects tamper without compromising video quality. Furthermore, we show that this approach is applicable to image generation models, enabling tamper detection in generated images as well. Codes and models are available at https://github.com/hurunyi/VideoShield. | video, watermarking, tamper localization | null | 986 | 2501.14195 | [
-0.02086896076798439,
-0.08454553782939911,
0.058975424617528915,
-0.0038172323256731033,
0.15163981914520264,
-0.04938889667391777,
-0.028945088386535645,
-0.1016332283616066,
0.04587505757808685,
-0.08083174377679825,
-0.04524322599172592,
-0.002025086898356676,
-0.0030497456900775433,
0.05675314739346504,
-0.06273924559354782,
-0.0348348505795002,
0.044730234891176224,
0.07374496012926102,
-0.09588385373353958,
-0.09685058146715164,
0.011169036850333214,
-0.10758133232593536,
0.014456948265433311,
-0.014824526384472847,
0.033093322068452835,
0.09779886901378632,
-0.0004652906209230423,
0.0328185074031353,
0.06119763106107712,
-0.03228731080889702,
0.08853507786989212,
-0.09036614745855331,
0.013050197623670101,
0.0059227789752185345,
-0.03168972581624985,
0.011677076108753681,
0.0025299922563135624,
0.007705843076109886,
-0.007179055362939835,
-0.06206753849983215,
0.032843250781297684,
0.03436565026640892,
0.022745011374354362,
0.008011741563677788,
0.02473665028810501,
0.040683481842279434,
0.037417102605104446,
0.022160829976201057,
-0.014462164603173733,
0.03243424743413925,
-0.0034644759725779295,
0.016557294875383377,
-0.05435646325349808,
-0.021885911002755165,
-0.0637887641787529,
-0.03692658245563507,
0.08625657856464386,
0.010538540780544281,
-0.00036617484875023365,
-0.007011889014393091,
0.006094371899962425,
-0.011385995894670486,
0.038604263216257095,
0.0016689208569005132,
-0.005756595637649298,
-0.019373290240764618,
0.06750475615262985,
0.04093090072274208,
-0.01810288242995739,
0.03184407949447632,
-0.08676877617835999,
0.06274605542421341,
0.008688997477293015,
0.05917348340153694,
-0.06407099217176437,
0.03494768962264061,
-0.010981569066643715,
-0.025775345042347908,
0.025859439745545387,
-0.08701671659946442,
0.10474994033575058,
-0.10464887320995331,
0.10783546417951584,
0.026266541332006454,
0.0563419871032238,
-0.02386443316936493,
0.036257728934288025,
0.0063633122481405735,
0.028661062940955162,
0.018888838589191437,
-0.11145099252462387,
0.0533965602517128,
0.06558756530284882,
0.01835549809038639,
0.052732378244400024,
0.017452046275138855,
-0.01562984101474285,
-0.011792502366006374,
0.00929524376988411,
0.05158478766679764,
-0.056664809584617615,
-0.039525099098682404,
-0.03550300747156143,
-0.030512453988194466,
0.14984019100666046,
0.01687147282063961,
0.05383412539958954,
0.06553915143013,
0.04905730113387108,
0.029283437877893448,
0.014135381206870079,
0.07233952730894089,
-0.010489231906831264,
-0.020068710669875145,
0.07427027821540833,
-0.0037111297715455294,
-0.020016612485051155,
0.034645549952983856,
-0.013609512709081173,
-0.010084982961416245,
0.04683404415845871,
-0.023757515475153923,
-0.099050372838974,
-0.023067200556397438,
0.004561562091112137,
-0.05390682443976402,
0.050919290632009506,
2.924805350643513e-34,
0.015099886804819107,
-0.042853277176618576,
-0.02099197544157505,
0.01134383026510477,
0.0671783834695816,
0.03783848509192467,
-0.0007249198388308287,
-0.0346081368625164,
-0.020867513492703438,
-0.09847259521484375,
0.03264718875288963,
-0.09890499711036682,
-0.043374743312597275,
0.07193373888731003,
0.063425213098526,
-0.06221795827150345,
-0.046682775020599365,
-0.058097101747989655,
-0.0024882873985916376,
0.0512770377099514,
0.034598179161548615,
-0.028832418844103813,
-0.02793928235769272,
-0.039857104420661926,
-0.01079635601490736,
-0.01444731280207634,
0.11373088508844376,
-0.05321443825960159,
0.06652146577835083,
0.011770639568567276,
-0.1099763810634613,
-0.06408099085092545,
0.08124129474163055,
0.003620811039581895,
0.017308171838521957,
0.03698981553316116,
-0.034872595220804214,
-0.03655347600579262,
-0.006819979753345251,
0.014630904421210289,
0.011871240101754665,
0.035047005861997604,
-0.1258310228586197,
-0.035365551710128784,
-0.12578430771827698,
0.023169104009866714,
0.029390525072813034,
0.060019489377737045,
-0.07050352543592453,
-0.03877229988574982,
0.04401852935552597,
0.016577882692217827,
0.003963594790548086,
-0.13729353249073029,
-0.057991404086351395,
-0.027224920690059662,
0.06907522678375244,
-0.09147527068853378,
0.04037521779537201,
-0.06309929490089417,
0.015932921320199966,
0.02802285924553871,
-0.029239386320114136,
0.04825395718216896,
0.006188694387674332,
-0.06355368345975876,
0.09607827663421631,
0.05502883344888687,
-0.009975455701351166,
-0.009058256633579731,
-0.04475537687540054,
0.003687582677230239,
-0.003953246865421534,
-0.010641475208103657,
-0.0544302836060524,
0.008290069177746773,
0.01931130141019821,
0.027378752827644348,
0.015289997681975365,
-0.03199315816164017,
-0.04610130935907364,
-0.08072862774133682,
0.06707073003053665,
-0.06404843926429749,
-0.0039543756283819675,
-0.06356225907802582,
-0.048121627420186996,
-0.05365316942334175,
-0.06188890337944031,
-0.0438925065100193,
0.00968370120972395,
-0.052250612527132034,
0.10452153533697128,
-0.020169753581285477,
0.0020262030884623528,
-4.298719933642237e-34,
-0.021273113787174225,
0.07505286484956741,
-0.02295340597629547,
0.09089528024196625,
-0.00008517397509422153,
-0.014063640497624874,
0.03013792634010315,
0.04746881499886513,
0.09270502626895905,
-0.061532702296972275,
-0.08302148431539536,
-0.03541726991534233,
-0.07165763527154922,
-0.012826777063310146,
-0.05722929909825325,
-0.06063997000455856,
0.03662111610174179,
0.0034369544591754675,
-0.07055798172950745,
-0.025576911866664886,
0.017767002806067467,
-0.02883123978972435,
0.03405112400650978,
0.05234384164214134,
0.01680009253323078,
0.06536214798688889,
0.028924543410539627,
0.03649745136499405,
0.024655839428305626,
-0.031408969312906265,
0.00880051963031292,
0.06619212031364441,
0.07635483890771866,
-0.08038379997015,
-0.08632365614175797,
0.03979819267988205,
0.05669460445642471,
0.002796681597828865,
-0.0430421456694603,
0.11709600687026978,
0.025104476138949394,
-0.01596415601670742,
-0.06478572636842728,
0.037656061351299286,
-0.013633772730827332,
0.009270068258047104,
-0.03879866376519203,
0.05603735148906708,
-0.0022626249119639397,
-0.028488442301750183,
0.00324096973054111,
0.02393808588385582,
-0.06554633378982544,
0.019902052357792854,
-0.04104069620370865,
-0.01594826579093933,
-0.10072092711925507,
0.04217452183365822,
0.09813321381807327,
0.053355056792497635,
0.018900273367762566,
-0.048638395965099335,
-0.01833326742053032,
-0.06897096335887909,
-0.02162092924118042,
0.007176471874117851,
-0.02318730391561985,
0.038753755390644073,
-0.05872062221169472,
0.019654467701911926,
0.048726242035627365,
-0.04417358338832855,
0.0438624732196331,
0.03585879132151604,
0.04859605059027672,
-0.08113852143287659,
-0.018753470852971077,
0.038097646087408066,
-0.003518042154610157,
-0.04329139366745949,
0.007539939135313034,
-0.019605355337262154,
0.02420029602944851,
-0.007213775999844074,
0.07523347437381744,
0.007323514204472303,
-0.015850557014346123,
-0.011747105978429317,
-0.016282860189676285,
0.08859258145093918,
-0.04813183471560478,
0.04373393580317497,
-0.045741114765405655,
0.028522154316306114,
-0.0023334291763603687,
-3.6966802952065336e-8,
-0.05431557074189186,
-0.00757433520630002,
0.0015062157763168216,
-0.0027175461873412132,
-0.0861053466796875,
0.018792612478137016,
0.05401751026511192,
0.04222489148378372,
0.10745710134506226,
-0.15099503099918365,
0.03260365128517151,
-0.010634029284119606,
-0.050063177943229675,
-0.07589199393987656,
-0.06543012708425522,
0.06692014634609222,
0.03841812163591385,
0.024900168180465698,
-0.02604968659579754,
0.03886883333325386,
-0.07685628533363342,
0.005918299313634634,
0.006351361516863108,
-0.035802438855171204,
0.026932179927825928,
0.03991563618183136,
-0.007498184684664011,
0.01525625679641962,
0.11430226266384125,
0.006905968300998211,
0.025189420208334923,
0.028908884152770042,
0.07793611288070679,
-0.02473476715385914,
-0.00278409942984581,
0.0610501728951931,
-0.007815161719918251,
-0.018492283299565315,
-0.035015229135751724,
-0.0036679788026958704,
0.005215856246650219,
0.024492669850587845,
0.042729128152132034,
-0.01797010935842991,
-0.004255443345755339,
0.03615683317184448,
0.07560764253139496,
-0.05038432404398918,
0.021627023816108704,
-0.016954997554421425,
0.02654249407351017,
-0.054558124393224716,
-0.04685789719223976,
0.06794769316911697,
0.058261774480342865,
-0.03745783120393753,
0.0585608035326004,
0.010952724143862724,
0.07254710793495178,
0.03959118574857712,
0.0338040255010128,
0.018275735899806023,
0.130000039935112,
0.03460131958127022
] |
PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks | https://openreview.net/forum?id=T5QLRRHyL1 | [
"Matthew Chang",
"Gunjan Chhablani",
"Alexander Clegg",
"Mikael Dallaire Cote",
"Ruta Desai",
"Michal Hlavac",
"Vladimir Karashchuk",
"Jacob Krantz",
"Roozbeh Mottaghi",
"Priyam Parashar",
"Siddharth Patki",
"Ishita Prasad",
"Xavier Puig",
"Akshara Rai",
"Ram Ramrakhya",
"Daniel Tran",
"Joanne Truong",
"John M Turner",
"Eric Undersander",
"Tsung-Yen Yang"
] | Poster | We present a benchmark for Planning And Reasoning Tasks in humaN-Robot collaboration (PARTNR) designed to study human-robot coordination in household activities. PARTNR tasks exhibit characteristics of everyday tasks, such as spatial, temporal, and heterogeneous agent capability constraints. We employ a semi-automated task generation pipeline using Large Language Models (LLMs), incorporating simulation-in-the-loop for the grounding and verification. PARTNR stands as the largest benchmark of its kind, comprising 100,000 natural language tasks, spanning 60 houses and 5,819 unique objects. We analyze state-of-the-art LLMs on PARTNR tasks, across the axes of planning, perception and skill execution. The analysis reveals significant limitations in SoTA models, such as poor coordination and failures in task tracking and recovery from errors. When LLMs are paired with 'real' humans, they require 1.5x as many steps as two humans collaborating and 1.1x more steps than a single human, underscoring the potential for improvement in these models. We further show that fine-tuning smaller LLMs with planning data can achieve performance on par with models 9 times larger, while being 8.6x faster at inference. Overall, PARTNR highlights significant challenges facing collaborative embodied agents and aims to drive research in this direction. | Human-Robot Collaboration, Planning, Embodied AI | Introducing a novel benchmark and in-depth analysis of top-performing planners for human-robot collaboration in household tasks. | 984 | 2411.00081 | [
0.03629715368151665,
-0.08418146520853043,
0.06131433695554733,
0.010158002376556396,
0.03857484459877014,
-0.0492340624332428,
-0.02765592373907566,
-0.0482938289642334,
0.003888071281835437,
-0.010461094789206982,
-0.04698316007852554,
-0.05021722614765167,
0.0200808085501194,
0.060712046921253204,
0.12072333693504333,
-0.02232990600168705,
0.09922090172767639,
-0.009019991382956505,
-0.07297954708337784,
-0.00312427943572402,
-0.0021487472113221884,
0.03416348993778229,
0.0562775582075119,
-0.006744499318301678,
0.025125563144683838,
-0.00746364938095212,
0.04042210429906845,
-0.10850214958190918,
0.0028506754897534847,
-0.02815384790301323,
0.06450306624174118,
-0.015438485890626907,
-0.005290542263537645,
-0.009969649836421013,
-0.000880095933098346,
0.1319001168012619,
-0.022019991651177406,
-0.00376436160877347,
0.055943336337804794,
-0.054721806198358536,
-0.054957315325737,
-0.03906995430588722,
0.04837874323129654,
-0.013758635148406029,
0.05888110399246216,
0.06937246769666672,
-0.02738657034933567,
-0.06852924823760986,
-0.003190202871337533,
0.019407907500863075,
-0.06731756776571274,
0.0061600590124726295,
0.054857950657606125,
0.012423517182469368,
-0.051372844725847244,
0.03354984521865845,
0.043475132435560226,
-0.005988273303955793,
0.039728663861751556,
-0.060742221772670746,
-0.00977642834186554,
-0.09804371744394302,
0.02570023387670517,
-0.040221087634563446,
-0.038211844861507416,
-0.042222242802381516,
-0.002707740059122443,
-0.015583722852170467,
-0.021165205165743828,
0.02259303443133831,
-0.003590164938941598,
-0.004346438217908144,
0.018455037847161293,
0.03658933937549591,
-0.0227613914757967,
0.03433268889784813,
-0.01326494850218296,
-0.045641787350177765,
0.10518968105316162,
-0.08363068848848343,
-0.05540071800351143,
0.0016881618648767471,
-0.03580234572291374,
0.05833951756358147,
0.04824069142341614,
0.012956012040376663,
-0.01427577156573534,
0.1275186836719513,
0.02556070126593113,
-0.0012169464025646448,
-0.03090132214128971,
-0.04618816450238228,
0.040627315640449524,
0.007369671016931534,
-0.01641049236059189,
0.07337292283773422,
-0.02246616780757904,
-0.055039092898368835,
0.0013470835983753204,
0.05860463157296181,
-0.027118993923068047,
0.06756230443716049,
0.03332557529211044,
-0.032577354460954666,
-0.019455686211586,
0.018174748867750168,
0.05841570347547531,
-0.023188527673482895,
0.011683972552418709,
-0.05350257828831673,
0.007839652709662914,
-0.050837837159633636,
0.07607161998748779,
0.09021283686161041,
0.035664789378643036,
-0.03444751352071762,
0.029754793271422386,
0.015327264554798603,
0.03164757043123245,
0.00851038470864296,
0.012788503430783749,
-0.033593401312828064,
0.09124588966369629,
0.05447067320346832,
0.09740161150693893,
-0.01066049188375473,
0.04062868654727936,
-1.7771516192477418e-34,
0.0506216324865818,
0.053505875170230865,
0.04743165150284767,
0.028393754735589027,
0.08809218555688858,
-0.006308531854301691,
0.036088068038225174,
0.022197293117642403,
-0.07552061975002289,
0.037892669439315796,
-0.0007285999599844217,
-0.06553167849779129,
-0.05619575083255768,
0.0648607388138771,
0.025447934865951538,
0.006089840549975634,
-0.003995582461357117,
0.04150278493762016,
-0.08233797550201416,
0.009732984006404877,
0.06474330276250839,
-0.029830122366547585,
-0.01193977240473032,
-0.006560432258993387,
0.0364837683737278,
0.022179055958986282,
0.08193942904472351,
-0.01818745769560337,
0.00640156539157033,
-0.02940857596695423,
-0.0999545082449913,
0.03464650735259056,
-0.02396746724843979,
0.10409161448478699,
-0.008377001620829105,
-0.05663701146841049,
-0.035975612699985504,
-0.07105065882205963,
0.008781280368566513,
0.03279814496636391,
0.01586516946554184,
0.04274071007966995,
0.031309179961681366,
-0.04780394956469536,
-0.06568491458892822,
-0.05192982405424118,
-0.010949836112558842,
0.08248461037874222,
-0.03015953302383423,
-0.012866368517279625,
0.019326452165842056,
0.08126091957092285,
-0.013904699124395847,
-0.0467001311480999,
-0.0045286244712769985,
-0.055614352226257324,
0.022028611972928047,
0.01836051605641842,
0.01621503010392189,
0.049310389906167984,
0.05209815874695778,
0.010388054884970188,
-0.033571865409612656,
0.07643326371908188,
0.09825067967176437,
0.00754405464977026,
0.0062726764008402824,
0.12298107147216797,
0.11718972027301788,
-0.044691301882267,
-0.021695023402571678,
-0.020788339897990227,
0.04154619947075844,
-0.02892647124826908,
0.018797291442751884,
-0.042391400784254074,
0.07016132771968842,
-0.08510933816432953,
0.001154188415966928,
0.04773422330617905,
-0.015249662101268768,
0.0314805768430233,
-0.052227530628442764,
-0.030612841248512268,
0.016545841470360756,
-0.07423389703035355,
0.0073390379548072815,
-0.07460116595029831,
-0.03435762971639633,
0.04864104837179184,
-0.035855699330568314,
-0.048204030841588974,
-0.07058820128440857,
0.013604450039565563,
-0.05766180157661438,
1.2827465874048387e-35,
-0.07061217725276947,
-0.07006222009658813,
-0.008117727003991604,
0.07894621789455414,
0.015280194580554962,
-0.12764616310596466,
0.023236287757754326,
-0.07268853485584259,
-0.0005482176784425974,
-0.02103634737432003,
-0.09425418078899384,
-0.07535673677921295,
0.028919018805027008,
0.0028736458625644445,
0.06412288546562195,
-0.056590691208839417,
0.007672495674341917,
-0.043781671673059464,
0.06049967184662819,
0.028203262016177177,
0.02217085100710392,
0.09310020506381989,
-0.0489371158182621,
0.02869335375726223,
-0.0231469813734293,
0.06377855688333511,
-0.025815704837441444,
-0.029719572514295578,
-0.00745118223130703,
0.05414663627743721,
0.00547760771587491,
-0.060761820524930954,
-0.015307010151445866,
-0.035724539309740067,
-0.02881082333624363,
-0.026553628966212273,
-0.03480721265077591,
0.02481917291879654,
0.010343887843191624,
0.01606071926653385,
0.0408572256565094,
-0.09110304713249207,
-0.058837324380874634,
0.0010994786862283945,
-0.019915809854865074,
0.003627006197348237,
-0.12184584885835648,
-0.011015962809324265,
-0.09390052407979965,
0.04588013514876366,
-0.0038026266265660524,
0.03559752553701401,
-0.11457047611474991,
-0.0926380455493927,
-0.00911098625510931,
-0.0763976126909256,
0.021985581144690514,
-0.05228767171502113,
0.03480280935764313,
-0.0237561222165823,
-0.09398942440748215,
0.04913707077503204,
0.04458209127187729,
0.04768577590584755,
0.007780460640788078,
0.04274808242917061,
-0.03434602543711662,
-0.02294188179075718,
0.05133878067135811,
0.005746850278228521,
0.013907029293477535,
-0.07305575907230377,
0.006701842416077852,
0.04148251935839653,
0.06941045820713043,
0.006659400183707476,
-0.038505919277668,
-0.019262762740254402,
0.07726399600505829,
-0.0333111509680748,
-0.03428091108798981,
0.012062775902450085,
0.003983377479016781,
0.06969161331653595,
-0.023276258260011673,
0.024763694033026695,
-0.047144316136837006,
0.1308015137910843,
-0.02416808158159256,
0.0386749692261219,
-0.010217803530395031,
0.026553738862276077,
0.03966698795557022,
0.06313056498765945,
-0.09967678040266037,
-4.350404836372945e-8,
-0.05884592607617378,
0.07194974273443222,
0.007600703742355108,
-0.01978922449052334,
0.01936446875333786,
-0.04450051486492157,
-0.025568341836333275,
0.032725732773542404,
0.023359596729278564,
0.06049833074212074,
0.07360047847032547,
-0.03394365683197975,
-0.005894574336707592,
-0.015541508793830872,
0.04531848430633545,
-0.006411746609956026,
0.04613902047276497,
0.0014381359796971083,
-0.03515363484621048,
-0.048211414366960526,
0.05030972510576248,
-0.0002972696674987674,
-0.06732994318008423,
0.0067785861901938915,
0.028795737773180008,
-0.02072400227189064,
-0.08009371161460876,
0.15526017546653748,
0.05821415036916733,
0.012736373580992222,
0.003089721780270338,
-0.07493021339178085,
-0.005589949432760477,
0.05696077272295952,
0.10833849012851715,
-0.0036491146311163902,
-0.003960212226957083,
-0.04590384289622307,
0.05582641065120697,
-0.06954260170459747,
0.08113063871860504,
0.08352245390415192,
-0.010787838138639927,
0.026096796616911888,
0.0409226194024086,
0.00462421914562583,
-0.13072921335697174,
-0.1268978714942932,
-0.01258205994963646,
-0.07888904958963394,
-0.08032455295324326,
-0.032551586627960205,
-0.06482362002134323,
0.08524276316165924,
0.03498564660549164,
0.04294591769576073,
0.008918216452002525,
-0.08203911036252975,
0.0632067397236824,
0.04621201008558273,
0.042031947523355484,
0.07233204692602158,
-0.10759095847606659,
-0.027049470692873
] |
Causal Identification for Complex Functional Longitudinal Studies | https://openreview.net/forum?id=96beVMeHh9 | [
"Andrew Ying"
] | Poster | Real-time monitoring in modern medical research introduces functional longitudinal data, characterized by continuous-time measurements of outcomes, treatments, and confounders. This complexity leads to uncountably infinite treatment-confounder feedbacks, which traditional causal inference methodologies cannot handle. Inspired by the coarsened data framework, we adopt stochastic process theory, measure theory, and net convergence to propose a nonparametric causal identification framework. This framework generalizes classical g-computation, inverse probability weighting, and doubly robust formulas, accommodating time-varying outcomes subject to mortality and censoring for functional longitudinal data. We examine our framework through Monte Carlo simulations. Our approach addresses significant gaps in current methodologies, providing a solution for functional longitudinal data and paving the way for future estimation work in this domain. | Causal Inference, Stochastic Process, Longitudinal Data; Functional Data, Continuous Time. | We propose a novel causal identification framework for functional longitudinal data. | 981 | 2206.12525 | [
-0.06220683455467224,
-0.024759462103247643,
0.04471677541732788,
0.08314212411642075,
0.0613301619887352,
-0.015880316495895386,
-0.021622443571686745,
0.0042963470332324505,
0.02965497225522995,
-0.04731232672929764,
-0.054988179355859756,
0.037042561918497086,
0.025984495878219604,
-0.001966852694749832,
-0.09260503947734833,
-0.07667440921068192,
0.09731322526931763,
0.02004126086831093,
-0.011557959951460361,
-0.04192350432276726,
-0.020555032417178154,
0.04812641069293022,
0.042515311390161514,
0.040486861020326614,
0.0021470305509865284,
-0.08752360194921494,
-0.02104753628373146,
-0.048546478152275085,
-0.042786747217178345,
0.022985167801380157,
0.08632521331310272,
0.031428009271621704,
-0.038367196917533875,
-0.016417119652032852,
-0.037766240537166595,
0.0464341901242733,
0.009517518803477287,
0.00008427244756603613,
-0.07806132733821869,
-0.022429708391427994,
-0.024295233190059662,
-0.003880186239257455,
-0.004629470873624086,
0.06581547111272812,
0.0515851229429245,
-0.07306525111198425,
0.02532552368938923,
-0.10225801169872284,
-0.05377437174320221,
0.07220926135778427,
-0.04453159496188164,
-0.0022665157448500395,
0.006677820347249508,
0.028705524280667305,
0.040783144533634186,
0.03262035548686981,
0.04587721452116966,
-0.0733010321855545,
-0.014950642362236977,
0.046651288866996765,
-0.049288176000118256,
-0.08954432606697083,
-0.06425420939922333,
-0.012836096808314323,
0.011618384160101414,
0.06883224099874496,
0.02176583744585514,
-0.015610480681061745,
-0.013886895030736923,
0.06547652930021286,
-0.010992266237735748,
0.013229604810476303,
-0.09220428764820099,
0.04333709552884102,
0.003528760513290763,
-0.051252130419015884,
0.007295422721654177,
0.03898421674966812,
0.06767550855875015,
-0.08348006010055542,
0.031651683151721954,
0.06232393532991409,
0.022132620215415955,
0.024503512308001518,
0.028657276183366776,
-0.012178502976894379,
0.0004995734198018909,
0.08715800195932388,
-0.040879473090171814,
0.030944423750042915,
0.015116779133677483,
0.103934146463871,
-0.015608259476721287,
-0.00017824115639086813,
-0.033947866410017014,
-0.025382259860634804,
-0.007102488074451685,
-0.014105451293289661,
0.039922088384628296,
0.04310184344649315,
-0.02526112087070942,
-0.010107805952429771,
0.07258949428796768,
0.07827060669660568,
-0.03123895265161991,
-0.06236165389418602,
-0.012395542114973068,
0.04775068163871765,
-0.008863900788128376,
0.016724146902561188,
-0.029247870668768883,
-0.05874476954340935,
0.12728670239448547,
-0.024164428934454918,
-0.02431795559823513,
-0.048602864146232605,
0.05131090432405472,
0.026418279856443405,
0.0960429385304451,
0.03690514340996742,
0.06363248080015182,
-0.04590310528874397,
0.002174241701141,
-0.024577995762228966,
0.024059418588876724,
0.1030493900179863,
0.003764493390917778,
5.022581548861428e-34,
0.05363164097070694,
-0.05964180454611778,
0.07100130617618561,
0.009750626049935818,
-0.025446884334087372,
-0.06413936614990234,
-0.0561508871614933,
-0.11351775377988815,
0.0030159836169332266,
-0.0039071133360266685,
0.030705992132425308,
0.04672625660896301,
-0.11109001934528351,
-0.008930983021855354,
-0.00270715425722301,
0.0827731043100357,
-0.024478571489453316,
0.14641857147216797,
0.01681375503540039,
-0.02766312099993229,
0.02655523642897606,
-0.0836532711982727,
-0.04153372347354889,
-0.009424163028597832,
-0.009690642356872559,
-0.014700323343276978,
-0.02739134058356285,
-0.005850780755281448,
-0.06276006996631622,
0.011072319932281971,
-0.016510017216205597,
0.02809729240834713,
0.050740886479616165,
-0.020595666021108627,
-0.044369589537382126,
-0.006597241852432489,
-0.00567522132769227,
-0.050535887479782104,
-0.04920103773474693,
-0.035799529403448105,
0.030934354290366173,
-0.00503377104178071,
0.07783836126327515,
-0.016398701816797256,
-0.09136727452278137,
-0.1101427897810936,
0.0608951561152935,
-0.09577023983001709,
-0.005622190423309803,
0.01209197100251913,
0.07921572774648666,
-0.03319195285439491,
0.00486319325864315,
-0.08883218467235565,
-0.0738665908575058,
0.08261324465274811,
-0.07256043702363968,
0.02587982453405857,
-0.06944098323583603,
0.039655692875385284,
0.033894751220941544,
0.0032686861231923103,
-0.016064148396253586,
-0.022720005363225937,
0.015585338696837425,
0.08086035400629044,
-0.0495234839618206,
-0.018709609284996986,
0.0751987099647522,
0.05613935366272926,
-0.023465938866138458,
0.044191211462020874,
0.003150068921968341,
0.002409572247415781,
0.02075367234647274,
-0.05337408557534218,
-0.026913857087492943,
-0.09304396063089371,
-0.01668597012758255,
0.03144891932606697,
0.051314231008291245,
-0.08707468956708908,
-0.07280915230512619,
0.06681942939758301,
0.04622023180127144,
-0.040237460285425186,
0.03274395316839218,
-0.003093490144237876,
-0.11205191165208817,
-0.06919805705547333,
-0.06322647631168365,
-0.00032860744977369905,
0.06069820001721382,
0.06860645860433578,
-0.01118200272321701,
-2.2826805412472996e-33,
-0.09393229335546494,
-0.002820788649842143,
0.023221660405397415,
-0.033325593918561935,
0.09741687774658203,
-0.03473509103059769,
-0.030017133802175522,
-0.07577371597290039,
0.12140598148107529,
-0.0866532176733017,
0.07880005240440369,
-0.12174246460199356,
0.02773965522646904,
0.023038960993289948,
0.008991869166493416,
0.010127969086170197,
-0.007900395430624485,
-0.027990346774458885,
-0.022317327558994293,
-0.022392259910702705,
0.062930628657341,
-0.03400016576051712,
0.022173311561346054,
0.0013541950611397624,
-0.011251210235059261,
-0.03829602524638176,
0.008609488606452942,
0.0736549124121666,
0.003936402965337038,
-0.029621170833706856,
-0.05182724818587303,
0.012709416449069977,
-0.004623569548130035,
-0.029005834832787514,
-0.026001567021012306,
0.03421275317668915,
0.054816026240587234,
0.05322595313191414,
-0.00661427341401577,
-0.09883946180343628,
0.09172248095273972,
0.01516767032444477,
-0.09147384017705917,
-0.05276651307940483,
0.06596390157938004,
-0.016934791579842567,
0.047183893620967865,
0.031981077045202255,
0.05526265129446983,
0.02061062678694725,
-0.0036847367882728577,
0.05986815318465233,
-0.01525135338306427,
0.07968012988567352,
-0.044372547417879105,
-0.11825153231620789,
0.051968228071928024,
-0.1265755146741867,
-0.002137740608304739,
0.03042697347700596,
-0.049482330679893494,
-0.012697230093181133,
-0.004225130192935467,
0.061057694256305695,
0.012719085440039635,
0.026386458426713943,
0.10806300491094589,
-0.05855169519782066,
-0.026696840301156044,
0.010208684019744396,
0.03145590424537659,
0.03565625101327896,
-0.048109035938978195,
-0.01704256609082222,
0.036019958555698395,
-0.024821793660521507,
0.00961125548928976,
0.024615926668047905,
-0.10802493989467621,
0.02607716992497444,
0.11955526471138,
-0.02083161659538746,
0.05828686058521271,
-0.05339770019054413,
-0.024932045489549637,
0.01201531570404768,
-0.0006691973540000618,
0.011695490218698978,
0.061335619539022446,
-0.009076223708689213,
-0.05915100499987602,
-0.044287219643592834,
-0.03412366658449173,
-0.018317943438887596,
-0.14470770955085754,
-4.09399767420382e-8,
0.10946457087993622,
0.09260604530572891,
-0.010380294173955917,
0.01568993739783764,
0.0029512241017073393,
0.043111588805913925,
-0.02597782388329506,
0.018357539549469948,
-0.005670987535268068,
0.020679660141468048,
0.08810477703809738,
0.01938622072339058,
0.07467194646596909,
-0.09212648123502731,
-0.02161601185798645,
0.039651043713092804,
0.011165016330778599,
0.02641918696463108,
-0.007937046699225903,
0.013535648584365845,
-0.01327645592391491,
0.03745272010564804,
-0.012235874310135841,
-0.03510101139545441,
-0.0004924723180010915,
0.014597130939364433,
-0.014673576690256596,
0.022602738812565804,
0.05548353120684624,
-0.0003300842363387346,
0.003715265542268753,
0.05356346443295479,
-0.01610124669969082,
0.03536895290017128,
-0.0214830469340086,
-0.06385646760463715,
0.08413362503051758,
0.01621435210108757,
-0.10564073920249939,
0.03190218284726143,
-0.011200349777936935,
0.09000220149755478,
-0.03991119563579559,
0.006336989812552929,
0.05269347131252289,
0.005992587190121412,
-0.011780557222664356,
-0.012634324841201305,
0.024888774380087852,
0.0877939760684967,
0.11660153418779373,
0.006157440599054098,
0.0409400500357151,
0.021389804780483246,
0.041856054216623306,
0.06426109373569489,
0.06477498263120651,
0.019792282953858376,
-0.02780061773955822,
-0.02438315749168396,
0.058256376534700394,
-0.01813124120235443,
-0.029536953195929527,
-0.0363163948059082
] |
Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation | https://openreview.net/forum?id=2RfWRKwxYh | [
"Sheng-Feng Yu",
"Jia-Jiun Yao",
"Wei-Chen Chiu"
] | Poster | Although larger datasets are crucial for training large deep models, the rapid growth of dataset size has brought a significant challenge in terms of considerable training costs, which even results in prohibitive computational expenses. Dataset Distillation becomes a popular technique recently to reduce the dataset size via learning a highly compact set of representative exemplars, where the model trained with these exemplars ideally should have comparable performance with respect to the one trained with the full dataset. While most of existing works upon dataset distillation focus on supervised datasets, \todo{we instead aim to distill images and their self-supervisedly trained representations into a distilled set. This procedure, named as Self-Supervised Dataset Distillation, effectively extracts rich information from real datasets, yielding the distilled sets with enhanced cross-architecture generalizability.} Particularly, in order to preserve the key characteristics of original dataset more faithfully and compactly, several novel techniques are proposed: 1) we introduce an innovative parameterization upon images and representations via distinct low-dimensional bases, where the base selection for parameterization is experimentally shown to play a crucial role; 2) we tackle the instability induced by the randomness of data augmentation -- a key component in self-supervised learning but being underestimated in the prior work of self-supervised dataset distillation -- by utilizing predetermined augmentations; 3) we further leverage a lightweight network to model the connections among the representations of augmented views from the same image, leading to more compact pairs of distillation. Extensive experiments conducted on various datasets validate the superiority of our approach in terms of distillation efficiency, cross-architecture generalization, and transfer learning performance. | dataset distillation, self-supervised learning | null | 969 | null | [
-0.07191573083400726,
-0.07523747533559799,
0.09784480184316635,
0.037346627563238144,
0.051775891333818436,
-0.0024788235314190388,
-0.011099344119429588,
-0.03880123421549797,
-0.019395027309656143,
-0.1275058388710022,
-0.029433224350214005,
-0.08570975065231323,
-0.00888720154762268,
0.009264054708182812,
-0.04807788133621216,
0.03995238617062569,
0.08229605108499527,
0.13710851967334747,
-0.1771264374256134,
-0.04837135598063469,
-0.005227869376540184,
-0.01284159254282713,
-0.0017088036984205246,
0.01617293246090412,
0.05940157547593117,
-0.011664148420095444,
0.05889861285686493,
0.011536307632923126,
0.035101622343063354,
-0.07042344659566879,
0.06206836178898811,
-0.0374143123626709,
-0.04994615539908409,
0.0191184189170599,
-0.029572071507573128,
0.05880073457956314,
0.0028135329484939575,
0.0832720696926117,
-0.01551023218780756,
-0.031727615743875504,
0.062067024409770966,
0.05577483028173447,
-0.000350525020621717,
0.07170537859201431,
0.04832368344068527,
0.07050523161888123,
0.010805608704686165,
-0.03157360479235649,
0.007957922294735909,
0.01227280218154192,
-0.05072275176644325,
0.027353301644325256,
-0.07394196093082428,
0.11131306737661362,
-0.006623539142310619,
-0.047063957899808884,
0.03539538010954857,
0.01379704661667347,
-0.03416075184941292,
0.024749401956796646,
0.023366909474134445,
-0.0643988624215126,
-0.009167948737740517,
-0.02390849031507969,
-0.015515164472162724,
0.0021532177925109863,
0.056334201246500015,
0.05185571685433388,
0.01780638098716736,
-0.025031769648194313,
-0.024342156946659088,
0.08443434536457062,
-0.025562407448887825,
0.06923739612102509,
0.0377797856926918,
0.049728792160749435,
0.027919845655560493,
0.04678606614470482,
0.045764386653900146,
-0.06996334344148636,
-0.007977905683219433,
0.016202077269554138,
0.05408647656440735,
0.008393950760364532,
0.08257439732551575,
-0.005912849213927984,
-0.023064937442541122,
0.003133653197437525,
0.034386876970529556,
0.018454376608133316,
-0.06786666810512543,
-0.014458888210356236,
-0.0009700056980364025,
-0.04911494627594948,
0.029278084635734558,
0.028889168053865433,
0.022064922377467155,
-0.08193641155958176,
-0.018781648948788643,
0.08409841358661652,
-0.014540711417794228,
0.05348462983965874,
0.04255012422800064,
0.014347829855978489,
0.05300125107169151,
-0.032160434871912,
0.04720770940184593,
0.0574706494808197,
0.04777950048446655,
-0.10074081271886826,
0.03781474381685257,
0.02071271650493145,
-0.05124964937567711,
-0.031722791492938995,
0.06524819135665894,
-0.008503050543367863,
-0.026714881882071495,
-0.01459316536784172,
-0.009850400499999523,
-0.018013419583439827,
-0.02543998695909977,
-0.01946362666785717,
-0.005730214528739452,
0.030899953097105026,
-0.017047038301825523,
-0.06319638341665268,
-0.10681552439928055,
6.254762829055175e-33,
0.03593965247273445,
0.050746358931064606,
0.04136423394083977,
-0.011331337504088879,
0.00788127351552248,
0.00632582139223814,
-0.020080236718058586,
0.034037139266729355,
-0.04297659918665886,
0.05601339042186737,
-0.03372425585985184,
-0.00871425773948431,
-0.06799524277448654,
0.07623957097530365,
0.05629908666014671,
-0.02466568350791931,
-0.07543964684009552,
0.06298049539327621,
-0.00444433931261301,
-0.06062346696853638,
0.03673495724797249,
0.03583342581987381,
-0.023669341579079628,
-0.027812043204903603,
-0.021908985450863838,
-0.0241326242685318,
0.07044373452663422,
-0.014189142733812332,
-0.029848329722881317,
-0.01850304938852787,
-0.06679273396730423,
0.011561891064047813,
0.058172557502985,
0.03082040697336197,
0.010363176465034485,
-0.04594257473945618,
-0.014172932133078575,
-0.02420133724808693,
0.06039804592728615,
0.004363690037280321,
0.04107112064957619,
0.05685293674468994,
0.0309585127979517,
0.022058594971895218,
-0.10381962358951569,
-0.06906601041555405,
-0.028503233566880226,
0.00964048970490694,
-0.04559176787734032,
-0.08142409473657608,
0.0651591345667839,
0.015613977797329426,
-0.03159908205270767,
-0.08094559609889984,
-0.07506897300481796,
-0.02443467266857624,
0.02413875050842762,
0.04760342836380005,
0.07556154578924179,
0.005361618008464575,
-0.024555783718824387,
0.045859239995479584,
-0.015351083129644394,
0.11731627583503723,
0.0004976096679456532,
-0.02399563416838646,
0.056963369250297546,
0.05833328142762184,
0.032762233167886734,
-0.006187197752296925,
-0.09712755680084229,
0.028726262971758842,
-0.037507314234972,
-0.08910524845123291,
0.06055038049817085,
-0.03794530779123306,
0.05506466329097748,
-0.06322325766086578,
-0.03545315936207771,
0.029056645929813385,
0.0004421309567987919,
0.0098728621378541,
-0.034460779279470444,
-0.056384023278951645,
-0.05074354261159897,
-0.005010365042835474,
0.02348611317574978,
-0.05354171618819237,
0.023623963817954063,
-0.040482502430677414,
-0.07731999456882477,
0.04020160436630249,
-0.04404497519135475,
-0.06269663572311401,
-0.019298214465379715,
-4.0384530865782014e-33,
0.05552566051483154,
0.07238885760307312,
-0.04683776944875717,
0.10646284371614456,
0.04480380564928055,
-0.02532638981938362,
-0.01034596748650074,
0.03778665140271187,
-0.016214916482567787,
-0.06387576460838318,
-0.006829504854977131,
-0.024838143959641457,
0.032009098678827286,
-0.06613562256097794,
-0.03606848418712616,
0.01734187640249729,
-0.03119160793721676,
0.001275390386581421,
-0.06767761707305908,
-0.005059176590293646,
-0.05068375542759895,
0.0864955335855484,
-0.04126259684562683,
0.0645855963230133,
-0.053045812994241714,
0.053932495415210724,
-0.036433376371860504,
0.08482971042394638,
0.03334221988916397,
-0.02218071185052395,
-0.06963962316513062,
0.0186836626380682,
-0.016806956380605698,
-0.07034486532211304,
-0.10576599836349487,
0.043202824890613556,
0.01881146989762783,
-0.059976235032081604,
-0.043717000633478165,
0.068600133061409,
0.005812224932014942,
0.00731767388060689,
-0.10450134426355362,
0.03463944420218468,
-0.0690009668469429,
-0.05191488564014435,
-0.06459960341453552,
-0.00590776139870286,
0.02190563827753067,
0.08449520915746689,
-0.01738518849015236,
0.010775471106171608,
-0.12725421786308289,
0.06913748383522034,
0.0066506690345704556,
-0.08093609660863876,
-0.004124719183892012,
-0.03204956650733948,
0.032793834805488586,
0.05458591878414154,
-0.07897187024354935,
-0.043286681175231934,
0.04735846444964409,
-0.07503325492143631,
-0.006163434125483036,
-0.08559863269329071,
-0.007154450286179781,
0.012524014338850975,
-0.01828099600970745,
0.038822073489427567,
0.01335185021162033,
-0.05596332624554634,
0.049912821501493454,
0.030084336176514626,
-0.05588291957974434,
-0.059083886444568634,
0.034979429095983505,
-0.013065963983535767,
0.012944377958774567,
-0.009929957799613476,
0.024138158187270164,
-0.030431192368268967,
0.02055903524160385,
0.0799388587474823,
0.13060986995697021,
0.06955667585134506,
0.015546671114861965,
0.049424219876527786,
0.03982733190059662,
0.034019485116004944,
-0.003891703439876437,
0.007835126481950283,
-0.010215611197054386,
0.11228340119123459,
0.005192324984818697,
-4.245429252591748e-8,
-0.08508512377738953,
0.07196573168039322,
0.01814718171954155,
0.049633245915174484,
-0.0034280268009752035,
-0.050401121377944946,
0.02897716872394085,
0.19077998399734497,
-0.02075621485710144,
-0.04503882676362991,
0.08774739503860474,
-0.005778863560408354,
-0.05919625237584114,
-0.03769707307219505,
0.005268295295536518,
0.0569133386015892,
0.042485788464546204,
0.045119695365428925,
0.003902900032699108,
0.040751952677965164,
0.0037174206227064133,
-0.005087736528366804,
-0.006879381835460663,
-0.06398418545722961,
0.10931213200092316,
-0.07874076813459396,
-0.026276199147105217,
0.09471151232719421,
-0.0019600586965680122,
0.027792835608124733,
0.0034699286334216595,
-0.0016280078561976552,
0.06235232576727867,
-0.09767316281795502,
-0.019974229857325554,
0.11407269537448883,
0.006104908883571625,
0.02440275438129902,
-0.07268105447292328,
-0.09829386323690414,
-0.020153217017650604,
0.07614627480506897,
-0.08726585656404495,
0.011128552258014679,
-0.017202673479914665,
-0.032979100942611694,
0.04960314929485321,
0.0029988919850438833,
0.015964791178703308,
0.045629505068063736,
0.04197278991341591,
-0.043572600930929184,
0.09735772758722305,
0.05832865089178085,
0.03132243454456329,
0.0033851920161396265,
-0.033877599984407425,
-0.001489496324211359,
0.023533713072538376,
0.038159389048814774,
-0.0034415768459439278,
-0.0522388257086277,
-0.016264531761407852,
0.0013895059237256646
] |
GALA: Geometry-Aware Local Adaptive Grids for Detailed 3D Generation | https://openreview.net/forum?id=KYOdZRR6nr | [
"Dingdong Yang",
"Yizhi Wang",
"Konrad Schindler",
"Ali Mahdavi Amiri",
"Hao Zhang"
] | Poster | We propose GALA, a novel representation of 3D shapes that (i) excels at capturing and reproducing complex geometry and surface details, (ii) is computationally efficient, and (iii) lends itself to 3D generative modelling with modern, diffusion-based schemes. The key idea of GALA is to exploit both the global sparsity of surfaces within a 3D volume and their local surface properties. *Sparsity* is promoted by covering only the 3D object boundaries, not empty space, with an ensemble of tree root voxels. Each voxel contains an octree to further limit storage and compute to regions that contain surfaces. *Adaptivity* is achieved by fitting one local and geometry-aware coordinate frame in each non-empty leaf node. Adjusting the orientation of the local grid, as well as the anisotropic scales of its axes, to the local surface shape greatly increases the amount of detail that can be stored in a given amount of memory, which in turn allows for quantization without loss of quality. With our optimized C++/CUDA implementation, GALA can be fitted to an object in less than 10 seconds. Moreover, the representation can efficiently be flattened and manipulated with transformer networks. We provide a cascaded generation pipeline capable of generating 3D shapes with great geometric detail. For more information, please visit our [project page](https://santisy.github.io/GALA/). | Generative Model, 3D, Computer Graphics | A new representation for 3D geometry generation | 963 | 2410.10037 | [
-0.00329400016926229,
-0.06382209062576294,
-0.015947144478559494,
0.01654008962213993,
0.06589325517416,
-0.07676918059587479,
-0.10981208086013794,
-0.07608238607645035,
0.03805771842598915,
-0.03534059599041939,
-0.04800442233681679,
-0.08183880895376205,
-0.01423767302185297,
0.033888183534145355,
-0.033237870782613754,
-0.04577212408185005,
0.013231612741947174,
0.08059012144804001,
-0.030226116999983788,
0.05334695056080818,
0.00550636975094676,
0.0007272763759829104,
-0.04968000948429108,
0.003981759771704674,
0.12235850840806961,
0.028777750208973885,
0.0032647105399519205,
-0.0682477205991745,
0.04890571907162666,
-0.055670056492090225,
0.08521116524934769,
0.04533986747264862,
-0.019557341933250427,
0.0398910753428936,
-0.02389451116323471,
0.0025525656528770924,
-0.0441807322204113,
-0.015315638855099678,
-0.019200848415493965,
-0.03650343790650368,
0.00536651536822319,
0.06899024546146393,
0.00580422580242157,
-0.012307979166507721,
0.06280240416526794,
0.040927328169345856,
0.006549876648932695,
0.002587956376373768,
-0.010629271157085896,
-0.037460438907146454,
0.004767261911183596,
-0.1297386735677719,
-0.05379721522331238,
-0.009864943102002144,
0.00014557127724401653,
0.07891526818275452,
0.050619564950466156,
-0.11985266208648682,
0.08788548409938812,
-0.008270017802715302,
0.030302003026008606,
-0.027181198820471764,
0.05847505107522011,
-0.003294049995020032,
-0.05633200705051422,
-0.057026103138923645,
0.013638203963637352,
-0.013627593405544758,
-0.004816295113414526,
0.023381253704428673,
0.0055291540920734406,
0.034589335322380066,
-0.060821183025836945,
0.004049181472510099,
0.04052169248461723,
0.01708833873271942,
0.02996116690337658,
-0.06702516973018646,
-0.014825332909822464,
0.017854483798146248,
-0.013164598494768143,
0.09458454698324203,
-0.03652125224471092,
-0.0013880433980375528,
-0.035577524453401566,
-0.034570079296827316,
-0.038072697818279266,
0.1011701449751854,
-0.06194131076335907,
-0.038134269416332245,
0.015816204249858856,
0.03360220417380333,
-0.10169515758752823,
-0.04987450689077377,
-0.08338537067174911,
-0.03176078200340271,
0.07275336235761642,
-0.07044446468353271,
0.00421663373708725,
0.040263764560222626,
-0.015172017738223076,
-0.02839423157274723,
0.045817986130714417,
0.02602696791291237,
0.031101439148187637,
-0.009902900084853172,
0.03320842608809471,
0.04977019876241684,
-0.03910251334309578,
0.062330394983291626,
0.05099620670080185,
-0.01234380342066288,
-0.045200929045677185,
0.018316898494958878,
-0.024656888097524643,
0.006734366994351149,
0.0011181952431797981,
-0.03514978662133217,
0.032673515379428864,
0.04273204132914543,
0.010293052531778812,
-0.00978381559252739,
-0.056885574012994766,
0.08492746949195862,
0.05202183872461319,
-0.054358143359422684,
-0.06277525424957275,
1.4386693525431584e-33,
0.014551034197211266,
0.05308195576071739,
-0.0013240630505606532,
0.07364937663078308,
0.14207011461257935,
0.039589300751686096,
0.02896488830447197,
-0.0365772508084774,
0.11109258234500885,
0.02925783395767212,
-0.1714242696762085,
0.04223894327878952,
-0.03674619644880295,
0.05927406996488571,
0.08594586700201035,
0.004298754036426544,
-0.02664649859070778,
0.04812627285718918,
-0.07735194265842438,
0.02365023083984852,
0.02919795550405979,
-0.040920425206422806,
-0.01894153468310833,
-0.010159184224903584,
0.009310398250818253,
0.03670770302414894,
0.07194197922945023,
-0.026889987289905548,
0.02129136398434639,
0.00027252372819930315,
-0.018229372799396515,
-0.037106312811374664,
-0.0024234075099229813,
0.003369803773239255,
0.039327915757894516,
0.026900792494416237,
-0.02382439747452736,
-0.061904046684503555,
-0.04139441251754761,
-0.029455162584781647,
-0.008351513184607029,
0.06651288270950317,
0.026942748576402664,
-0.02221882902085781,
-0.08626571297645569,
0.06497642397880554,
0.010205467231571674,
0.017501192167401314,
-0.0666317567229271,
0.013401065953075886,
0.09072116017341614,
0.044021837413311005,
-0.05084463953971863,
-0.02656484954059124,
0.00028902749181725085,
-0.03212437778711319,
-0.03995528817176819,
-0.010520212352275848,
0.07745540142059326,
0.06316252052783966,
0.02013777196407318,
0.037991002202034,
-0.02468409761786461,
0.024634670466184616,
-0.02608651854097843,
0.0033537789713591337,
0.04666562005877495,
0.07485397160053253,
0.04823196679353714,
0.01483068149536848,
-0.07448133826255798,
-0.02898668497800827,
0.003873395500704646,
0.011723452247679234,
-0.019436417147517204,
-0.02514810487627983,
0.04504610598087311,
-0.047756507992744446,
0.05871567875146866,
-0.019102046266198158,
-0.05756644159555435,
0.013240688480436802,
-0.028377078473567963,
-0.07606885582208633,
-0.03654623031616211,
-0.10680216550827026,
0.0429258793592453,
0.03010084666311741,
-0.04063041880726814,
-0.06422986835241318,
-0.03759228065609932,
-0.06938839703798294,
0.0651094987988472,
-0.04521359130740166,
-0.08024868369102478,
-2.7343578170660325e-33,
0.025883549824357033,
-0.06845355778932571,
0.030118007212877274,
-0.0031413636170327663,
-0.0007416745065711439,
0.000028453006962081417,
-0.09813211113214493,
-0.05506327748298645,
-0.03441547974944115,
-0.10519728809595108,
-0.07371143251657486,
0.07555381953716278,
0.0722251608967781,
0.02480395697057247,
0.04209623113274574,
0.02169479802250862,
-0.02516304701566696,
-0.027500001713633537,
-0.14457589387893677,
0.02366478741168976,
-0.0331951305270195,
0.07411341369152069,
-0.03320075199007988,
-0.004452356602996588,
-0.10988698899745941,
0.03715648874640465,
-0.034699730575084686,
-0.012394800782203674,
0.016973793506622314,
0.047023527324199677,
0.007480408996343613,
-0.025946397334337234,
0.007856064476072788,
-0.06873508542776108,
-0.019191432744264603,
0.0004130334418732673,
0.01126904133707285,
0.02691192366182804,
0.012530322186648846,
-0.005983033217489719,
0.06838057935237885,
-0.03315052390098572,
-0.0400017648935318,
0.0725378468632698,
0.007303759455680847,
0.05158359184861183,
-0.02820967137813568,
0.04456396400928497,
0.007277381140738726,
0.08824267238378525,
-0.06169406697154045,
-0.017568163573741913,
0.03475702926516533,
0.00813737791031599,
0.09021850675344467,
-0.05439022183418274,
-0.12141097337007523,
-0.056927021592855453,
0.09393562376499176,
-0.06834731996059418,
-0.0812758281826973,
-0.08835116773843765,
0.000307444017380476,
-0.03010331653058529,
-0.027629762887954712,
-0.0037042475305497646,
0.017761705443263054,
-0.05944062024354935,
0.020088432356715202,
0.08527033776044846,
0.05678945034742355,
0.01709475927054882,
-0.05731156840920448,
0.013518983498215675,
-0.007521235384047031,
-0.02542887255549431,
-0.008338039740920067,
-0.009687287732958794,
0.048708803951740265,
-0.03658289462327957,
0.017275286838412285,
0.08157142996788025,
0.02062608301639557,
0.08392133563756943,
0.12940040230751038,
-0.022262193262577057,
0.03305533155798912,
0.04538588970899582,
0.028393905609846115,
0.052684884518384933,
-0.010947362519800663,
0.08214074373245239,
0.0064884815365076065,
0.011721917428076267,
-0.007329200394451618,
-4.2751022277798256e-8,
-0.0074528842233121395,
0.0656837522983551,
0.03596816956996918,
-0.068796306848526,
-0.023022912442684174,
-0.0583520382642746,
0.03803211823105812,
0.08023963868618011,
-0.02568790875375271,
-0.018244722858071327,
0.09967117011547089,
-0.0537225715816021,
-0.023657629266381264,
-0.02063189446926117,
0.07974733412265778,
0.1310786008834839,
0.014022729359567165,
-0.010395783931016922,
0.014111775904893875,
0.0872780829668045,
-0.062120307236909866,
0.07749227434396744,
0.003470662049949169,
0.03392777964472771,
0.045505281537771225,
-0.03874224051833153,
-0.0008071923512034118,
0.08022656291723251,
0.08182195574045181,
-0.06580816209316254,
0.06758009642362595,
0.023595597594976425,
0.03940770775079727,
0.09382367134094238,
-0.06758256256580353,
0.02482636086642742,
0.02498706616461277,
0.025815926492214203,
-0.016429483890533447,
-0.08083236217498779,
-0.017903326079249382,
-0.018207605928182602,
-0.011143279261887074,
-0.04202951118350029,
-0.05247402563691139,
0.08148518949747086,
0.02519160509109497,
-0.012121782638132572,
-0.05328162759542465,
0.07868070900440216,
-0.0833326131105423,
-0.015565648674964905,
-0.04932454973459244,
0.0275094173848629,
0.009760702028870583,
0.02921980991959572,
-0.07656880468130112,
-0.03431927412748337,
0.11111763119697571,
0.015990233048796654,
-0.002008827170357108,
-0.04550064355134964,
-0.04308824986219406,
-0.03506462275981903
] |
MMEgo: Towards Building Egocentric Multimodal LLMs for Video QA | https://openreview.net/forum?id=67sSPPAZiG | [
"Hanrong Ye",
"Haotian Zhang",
"Erik Daxberger",
"Lin Chen",
"Zongyu Lin",
"Yanghao Li",
"Bowen Zhang",
"Haoxuan You",
"Dan Xu",
"Zhe Gan",
"Jiasen Lu",
"Yinfei Yang"
] | Poster | This research aims to comprehensively explore building a multimodal foundation model for egocentric video understanding.
To achieve this goal, we work on three fronts.
First, as there is a lack of QA data for egocentric video understanding, we automatically generate 7M high-quality QA samples for egocentric videos ranging from 30 seconds to one hour long in Ego4D based on human-annotated data.
This is one of the largest egocentric QA datasets.
Second, we contribute a challenging egocentric QA benchmark with 629 videos and 7,026 questions to evaluate the models' ability in recognizing and memorizing visual details across videos of varying lengths. We introduce a new de-biasing evaluation method to help mitigate the unavoidable language bias present in the models being evaluated.
Third, we propose a specialized multimodal architecture featuring a novel ``Memory Pointer Prompting" mechanism. This design includes a global glimpse step to gain an overarching understanding of the entire video and identify key visual information, followed by a fallback step that utilizes the key visual information to generate responses. This enables the model to more effectively comprehend extended video content.
With the data, benchmark, and model, we build MM-Ego, an egocentric multimodal LLM that shows powerful performance on egocentric video understanding. | multimodal models | MM-Ego, an egocentric multimodal LLM that shows powerful performance on egocentric video understanding with the contribution on: (i) Egocentric QA Data Engine (ii) Memory Pointer Prompting Mechanism (iii) EgoMemoria benchmark. | 962 | null | [
0.030962655320763588,
-0.12088984251022339,
-0.020337393507361412,
-0.05418688431382179,
0.09911202639341354,
-0.001483980449847877,
-0.02086091972887516,
0.041373398154973984,
0.08859771490097046,
-0.04753314331173897,
-0.02930631674826145,
-0.08304160833358765,
-0.0627712532877922,
0.07551487535238266,
0.010804724879562855,
-0.00765090761706233,
0.033048711717128754,
0.0756249874830246,
-0.11894582957029343,
-0.009515657089650631,
0.029438477009534836,
-0.06357789784669876,
0.054709479212760925,
0.006051086355000734,
0.03644530475139618,
0.05010012537240982,
0.030195750296115875,
-0.009482462890446186,
0.005252073984593153,
-0.056968577206134796,
0.03408082574605942,
-0.004833622369915247,
0.05571812018752098,
0.08067834377288818,
-0.043458785861730576,
0.021538425236940384,
-0.07648343592882156,
0.009259587153792381,
-0.09445235878229141,
-0.0790722593665123,
0.008248839527368546,
0.04655127227306366,
0.024078402668237686,
0.00927121751010418,
0.06933991611003876,
0.03600529208779335,
-0.038967832922935486,
-0.021983889862895012,
0.035467296838760376,
-0.016275553032755852,
-0.11139317601919174,
-0.06585122644901276,
-0.026407448574900627,
0.04022883251309395,
-0.05268123000860214,
0.047174565494060516,
0.03869090601801872,
-0.01690700836479664,
0.02162707783281803,
0.026639405637979507,
0.014170081354677677,
-0.10232532769441605,
0.03069022297859192,
0.007862869650125504,
-0.009857605211436749,
-0.07375410944223404,
0.09994381666183472,
-0.024296969175338745,
-0.005104479379951954,
-0.01620348170399666,
-0.055390771478414536,
0.025589171797037125,
0.019027426838874817,
-0.004301135893911123,
0.022540370002388954,
-0.01022390741854906,
-0.030147437006235123,
-0.042006704956293106,
0.026942111551761627,
-0.11712642759084702,
0.06764085590839386,
-0.036292664706707,
0.06131550669670105,
0.00815454963594675,
0.11146518588066101,
0.01677406020462513,
0.00967472419142723,
0.0631585642695427,
-0.06638120114803314,
0.062293656170368195,
-0.0911623165011406,
-0.05483114346861839,
-0.03324798494577408,
-0.08736880123615265,
0.07555507123470306,
0.03266189992427826,
0.024452265352010727,
-0.11492449790239334,
-0.03420562297105789,
0.023651134222745895,
0.05094835162162781,
0.031246203929185867,
-0.0375339537858963,
-0.01619892194867134,
0.05032578855752945,
0.01067088171839714,
0.08202067762613297,
0.008923313580453396,
-0.005384726915508509,
-0.043476592749357224,
-0.0036556562408804893,
-0.014083653688430786,
0.030488591641187668,
-0.019734380766749382,
0.04436412826180458,
0.007359436247497797,
0.1004888191819191,
0.03447273373603821,
0.041900765150785446,
-0.01928870379924774,
0.019631873816251755,
-0.04148726165294647,
-0.0037288751918822527,
0.004383913706988096,
0.044556763023138046,
-0.019025737419724464,
-0.029422789812088013,
1.027904422920826e-33,
0.05238192155957222,
0.038249216973781586,
0.028784507885575294,
0.050337158143520355,
0.08131592720746994,
-0.021684855222702026,
0.016120335087180138,
0.05860113725066185,
-0.035437680780887604,
-0.038024142384529114,
0.06585875898599625,
-0.01392135489732027,
-0.022110672667622566,
0.09012306481599808,
-0.017505710944533348,
-0.012765699066221714,
-0.07528387755155563,
0.010152667760848999,
-0.02682643011212349,
0.04211216792464256,
0.013271735981106758,
-0.01956542208790779,
0.00894508883357048,
-0.0013482419308274984,
0.008338745683431625,
0.02277401089668274,
0.09730236977338791,
0.011424096301198006,
-0.03818296641111374,
-0.009457966312766075,
-0.13352610170841217,
-0.018545592203736305,
-0.043160177767276764,
-0.04847468063235283,
0.05146506056189537,
-0.03730478510260582,
-0.058564674109220505,
-0.010412820614874363,
0.019535500556230545,
-0.04602698236703873,
-0.02660374343395233,
0.039373453706502914,
-0.018096229061484337,
-0.057456016540527344,
-0.14586322009563446,
-0.017018647864460945,
0.07209587097167969,
0.09854019433259964,
-0.05333646759390831,
-0.00429105618968606,
0.08948558568954468,
0.02485303021967411,
-0.04985673353075981,
-0.10173850506544113,
-0.07426553219556808,
-0.006355335004627705,
0.07632959634065628,
0.024630660191178322,
-0.014471550472080708,
-0.01085230428725481,
0.010492308996617794,
0.10331197828054428,
-0.010623055510222912,
0.08559887111186981,
0.037300173193216324,
-0.022438960149884224,
-0.03672120347619057,
0.12850512564182281,
0.06411319971084595,
0.06018152832984924,
-0.0026187750045210123,
0.00036642709164880216,
0.041223566979169846,
-0.029931416735053062,
-0.03561608865857124,
0.04603549465537071,
0.05291061848402023,
-0.041069578379392624,
0.04105781018733978,
0.04854578524827957,
0.05401831865310669,
-0.001631452003493905,
0.040498483926057816,
-0.036802150309085846,
-0.03626031056046486,
-0.04323889687657356,
0.04942701756954193,
-0.1177879273891449,
-0.0628080889582634,
0.01691405102610588,
-0.012772793881595135,
-0.006202119868248701,
0.025137515738606453,
0.011067267507314682,
-0.07986808568239212,
-3.7639960482365732e-34,
0.06629370898008347,
0.04010966047644615,
-0.01855388656258583,
0.05839661508798599,
0.017036499455571175,
-0.04872599244117737,
0.005086075514554977,
0.03185826912522316,
0.013536506332457066,
-0.057501085102558136,
-0.03829759731888771,
-0.02159886434674263,
0.019339120015501976,
0.015184730291366577,
0.031781215220689774,
-0.010818700306117535,
0.016874250024557114,
-0.029761221259832382,
-0.05947062745690346,
0.018775885924696922,
-0.0004093357711099088,
0.041893310844898224,
-0.03986223787069321,
-0.007636609021574259,
-0.09956233203411102,
0.07465611398220062,
-0.014835883863270283,
0.016384458169341087,
-0.032472845166921616,
-0.0849238932132721,
0.01897791400551796,
-0.04468280076980591,
-0.026089632883667946,
-0.09866447001695633,
-0.043234799057245255,
0.03396831452846527,
0.05308852344751358,
-0.05057080090045929,
-0.028322715312242508,
0.08839982002973557,
0.07113045454025269,
0.014324006624519825,
-0.057770781219005585,
-0.009362732991576195,
0.0007560167578049004,
-0.05232350155711174,
-0.042149174958467484,
0.035620804876089096,
-0.07213402539491653,
0.10204334557056427,
-0.05184344947338104,
0.060855839401483536,
-0.05988452956080437,
0.010351505130529404,
-0.055137749761343,
-0.08592770993709564,
-0.06526977568864822,
0.03645956143736839,
0.10022146254777908,
0.019876379519701004,
-0.014413155615329742,
-0.041154541075229645,
-0.04073890298604965,
-0.049264661967754364,
0.02741030789911747,
0.003115798346698284,
-0.011496198363602161,
-0.03157957270741463,
-0.035327501595020294,
0.0007611713372170925,
0.08219469338655472,
-0.015310524962842464,
-0.0923956111073494,
0.013086755760014057,
0.04192578047513962,
0.049156107008457184,
-0.0023986506275832653,
-0.02660420723259449,
0.03765786811709404,
-0.06341007351875305,
-0.011994141153991222,
-0.07474538683891296,
0.022326195612549782,
0.09723109006881714,
0.033842533826828,
0.09755896031856537,
0.06896042078733444,
0.07523535192012787,
-0.01292380876839161,
0.06483085453510284,
0.003783063031733036,
0.04485122114419937,
-0.020011402666568756,
0.07021691650152206,
-0.021614234894514084,
-3.51035680523637e-8,
-0.1299004852771759,
0.010063634254038334,
-0.01321741845458746,
0.01046659704297781,
-0.02211269736289978,
-0.12488091737031937,
0.0030092280358076096,
0.04735906422138214,
0.038451891392469406,
-0.03428027778863907,
0.0556839182972908,
-0.005921525880694389,
-0.01632111333310604,
0.044208358973264694,
-0.019079603254795074,
0.064511738717556,
0.020183579996228218,
0.07405813038349152,
-0.00945338699966669,
0.03061431087553501,
0.05004272237420082,
0.02048673667013645,
-0.007952774874866009,
-0.05362787842750549,
0.004285065922886133,
-0.023765437304973602,
0.007819360122084618,
0.11026623845100403,
0.043970707803964615,
-0.03718799352645874,
0.013381035067141056,
0.01841377280652523,
-0.005171286873519421,
-0.09535127878189087,
-0.01939353346824646,
0.010794667527079582,
0.03813435509800911,
-0.019202616065740585,
0.002360207960009575,
-0.004951287526637316,
0.003617939306423068,
-0.0034963854122906923,
-0.026959743350744247,
0.07862179726362228,
-0.06307899206876755,
0.010719113983213902,
-0.03861590847373009,
-0.12493722885847092,
-0.021366560831665993,
-0.04465383663773537,
-0.010376032441854477,
-0.06171296164393425,
0.04410352557897568,
0.11139585077762604,
0.05274132266640663,
-0.017825020477175713,
0.09762439131736755,
0.046937812119722366,
0.03025182895362377,
0.05008593201637268,
0.04133941978216171,
0.029196245595812798,
-0.08344604820013046,
0.03215732425451279
] |
Beyond correlation: The impact of human uncertainty in measuring the effectiveness of automatic evaluation and LLM-as-a-judge | https://openreview.net/forum?id=E8gYIrbP00 | [
"Aparna Elangovan",
"Lei Xu",
"Jongwoo Ko",
"Mahsa Elyasi",
"Ling Liu",
"Sravan Babu Bodapati",
"Dan Roth"
] | Poster | The effectiveness of automatic evaluation of generative models is typically measured by comparing the labels generated via automation with human labels using correlation metrics.
However, metrics like Krippendorff's $\alpha$ and Randolph's $\kappa$ were originally designed to measure the reliability of human labeling, thus make assumptions about typical human labeling behavior, and these assumptions may not be applicable to machine generated labels.
In this paper, we show how *relying on a single aggregate correlation score* can obscure fundamental differences between human labels and those from automatic evaluation, including LLM-as-a-Judge.
Specifically, we demonstrate that when the proportion of samples with variation or uncertainty in human assigned labels is relatively high, machine labels (generated by automatic evaluation methods) may superficially appear to have similar or better correlation with the human majority label compared to the human-to-human (HH) correlation.
This can create the illusion that labels from automatic evaluation approximates the human majority label.
However, as the proportion of samples with consistent human labels increases, the correlation between machine and human labels fall well below HH correlation.
Based on these findings, we first propose *stratifying data by human label uncertainty* to provide a more robust analysis of automatic evaluation performance. Second, recognizing that uncertainty and variation are inherent in perception-based human evaluations, such as those involving attitudes or preferences, we introduce a new metric -*binned Jensen-Shannon Divergence for perception* for such scenarios to better measure the effectiveness of automatic evaluations. Third, we present visualization techniques -- *perception charts*, to contextualize correlation measures appropriately and to show the strengths and limitations of automatic evaluation. We have open-sourced our analysis and visualization tools at https://github.com/amazon-science/BeyondCorrelation. | Automated evaluation, LLM as a judge, correlation measures | Impact of noise when measuring effectiveness of automated evaluation | 956 | null | [
-0.020016836002469063,
-0.046519119292497635,
-0.062001824378967285,
0.027185047045350075,
0.013533988036215305,
-0.009774674661457539,
0.04096326231956482,
0.03139720857143402,
0.05395932495594025,
-0.08117877691984177,
-0.04520531743764877,
-0.1221427470445633,
0.08316607028245926,
-0.0024673037696629763,
-0.051609400659799576,
-0.027429599314928055,
0.13224448263645172,
0.006209766957908869,
-0.035169679671525955,
-0.05666540190577507,
-0.04588448628783226,
0.04392846301198006,
0.056875936686992645,
0.024494657292962074,
0.006334815640002489,
-0.06521279364824295,
-0.0732354000210762,
0.04489487409591675,
0.03165466710925102,
-0.011421624571084976,
-0.015171322040259838,
0.010329078882932663,
0.0731860101222992,
-0.005916036199778318,
-0.026887530460953712,
-0.02822233736515045,
-0.03438405692577362,
0.09488026797771454,
0.0688667744398117,
-0.04915894195437431,
-0.006706957705318928,
-0.028992142528295517,
0.05454881861805916,
0.03547882288694382,
-0.028750453144311905,
0.08255983144044876,
-0.04990777745842934,
-0.07683635503053665,
-0.14472727477550507,
0.08619153499603271,
-0.08519716560840607,
0.010215381160378456,
0.048406776040792465,
-0.0716799721121788,
-0.0718662217259407,
0.07391082495450974,
0.07562196254730225,
0.02292652800679207,
0.009107004851102829,
0.04629823938012123,
-0.08086130023002625,
-0.09471394121646881,
-0.0843287780880928,
0.04359797388315201,
0.08465268462896347,
0.03132086247205734,
0.024946430698037148,
-0.0247378908097744,
-0.03731611743569374,
0.05908524990081787,
0.011201784014701843,
0.08498504012823105,
-0.028009431436657906,
0.09871824085712433,
-0.03204173967242241,
0.04289498180150986,
-0.03466203436255455,
-0.028484592214226723,
0.08182825148105621,
-0.017597584053874016,
-0.0032613174989819527,
0.05563640221953392,
0.018816137686371803,
0.02591514028608799,
0.10492187738418579,
0.02585979737341404,
0.04349322244524956,
0.014868649654090405,
-0.06850277632474899,
0.02562677673995495,
-0.04789119213819504,
-0.050726890563964844,
0.04478053003549576,
0.0020795497111976147,
0.006775501649826765,
0.03664455935359001,
-0.01825709640979767,
0.006852858234196901,
0.06651115417480469,
0.06352139264345169,
-0.014456716366112232,
0.054024435579776764,
-0.06565779447555542,
-0.05522167310118675,
0.028491700068116188,
-0.01672726683318615,
0.013714248314499855,
0.008826352655887604,
0.06856204569339752,
-0.04256322607398033,
-0.020428944379091263,
-0.04354288801550865,
-0.0307795200496912,
0.006912400480359793,
0.02660985477268696,
-0.04579467326402664,
0.04769030958414078,
-0.002461905125528574,
0.03884314373135567,
-0.015868984162807465,
-0.04603230580687523,
-0.032685212790966034,
0.04570205509662628,
0.011043519712984562,
0.029656581580638885,
0.014190404675900936,
0.009189958684146404,
3.4591272261647454e-33,
0.05709455907344818,
-0.061889808624982834,
0.046942442655563354,
0.053462035953998566,
0.0003907468926627189,
-0.03558214753866196,
-0.11472728848457336,
-0.023097988218069077,
-0.022419584915041924,
0.056833188980817795,
0.021300319582223892,
0.0776943787932396,
-0.04466143250465393,
0.054799240082502365,
0.031072037294507027,
0.07790250331163406,
-0.005812079645693302,
0.08219205588102341,
-0.1287846863269806,
-0.018572639673948288,
0.041017357259988785,
-0.01507509034126997,
0.030063055455684662,
-0.01928100362420082,
0.010296888649463654,
-0.03841915354132652,
0.011309671215713024,
0.06768395751714706,
0.009588472545146942,
-0.00788224395364523,
-0.08381037414073944,
-0.04369521513581276,
0.10587064921855927,
0.009792502969503403,
-0.011309025809168816,
-0.06428483873605728,
-0.0044479151256382465,
0.014122633263468742,
-0.004551207181066275,
-0.0009396917303092778,
0.008166646584868431,
0.04198839142918587,
0.052737798541784286,
-0.04020892083644867,
-0.05931040272116661,
0.02251194603741169,
0.019803360104560852,
-0.003066487843170762,
-0.020858246833086014,
-0.03651433065533638,
-0.005185284186154604,
0.0536469966173172,
0.05937950685620308,
0.024975737556815147,
-0.08411628752946854,
0.004667640198022127,
-0.019725117832422256,
-0.008376000449061394,
-0.037844568490982056,
-0.03808291628956795,
-0.03835074603557587,
0.09845168888568878,
0.0028977852780371904,
0.019206931814551353,
0.06339923292398453,
0.09253644198179245,
0.025162791833281517,
-0.00022527902910951525,
0.04733067750930786,
0.02614942193031311,
0.011396029964089394,
-0.006267040967941284,
-0.07520764321088791,
-0.06179042533040047,
0.0796942263841629,
-0.021329842507839203,
-0.000951254740357399,
0.01135545689612627,
-0.025620926171541214,
-0.04065214470028877,
-0.01360312383621931,
-0.03546790033578873,
-0.04268594831228256,
-0.06547366827726364,
-0.012545019388198853,
-0.029027655720710754,
0.03612702712416649,
-0.030564799904823303,
-0.022894615307450294,
0.037051234394311905,
0.03914265334606171,
-0.04428311809897423,
-0.030451199039816856,
0.05263226106762886,
-0.07942093163728714,
-4.811666454584224e-33,
-0.1333313286304474,
0.027547238394618034,
0.05361323431134224,
0.055915821343660355,
-0.06400939077138901,
-0.000775062944740057,
0.09159144759178162,
0.02428237721323967,
-0.02028513327240944,
-0.06295812875032425,
-0.05466843396425247,
-0.050050511956214905,
0.0022463046479970217,
0.0029022342059761286,
0.015071204863488674,
-0.04229451343417168,
-0.060331664979457855,
0.006068466231226921,
0.04265749454498291,
0.031762123107910156,
0.10909051448106766,
0.08436545729637146,
-0.08286336809396744,
0.04985015466809273,
-0.04710441827774048,
0.04151506721973419,
-0.0023894391488283873,
-0.009603524580597878,
0.007009527180343866,
-0.020740512758493423,
0.009609666652977467,
-0.05355055257678032,
-0.014723571017384529,
-0.0204154159873724,
-0.04752584174275398,
-0.04618965461850166,
0.06930384784936905,
0.009873218834400177,
-0.04898110404610634,
0.08286719024181366,
-0.08092084527015686,
0.029763299971818924,
-0.16820846498012543,
-0.07773151248693466,
-0.006663879845291376,
-0.03129038214683533,
-0.0489802323281765,
-0.07809173315763474,
-0.04958281293511391,
-0.011691853404045105,
-0.034518156200647354,
0.04387514665722847,
-0.09083089977502823,
0.029667604714632034,
-0.1268499493598938,
-0.11788121610879898,
-0.05462919548153877,
-0.061915747821331024,
0.014732602052390575,
0.10906990617513657,
-0.09134518355131149,
0.011001883074641228,
-0.01457900833338499,
0.0021393627393990755,
-0.020994968712329865,
0.011812841519713402,
-0.01934165693819523,
-0.02647055871784687,
-0.03139939159154892,
0.04052697494626045,
0.07907786220312119,
0.021895911544561386,
-0.03299860283732414,
-0.0070411935448646545,
0.01647811196744442,
-0.00248484848998487,
0.023359185084700584,
0.0003545962681528181,
-0.018000612035393715,
-0.04704922437667847,
0.02612898126244545,
-0.038344856351614,
0.03924385830760002,
0.06157815828919411,
-0.04791368916630745,
0.006733355578035116,
0.07941693067550659,
0.023606091737747192,
-0.004391001537442207,
0.05720933526754379,
-0.0006016900297254324,
-0.006462644319981337,
-0.07019639015197754,
0.014083385467529297,
-0.09474017471075058,
-4.896270766607813e-8,
-0.024639295414090157,
0.04184760898351669,
-0.006942251697182655,
0.09820766746997833,
0.0013793464750051498,
0.01872537098824978,
-0.0007908542174845934,
0.04467499256134033,
-0.054360177367925644,
0.004276173654943705,
0.035643547773361206,
-0.03478296473622322,
-0.09188303351402283,
0.011857173405587673,
0.0136349406093359,
0.016849299892783165,
0.006671423092484474,
0.097434401512146,
-0.006556608248502016,
0.010107842274010181,
0.03616001084446907,
0.01669287122786045,
0.01987711526453495,
-0.057781122624874115,
0.029935766011476517,
-0.044055331498384476,
-0.06324172765016556,
0.06755856424570084,
-0.021277908235788345,
0.041869156062603,
0.042032450437545776,
0.012610429897904396,
-0.0014082012930884957,
0.039869099855422974,
0.08819565922021866,
0.06766315549612045,
-0.05413315072655678,
-0.043830033391714096,
0.04933290556073189,
0.0532778725028038,
-0.03879856690764427,
0.056484561413526535,
-0.10384628176689148,
0.07724905759096146,
0.10178237408399582,
-0.003244231455028057,
-0.021809671074151993,
-0.048635125160217285,
-0.017285794019699097,
-0.05062926560640335,
0.025392288342118263,
-0.06376136094331741,
-0.043096110224723816,
0.019273152574896812,
0.06731598824262619,
-0.051780104637145996,
0.04037893936038017,
-0.04818751662969589,
0.011980435810983181,
0.05848291888833046,
0.093327097594738,
0.031393926590681076,
-0.006651776377111673,
-0.10670282691717148
] |
Federated Q-Learning with Reference-Advantage Decomposition: Almost Optimal Regret and Logarithmic Communication Cost | https://openreview.net/forum?id=FoUpv84hMw | [
"Zhong Zheng",
"Haochen Zhang",
"Lingzhou Xue"
] | Poster | In this paper, we consider model-free federated reinforcement learning for tabular episodic Markov decision processes. Under the coordination of a central server, multiple agents collaboratively explore the environment and learn an optimal policy without sharing their raw data. Despite recent advances in federated $Q$-learning algorithms achieving near-linear regret speedup with low communication cost, existing algorithms only attain suboptimal regrets compared to the information bound. We propose a novel model-free federated $Q$-Learning algorithm, termed FedQ-Advantage. Our algorithm leverages reference-advantage decomposition for variance reduction and adopts three novel designs: separate event-triggered communication and policy switching, heterogeneous communication triggering conditions, and optional forced synchronization. We prove that our algorithm not only requires a lower logarithmic communication cost but also achieves an almost optimal regret, reaching the information bound up to a logarithmic factor and near-linear regret speedup compared to its single-agent counterpart when the time horizon is sufficiently large. | Federated Learning, Reinforcement Learning, variance reduction, communication cost | This paper designs a federated Q-learning algorithm with variance reduction and reaches an almost optimal regret and a logarithmic communication cost. | 952 | null | [
-0.0027893129736185074,
-0.02242281846702099,
0.032034117728471756,
-0.004488927312195301,
0.04053566977381706,
0.0020493152551352978,
0.012970898300409317,
-0.014536230824887753,
0.118720144033432,
0.07500862330198288,
-0.0656193345785141,
0.01695926859974861,
0.01898413524031639,
0.028064120560884476,
0.04415423050522804,
0.013229825533926487,
0.04531736299395561,
-0.007155470550060272,
-0.02022813819348812,
-0.014107109047472477,
-0.04015251621603966,
-0.027564847841858864,
0.04448173940181732,
-0.00464941980317235,
0.07405327260494232,
0.003938053734600544,
-0.04839275777339935,
0.02331388182938099,
-0.006155115552246571,
-0.010236109606921673,
-0.0062360805459320545,
-0.02394932694733143,
-0.007151395082473755,
-0.1366652101278305,
-0.06752806901931763,
0.07741539180278778,
-0.09237143397331238,
0.014558421447873116,
-0.020737014710903168,
0.03935658186674118,
-0.004437824245542288,
0.003088858909904957,
-0.02932056598365307,
0.05311412736773491,
0.02072521299123764,
-0.0014574327506124973,
0.02321169152855873,
-0.00716799683868885,
-0.0036093334201723337,
-0.026934729889035225,
0.015276861377060413,
-0.005555641371756792,
0.007746836636215448,
0.04322236403822899,
0.02896093763411045,
0.03620521351695061,
-0.051596373319625854,
0.006124571431428194,
-0.05045867711305618,
-0.016410674899816513,
-0.08850223571062088,
-0.11653974652290344,
-0.020546870306134224,
-0.04525165259838104,
-0.07570507377386093,
-0.0035307416692376137,
-0.004198511131107807,
0.07276546955108643,
-0.0023176479153335094,
0.033149946480989456,
-0.06977178901433945,
0.02443360723555088,
0.02395053580403328,
-0.06931362301111221,
-0.00837993435561657,
0.0015530826058238745,
0.0021764172706753016,
-0.005773407872766256,
-0.0016353103565052152,
-0.028286755084991455,
-0.031246675178408623,
0.035987745970487595,
-0.006624991539865732,
-0.029980069026350975,
0.041184429079294205,
-0.04258616268634796,
0.02633034996688366,
0.05220711603760719,
0.08839494735002518,
-0.07943329215049744,
-0.012042669579386711,
0.055889248847961426,
0.0375964380800724,
-0.015565325506031513,
0.00027389690512791276,
0.05871911719441414,
-0.03762500360608101,
-0.04158873111009598,
-0.057219378650188446,
0.08090426027774811,
0.0524117685854435,
0.08522646874189377,
-0.024081263691186905,
0.04794174060225487,
0.023193560540676117,
-0.0268411822617054,
-0.006490555591881275,
0.054501794278621674,
0.040846895426511765,
-0.07597526907920837,
-0.009772269986569881,
0.03688783943653107,
0.02995804324746132,
0.04247232526540756,
-0.07104439288377762,
0.06060493737459183,
0.01993011310696602,
0.005267669912427664,
0.02338002808392048,
0.09365925937891006,
-0.03046514466404915,
-0.08344364166259766,
-0.011885048821568489,
0.05395034700632095,
-0.021950706839561462,
-0.0515342578291893,
-0.02628788724541664,
1.8504977372110004e-33,
0.0830259770154953,
0.05720468610525131,
0.03705710917711258,
-0.13936862349510193,
0.1146366223692894,
0.029102463275194168,
0.0773996040225029,
-0.03047015890479088,
-0.08322373777627945,
-0.027834327891469002,
0.06524112820625305,
0.015706779435276985,
-0.006769258063286543,
0.03489508479833603,
0.07610145956277847,
-0.038146380335092545,
-0.11902697384357452,
0.05376051366329193,
0.13725052773952484,
-0.14997941255569458,
0.01346499565988779,
-0.051077648997306824,
-0.05443882942199707,
0.012967102229595184,
0.03278130665421486,
-0.012961489148437977,
0.011713244952261448,
0.041091300547122955,
0.052477821707725525,
-0.008242384530603886,
-0.031182456761598587,
-0.0038533571641892195,
-0.06412049382925034,
-0.018720293417572975,
0.0021116400603204966,
0.009045479819178581,
-0.09830282628536224,
-0.00918486062437296,
-0.04548073559999466,
0.014023179188370705,
0.01602247543632984,
-0.045678429305553436,
0.025959288701415062,
-0.049525704234838486,
-0.14291153848171234,
-0.10451290756464005,
-0.01766171120107174,
0.004371479153633118,
-0.010494265705347061,
-0.11040670424699783,
0.05336419865489006,
-0.04554247483611107,
-0.029932763427495956,
-0.009499459527432919,
-0.007456191349774599,
0.027583036571741104,
0.02552189864218235,
0.07590138167142868,
-0.07018975168466568,
0.08606060594320297,
-0.015564736910164356,
-0.024189265444874763,
0.03966975957155228,
0.053790174424648285,
-0.0030182877089828253,
0.07644742727279663,
-0.014565159566700459,
0.0539555624127388,
0.07466550171375275,
-0.03904236853122711,
0.003711082972586155,
0.06200404092669487,
0.05783235281705856,
-0.04431375861167908,
0.11761637032032013,
0.006892466451972723,
0.023977451026439667,
-0.03688181936740875,
0.05786275863647461,
0.022882129997015,
-0.024107206612825394,
0.05158720165491104,
-0.08368432521820068,
0.014618603512644768,
-0.010186298750340939,
-0.07899903506040573,
-0.02212001010775566,
-0.0289035364985466,
-0.0720597580075264,
0.076869435608387,
-0.13144317269325256,
0.034520626068115234,
0.04771910980343819,
0.07004497200250626,
-0.009472860023379326,
-2.370196095666019e-33,
-0.006493192631751299,
-0.06665044277906418,
0.039954911917448044,
0.008295380510389805,
0.03489481657743454,
0.01027654204517603,
-0.034305546432733536,
-0.04475700482726097,
-0.019804541021585464,
-0.022553956136107445,
-0.0959242582321167,
0.03617795556783676,
0.0786946639418602,
0.029202036559581757,
-0.023707574233412743,
0.008209515362977982,
0.02643408067524433,
0.0011971383355557919,
-0.0180891714990139,
-0.014832671731710434,
0.015238186344504356,
0.042515356093645096,
-0.05764695256948471,
0.09509763866662979,
0.08653879165649414,
0.016189182177186012,
0.007973920553922653,
0.08989129215478897,
-0.009236625395715237,
-0.031108560040593147,
-0.10548723489046097,
-0.10546914488077164,
-0.027579329907894135,
-0.08046001940965652,
-0.026702387258410454,
0.1002226173877716,
0.02560245804488659,
0.031114595010876656,
-0.08624232560396194,
0.09201066195964813,
0.03586570546030998,
-0.10726819187402725,
-0.042344771325588226,
-0.00025158157222904265,
0.03803350031375885,
-0.01722870208323002,
-0.08143649250268936,
-0.03987308591604233,
0.04420893266797066,
0.012449448928236961,
0.013593276031315327,
0.02269742265343666,
-0.012971784919500351,
0.038441237062215805,
-0.07662243396043777,
-0.03426907956600189,
0.03460746631026268,
0.07989683002233505,
0.08636774122714996,
0.04945966973900795,
0.030201014131307602,
-0.015501698479056358,
-0.05498683452606201,
0.0666336640715599,
0.07211466878652573,
-0.017533808946609497,
0.021731434389948845,
-0.008214338682591915,
0.028819257393479347,
-0.015801655128598213,
0.01759280636906624,
0.029994333162903786,
-0.0289725661277771,
-0.02812678925693035,
0.14860799908638,
0.07179231196641922,
-0.020975491032004356,
-0.001493200659751892,
-0.07923734933137894,
0.015509791672229767,
0.051438190042972565,
0.05876684561371803,
0.005158214829862118,
0.02275153063237667,
0.04727131500840187,
0.057358596473932266,
0.042149338871240616,
0.045293040573596954,
0.0140977269038558,
-0.051415055990219116,
0.006005327217280865,
-0.031653691083192825,
0.042875371873378754,
0.024855773895978928,
-0.056089770048856735,
-4.2324614923927584e-8,
-0.02999851666390896,
0.008000842295587063,
0.05411160737276077,
-0.015278191305696964,
0.058181602507829666,
-0.010256147012114525,
0.02064524032175541,
0.00031919023604132235,
-0.0305715873837471,
-0.03566843643784523,
-0.0031076481100171804,
0.05286124721169472,
0.04248899221420288,
-0.004951709881424904,
-0.01597297564148903,
0.03646495193243027,
-0.01604882813990116,
-0.11652068793773651,
0.017138943076133728,
0.03488684073090553,
0.0480358749628067,
-0.030555449426174164,
-0.028499342501163483,
-0.0293793436139822,
0.00903404038399458,
-0.04499000310897827,
0.04293875768780708,
0.08092988282442093,
0.0714992880821228,
0.009324113838374615,
-0.0007351679960265756,
-0.025527626276016235,
0.015241028741002083,
0.004750734660774469,
-0.02035108022391796,
-0.005946046207100153,
0.013692802749574184,
-0.12111183255910873,
-0.01129060611128807,
-0.007719560991972685,
0.0076500363647937775,
0.014423085376620293,
-0.06505443155765533,
-0.020043522119522095,
0.027631321921944618,
0.009818781167268753,
-0.12600310146808624,
-0.08481265604496002,
0.0013296473771333694,
-0.03862366825342178,
0.07065293192863464,
-0.04299177974462509,
0.042474377900362015,
0.023711103945970535,
0.08080790936946869,
-0.0614413246512413,
0.07514074444770813,
-0.046509478241205215,
0.10384023934602737,
0.043794628232717514,
0.05901995673775673,
-0.029637925326824188,
-0.07721170037984848,
-0.004123782739043236
] |
Remove Symmetries to Control Model Expressivity and Improve Optimization | https://openreview.net/forum?id=Gv0TOAigIY | [
"Liu Ziyin",
"Yizhou Xu",
"Isaac L. Chuang"
] | Poster | When symmetry is present in the loss function, the model is likely to be trapped in a low-capacity state that is sometimes known as a ``collapse." Being trapped in these low-capacity states can be a major obstacle to training across many scenarios where deep learning technology is applied. We first prove two concrete mechanisms through which symmetries lead to reduced capacities and ignored features during training and inference. We then propose a simple and theoretically justified algorithm, \textit{syre}, to remove almost all symmetry-induced low-capacity states in neural networks. When this type of entrapment is especially a concern, removing symmetries with the proposed method is shown to correlate well with improved optimization or performance. A remarkable merit of the proposed method is that it is model-agnostic and does not require any knowledge of the symmetry. | model capacity, symmetry, optimization | null | 945 | 2408.15495 | [
-0.04959029704332352,
-0.03842997923493385,
0.05295821651816368,
0.02060825377702713,
0.000774436048232019,
0.00012869412603322417,
-0.05346454679965973,
-0.07613895833492279,
-0.03746207058429718,
-0.04336000978946686,
-0.0616113543510437,
0.037161462008953094,
-0.028779713436961174,
0.010568032041192055,
-0.020578917115926743,
0.02209971286356449,
0.04602514207363129,
0.10002046078443527,
-0.11862513422966003,
-0.0019217608496546745,
0.03246598690748215,
-0.05216339975595474,
-0.050369370728731155,
0.030957594513893127,
0.04126106947660446,
-0.0094612967222929,
-0.01571580022573471,
-0.07264888286590576,
-0.021423624828457832,
-0.08860713988542557,
0.059448160231113434,
-0.05784664303064346,
-0.060342706739902496,
-0.04983964189887047,
-0.0629056990146637,
0.0889398604631424,
-0.04437185823917389,
-0.005576242692768574,
0.0590190589427948,
-0.07959337532520294,
-0.014574892818927765,
0.046973273158073425,
-0.025145061314105988,
0.007524563930928707,
0.052384816110134125,
0.09688536077737808,
0.043320707976818085,
-0.04429619386792183,
0.003602709388360381,
-0.042467840015888214,
0.04762696847319603,
0.09245461225509644,
-0.0692879855632782,
0.04042641445994377,
-0.011487342417240143,
-0.0068674394860863686,
0.025380773469805717,
-0.015593701042234898,
-0.09933862835168839,
0.009799428284168243,
0.02723594754934311,
-0.10896706581115723,
-0.009408582001924515,
-0.04240245372056961,
0.012083129957318306,
-0.006471790838986635,
0.08056280016899109,
0.016898036003112793,
0.020203011110424995,
0.05725123733282089,
0.04831654950976372,
0.07954467087984085,
-0.078382708132267,
0.0017107968451455235,
0.08830729126930237,
0.05050301551818848,
0.072644904255867,
-0.006864907220005989,
0.039524223655462265,
0.022144384682178497,
-0.013474897481501102,
0.0014637045096606016,
-0.024638507515192032,
-0.009053255431354046,
0.09542117267847061,
-0.06684992462396622,
-0.08322025835514069,
-0.03692834824323654,
0.08014580607414246,
-0.01882718876004219,
-0.027090823277831078,
0.041548214852809906,
-0.009306889027357101,
-0.04301867634057999,
0.03887200728058815,
0.03139783442020416,
-0.009047497063875198,
-0.014308067969977856,
-0.1037231832742691,
0.1313754916191101,
0.0163129810243845,
0.030007904395461082,
0.08745893836021423,
-0.0246138758957386,
0.08421701192855835,
-0.06234877556562424,
0.048039767891168594,
-0.01750766672194004,
0.04931122437119484,
-0.1036834716796875,
0.05538766086101532,
0.018176376819610596,
0.08657749742269516,
0.06391443312168121,
-0.012498791329562664,
0.04904092475771904,
-0.050115715712308884,
-0.019341034814715385,
0.0016012242995202541,
0.06052085757255554,
-0.07290477305650711,
-0.017477627843618393,
-0.041819218546152115,
0.04876711964607239,
-0.01013900525867939,
0.02068091183900833,
-0.09421779215335846,
3.1904312580858484e-33,
0.0051940870471298695,
0.03074808418750763,
0.0395360141992569,
-0.0277091097086668,
0.08817560970783234,
0.019368821755051613,
0.01092164684087038,
-0.00467430055141449,
-0.02110755257308483,
0.04960014671087265,
-0.12135745584964752,
-0.060293592512607574,
-0.04180632904171944,
0.06884322315454483,
-0.011365092359483242,
0.010774307884275913,
-0.002580297412350774,
0.06201652064919472,
-0.004887640010565519,
-0.13967272639274597,
0.04765254631638527,
-0.009024941362440586,
-0.1025003120303154,
0.012767470441758633,
-0.08823838829994202,
-0.015143907628953457,
0.06115745007991791,
0.03872263431549072,
-0.07247884571552277,
0.004888486582785845,
-0.016991565003991127,
0.028380854055285454,
-0.036385178565979004,
0.05425618961453438,
0.08983579277992249,
-0.006524799391627312,
0.029000867158174515,
-0.0079692667350173,
0.0970468744635582,
-0.015201899223029613,
0.029102114960551262,
0.09183706343173981,
-0.036156684160232544,
-0.03767478093504906,
-0.015378113836050034,
-0.04842430725693703,
0.09593123197555542,
-0.030033666640520096,
-0.14386188983917236,
-0.08114312589168549,
0.025138072669506073,
-0.00432060519233346,
-0.010358651168644428,
-0.04615204408764839,
0.015116183087229729,
-0.04425833001732826,
0.08627991378307343,
0.055140312761068344,
-0.017858335748314857,
0.09356293082237244,
-0.004484011325985193,
-0.019715161994099617,
-0.007626884616911411,
-0.00668525043874979,
-0.028172791004180908,
0.01351387519389391,
-0.02172367461025715,
-0.0031805618200451136,
0.050827328115701675,
-0.045330580323934555,
-0.014347705990076065,
0.04343806207180023,
-0.08273252844810486,
-0.04501562938094139,
0.07827361673116684,
-0.04714571312069893,
0.08998546749353409,
-0.0822814553976059,
-0.026517078280448914,
0.009046374820172787,
-0.010982048697769642,
0.11908145993947983,
-0.048167016357183456,
0.046789079904556274,
-0.0438530407845974,
-0.06189341098070145,
0.018039049580693245,
-0.13056497275829315,
-0.012501344084739685,
-0.0019379385048523545,
-0.07830343395471573,
-0.06028546020388603,
-0.058793939650058746,
-0.020367180928587914,
-0.017345190048217773,
-3.058165919428271e-33,
-0.05653926730155945,
-0.005983429029583931,
-0.03373148664832115,
0.015487480908632278,
-0.05609555542469025,
-0.0006765160360373557,
0.033414389938116074,
0.0566687174141407,
-0.04596693813800812,
-0.05104074254631996,
0.04875299707055092,
-0.06018606200814247,
0.0010935376631096005,
-0.03735065832734108,
0.007185191847383976,
0.024457069113850594,
-0.014023921452462673,
-0.05044408142566681,
0.01795339025557041,
0.025877701118588448,
-0.004955383483320475,
0.08056741207838058,
-0.03872565180063248,
0.05727636441588402,
0.039064131677150726,
0.015015671029686928,
-0.06395450234413147,
0.06590345501899719,
0.10620847344398499,
0.02290262095630169,
-0.0713006779551506,
-0.026483558118343353,
-0.06819120794534683,
0.032961394637823105,
-0.029176315292716026,
0.04834911972284317,
0.00789290014654398,
0.07323269546031952,
-0.04797150194644928,
0.07377853989601135,
-0.009956044144928455,
0.011816000565886497,
-0.049861956387758255,
0.0912424847483635,
0.022109810262918472,
-0.04665297642350197,
0.024462008848786354,
-0.037937991321086884,
-0.05915984511375427,
0.042281970381736755,
0.001725691370666027,
-0.06442325562238693,
-0.029422737658023834,
-0.011341842822730541,
-0.03637569770216942,
0.011141089722514153,
0.014338945038616657,
0.0030592407565563917,
0.08014527708292007,
0.03226069360971451,
-0.05492992699146271,
-0.03929438814520836,
-0.015841972082853317,
-0.03361440449953079,
-0.06140751764178276,
0.034608252346515656,
0.010150207206606865,
0.06244954094290733,
0.07453828305006027,
0.04542337730526924,
0.07394377142190933,
0.041439030319452286,
0.029017483815550804,
-0.010208463296294212,
0.03680757060647011,
-0.0006270519224926829,
-0.007713704369962215,
-0.02738812379539013,
-0.020306188613176346,
-0.05387980490922928,
-0.01890713907778263,
-0.0005882537225261331,
0.019274944439530373,
0.014575534500181675,
0.05031852796673775,
0.014967869967222214,
0.014493392780423164,
0.06000729650259018,
-0.056549184024333954,
0.020954245701432228,
-0.07017206400632858,
0.05590870976448059,
0.03525235503911972,
0.06894844025373459,
-0.054105401039123535,
-4.714837942287886e-8,
-0.05579756200313568,
0.036179278045892715,
0.03660856932401657,
0.03416440635919571,
0.004330911207944155,
-0.04043756052851677,
0.008090894669294357,
0.07797113060951233,
-0.030808933079242706,
0.008199425414204597,
0.042024917900562286,
-0.016660591587424278,
-0.06697530299425125,
-0.018306534737348557,
-0.020012786611914635,
0.11240428686141968,
0.016644813120365143,
0.03579599782824516,
-0.023771366104483604,
0.018920332193374634,
0.01599559374153614,
0.008479777723550797,
-0.03446516767144203,
0.0099289296194911,
0.09182940423488617,
-0.09753590822219849,
-0.04579097405076027,
0.012632585130631924,
0.06972819566726685,
0.0182202085852623,
-0.022092735394835472,
-0.03166951239109039,
0.09387768059968948,
0.0018738856306299567,
0.03224793076515198,
0.15138573944568634,
-0.020803578197956085,
-0.0012106562498956919,
-0.03954776003956795,
-0.06867939233779907,
-0.0013658027164638042,
0.07003758102655411,
-0.004109962843358517,
0.0337466262280941,
0.040830984711647034,
0.014458740130066872,
0.04320733621716499,
-0.021252382546663284,
0.02220817282795906,
-0.022959725931286812,
0.034001924097537994,
0.04667649790644646,
0.00036960566649213433,
0.051468681544065475,
-0.0359191969037056,
0.024417929351329803,
0.031189151108264923,
-0.06940540671348572,
-0.009018144570291042,
0.09304571151733398,
-0.00959861185401678,
0.039524056017398834,
-0.11462884396314621,
-0.019149845466017723
] |
ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models | https://openreview.net/forum?id=2OegVbwvY2 | [
"Seonghwan Park",
"Jaehyeon Jeong",
"Yongjun Kim",
"Jaeho Lee",
"Namhoon Lee"
] | Poster | Recent studies have introduced various approaches for prompt-tuning black-box vision-language models, referred to as black-box prompt-tuning (BBPT). While BBPT has demonstrated considerable potential, it is often found that many existing methods require an excessive number of queries (i.e., function evaluations), which poses a significant challenge in real-world scenarios where the number of allowed queries is limited. To tackle this issue, we propose Zeroth-order Intrinsic-dimensional Prompt-tuning (ZIP), a novel approach that enables efficient and robust prompt optimization in a purely black-box setting. The key idea of ZIP is to reduce the problem dimensionality and the variance of zeroth-order gradient estimates, such that the training is done fast with far less queries. We achieve this by re-parameterizing prompts in low-rank representations and designing intrinsic-dimensional clipping of estimated gradients. We evaluate ZIP on 13+ vision-language tasks in standard benchmarks and show that it achieves an average improvement of approximately 6% in few-shot accuracy and 48% in query efficiency compared to the best-performing alternative BBPT methods, establishing a new state of the art. Our ablation analysis further shows that the proposed clipping mechanism is robust and nearly optimal, without the need to manually select the clipping threshold, matching the result of expensive hyperparameter search. | vision-language models, prompt-tuning, black-box optimization, zeroth-order optimization | We propose Zeroth-order Intrinsic-dimensional Prompt-tuning (ZIP), a method that reduces query demands in black-box prompt-tuning by optimizing in a lower-dimensional space with a robust clipping mechanism. | 939 | null | [
0.026197968050837517,
-0.03858819231390953,
-0.011922103352844715,
0.056917399168014526,
0.03911866247653961,
0.046515822410583496,
-0.015186145901679993,
-0.024291368201375008,
0.0035079296212643385,
-0.05726821348071098,
-0.09474391490221024,
-0.08174306899309158,
0.036192625761032104,
-0.006852432619780302,
-0.05819428339600563,
-0.009252103976905346,
0.1108805313706398,
0.04533147066831589,
-0.08695358037948608,
-0.0323021225631237,
-0.0124214394018054,
-0.010192773304879665,
0.058020636439323425,
-0.021092690527439117,
0.011821324005723,
-0.025494106113910675,
-0.04110312834382057,
-0.056961338967084885,
0.0083714434877038,
-0.011469356715679169,
-0.026214785873889923,
0.0060041374526917934,
0.05024329572916031,
0.04065675288438797,
-0.04979798197746277,
0.017707109451293945,
-0.02528800256550312,
0.01437936257570982,
-0.027814270928502083,
-0.05831444263458252,
-0.07983613759279251,
0.032604146748781204,
-0.0011076381197199225,
0.04620563983917236,
0.08085133135318756,
-0.037404615432024,
0.037592560052871704,
-0.05162078142166138,
-0.02138482592999935,
-0.012254505418241024,
-0.09721074998378754,
-0.038199976086616516,
-0.048995405435562134,
0.07269351184368134,
-0.043307047337293625,
0.048814479261636734,
0.018940865993499756,
0.02058332972228527,
0.030967898666858673,
0.005548775661736727,
-0.11035802960395813,
-0.12969280779361725,
0.01938926801085472,
0.02105833776295185,
0.012481183744966984,
-0.007494983728975058,
0.03440876305103302,
-0.046275604516267776,
-0.021386394277215004,
0.013971840031445026,
0.011946330778300762,
0.11780206114053726,
-0.01890721544623375,
0.044693320989608765,
0.007533719297498465,
0.10049892216920853,
0.030571941286325455,
-0.08444487303495407,
0.04362824931740761,
-0.019655289128422737,
-0.0021025766618549824,
-0.03908471763134003,
0.028029220178723335,
0.05196303501725197,
0.12689723074436188,
0.0022595259360969067,
-0.0022357702255249023,
-0.00824283529073,
0.04412242770195007,
0.017596520483493805,
-0.059067849069833755,
0.004326268099248409,
-0.03352256119251251,
-0.006846849340945482,
-0.040351107716560364,
0.028045443817973137,
0.004760386887937784,
-0.13231991231441498,
-0.04258715361356735,
0.0997004434466362,
-0.02949422411620617,
0.0069074975326657295,
0.04477037116885185,
-0.06399879604578018,
-0.002115790266543627,
-0.008693784475326538,
0.06768748164176941,
0.04607890918850899,
0.0452236570417881,
-0.07227005064487457,
0.0377718023955822,
0.02164328843355179,
0.047932982444763184,
-0.011429917067289352,
-0.004025311209261417,
0.03371613845229149,
-0.0080302394926548,
-0.04625653475522995,
0.0419250912964344,
0.05635329335927963,
0.020889775827527046,
0.01019822247326374,
-0.046197496354579926,
0.04340122640132904,
0.04159679263830185,
-0.004262094385921955,
-0.005193376448005438,
1.4079417006239234e-34,
0.04288528859615326,
0.06216970086097717,
0.07437741756439209,
0.008180209435522556,
0.11117641627788544,
0.028265543282032013,
0.08502356708049774,
0.0432298369705677,
-0.05983426794409752,
0.026600796729326248,
-0.02553027868270874,
-0.035596106201410294,
-0.0617687962949276,
0.08180378377437592,
0.06837549805641174,
0.005725228227674961,
0.00958881713449955,
0.06194354221224785,
-0.07785532623529434,
0.03467637300491333,
0.06059318035840988,
-0.032128870487213135,
-0.008380427956581116,
-0.00967765599489212,
0.08723747730255127,
-0.02577146142721176,
0.0528942234814167,
-0.0230867937207222,
-0.03763808682560921,
0.006889460142701864,
-0.12025139480829239,
-0.05199341103434563,
-0.014866279438138008,
-0.006368244532495737,
0.05584445968270302,
-0.039856359362602234,
-0.029914986342191696,
0.012610569596290588,
0.0009988525416702032,
0.03917441889643669,
-0.006977546960115433,
0.026902804151177406,
0.012852294370532036,
-0.028896398842334747,
-0.09075701981782913,
-0.06492312252521515,
0.0013057052856311202,
0.036397743970155716,
-0.0932716429233551,
-0.006355151534080505,
0.01579253375530243,
0.03693552315235138,
-0.039106693118810654,
-0.019272778183221817,
-0.0007971361628733575,
-0.05301111564040184,
0.08657519519329071,
0.023287693038582802,
-0.005798811092972755,
0.042613815516233444,
0.006501539144665003,
0.013798347674310207,
-0.013467195443809032,
0.024048004299402237,
0.002575581893324852,
-0.033535052090883255,
-0.006410789210349321,
0.034366175532341,
0.06039838492870331,
0.0037952319253236055,
-0.03766018897294998,
0.0018738129874691367,
-0.006291206926107407,
-0.06509947776794434,
0.0016262774588540196,
0.03211590647697449,
0.05737237632274628,
-0.07883132249116898,
0.1070869192481041,
0.026846716180443764,
0.03859160840511322,
-0.008617249317467213,
0.008438088931143284,
-0.026964975520968437,
-0.014086027629673481,
-0.09435306489467621,
0.07298208028078079,
-0.013085625134408474,
-0.030982818454504013,
-0.09085460752248764,
-0.03768881410360336,
-0.016537291929125786,
-0.056078191846609116,
-0.050273820757865906,
0.008804012089967728,
3.822973263046716e-34,
0.004029730334877968,
0.0038069426082074642,
-0.0355902723968029,
0.04713936150074005,
-0.07114249467849731,
-0.04476448521018028,
0.02977536804974079,
0.010378262028098106,
0.017200184985995293,
-0.07657621800899506,
-0.022851457819342613,
-0.02066844329237938,
0.020679710432887077,
0.05918239429593086,
0.0654536560177803,
0.04782697185873985,
-0.05550648272037506,
0.040364425629377365,
0.005454636178910732,
0.03391557186841965,
-0.0074639529921114445,
0.09951119869947433,
-0.12156891822814941,
0.041285645216703415,
-0.10281773656606674,
-0.02190864086151123,
-0.0099010756239295,
0.05338571220636368,
-0.01412753015756607,
0.007690432481467724,
-0.014512917026877403,
0.0220336951315403,
-0.046675216406583786,
0.04890265688300133,
-0.05960327386856079,
0.0624915212392807,
0.03507949411869049,
-0.04195452854037285,
-0.001678174245171249,
0.15765775740146637,
0.09799794852733612,
-0.05886104330420494,
-0.05872602388262749,
-0.011666485108435154,
-0.05627592280507088,
-0.04116881638765335,
-0.04158398509025574,
0.008886629715561867,
-0.014494297094643116,
0.042757656425237656,
-0.1310470998287201,
0.08097849786281586,
-0.056287407875061035,
0.031078346073627472,
-0.060652974992990494,
-0.11976794898509979,
-0.11017359048128128,
0.004798318259418011,
0.14168311655521393,
-0.024960404261946678,
0.005046993959695101,
0.04687712714076042,
0.005574239417910576,
-0.09508578479290009,
0.0006272854516282678,
-0.03765860199928284,
-0.017870353534817696,
0.0028467876836657524,
0.013322519138455391,
-0.043187227100133896,
0.01720491237938404,
-0.06187625601887703,
0.03339235112071037,
0.05311077833175659,
-0.02262146770954132,
0.035454705357551575,
0.03363322466611862,
-0.013559291139245033,
0.027362141758203506,
0.010789845138788223,
0.05622829124331474,
0.07960988581180573,
0.08110928535461426,
0.08267926424741745,
0.04431368410587311,
0.10652690380811691,
0.07812386006116867,
0.1122875064611435,
0.047014158219099045,
-0.02926487661898136,
-0.046415094286203384,
-0.013687274418771267,
0.035511162132024765,
0.08898754417896271,
-0.06934187561273575,
-4.073592307918261e-8,
-0.12013626098632812,
0.07414952665567398,
0.007408943958580494,
0.05742757394909859,
0.044236574321985245,
-0.03658435121178627,
-0.021462012082338333,
0.07495231926441193,
0.009589061141014099,
-0.03961799293756485,
0.04951448366045952,
0.005386373959481716,
-0.041510649025440216,
-0.04992847889661789,
-0.015917034819722176,
0.06722576916217804,
-0.021252354606986046,
-0.010855654254555702,
-0.01204331498593092,
-0.017802581191062927,
-0.05963006615638733,
0.05790678411722183,
0.03307699039578438,
-0.04068935662508011,
0.006350155454128981,
-0.07024667412042618,
-0.02695336565375328,
0.08668813109397888,
0.12483612447977066,
-0.02133762091398239,
-0.02711516246199608,
0.032047778367996216,
0.03452073410153389,
-0.051655035465955734,
0.04649188742041588,
0.06539524346590042,
-0.012321989051997662,
0.014494764618575573,
-0.004620391875505447,
-0.05182933062314987,
0.036501482129096985,
-0.0032294169068336487,
-0.02386000193655491,
-0.02124154567718506,
-0.03376040235161781,
-0.020950106903910637,
0.020155325531959534,
-0.13038650155067444,
0.030677959322929382,
-0.04510638490319252,
0.02923763170838356,
-0.05492829158902168,
-0.05537092313170433,
0.09790743142366409,
0.021423744037747383,
0.0048867701552808285,
-0.012104768306016922,
-0.017111506313085556,
-0.03021230734884739,
0.10740712285041809,
-0.01157840434461832,
0.05240292847156525,
-0.05951286107301712,
-0.033664800226688385
] |
IFORMER: INTEGRATING CONVNET AND TRANSFORMER FOR MOBILE APPLICATION | https://openreview.net/forum?id=4ytHislqDS | [
"Chuanyang Zheng"
] | Poster | We present a new family of mobile hybrid vision networks, called iFormer, with a focus on optimizing latency and accuracy on mobile applications. iFormer effectively integrates the fast local representation capacity of convolution with the efficient global modeling ability of self-attention. The local interactions are derived from transforming a standard convolutional network, \textit{i.e.}, ConvNeXt, to design a more lightweight mobile network. Our newly introduced mobile modulation attention removes memory-intensive operations in MHA and employs an efficient modulation mechanism to boost dynamic global representational capacity. We conduct comprehensive experiments demonstrating that iFormer outperforms existing lightweight networks across various tasks. Notably, iFormer achieves an impressive Top-1 accuracy of 80.4% on ImageNet-1k with a latency of only 1.10 ms on an iPhone 13, surpassing the recently proposed MobileNetV4 under similar latency constraints. Additionally, our method shows significant improvements in downstream tasks, including COCO object detection, instance segmentation, and ADE20k semantic segmentation, while still maintaining low latency on mobile devices for high-resolution inputs in these scenarios. Code and models are available at: https://github.com/ChuanyangZheng/iFormer. | Lightweight Networks, Efficient Networks, Vision Transformers, Classification | null | 938 | null | [
-0.014353666454553604,
-0.0009589125402271748,
0.03755492717027664,
0.02885330282151699,
0.08862140029668808,
-0.05325033888220787,
-0.03529931604862213,
-0.02310570888221264,
0.04334965720772743,
-0.03717821463942528,
0.00034213176695629954,
-0.1014845222234726,
0.0003072675608564168,
0.08077345043420792,
-0.01949240453541279,
0.001153340796008706,
0.11804395169019699,
0.029321035370230675,
-0.14413020014762878,
-0.02521355450153351,
0.007037079893052578,
-0.02420216053724289,
0.05610225349664688,
0.004254586528986692,
0.11714214086532593,
-0.032498594373464584,
-0.062418561428785324,
-0.11972306668758392,
0.022207241505384445,
-0.03372407704591751,
-0.0014810821739956737,
0.029129603877663612,
-0.035787444561719894,
0.10233907401561737,
-0.07545654475688934,
0.01506910752505064,
-0.01296838279813528,
-0.004149076994508505,
-0.04518279805779457,
-0.09773179143667221,
-0.021213212981820107,
-0.0032563405111432076,
0.026713915169239044,
0.07024016231298447,
0.011598550714552402,
0.06826288253068924,
0.05192399024963379,
0.015562348067760468,
0.057379841804504395,
-0.03983753174543381,
0.026330500841140747,
0.0261221993714571,
-0.006035150028765202,
0.03002779558300972,
-0.05846983939409256,
0.05570816993713379,
-0.004120852332562208,
0.01801786758005619,
0.06326209008693695,
0.08625099062919617,
0.020131856203079224,
0.022335687652230263,
0.043015558272600174,
0.016900086775422096,
-0.07768318802118301,
-0.09728691726922989,
0.07567992806434631,
0.0156497061252594,
0.0021953138057142496,
-0.06006346642971039,
0.047368958592414856,
0.05204881355166435,
0.025851069018244743,
0.01882467046380043,
0.04429265856742859,
0.018357791006565094,
0.07577764987945557,
-0.030166855081915855,
0.026376960799098015,
-0.037499357014894485,
0.06951349228620529,
-0.044817354530096054,
0.06592488288879395,
0.007986478507518768,
0.1334974467754364,
-0.03488010913133621,
-0.05912729352712631,
0.08240120857954025,
-0.029077477753162384,
-0.021537944674491882,
-0.07137659192085266,
-0.04027589410543442,
-0.010105671361088753,
-0.04100834205746651,
0.046166084706783295,
-0.00036639769678004086,
-0.08373264223337173,
-0.09558357298374176,
-0.07257459312677383,
0.09455376118421555,
-0.02591441385447979,
-0.048211876302957535,
0.04716348648071289,
-0.028391920030117035,
0.04222577065229416,
0.0035677540581673384,
0.10567817091941833,
0.04487080127000809,
0.03894443064928055,
0.03618520498275757,
0.03945586457848549,
0.012951618060469627,
0.06095544621348381,
-0.0458916611969471,
0.04156625270843506,
0.044495128095149994,
-0.021595366299152374,
0.015279141254723072,
0.05868542194366455,
-0.017431920394301414,
-0.07622344046831131,
-0.029421543702483177,
-0.07604158669710159,
-0.0012664843816310167,
0.058034975081682205,
-0.008227074518799782,
-0.05728049948811531,
2.4210202469627513e-33,
-0.0019247605232521892,
0.02478807419538498,
-0.003377336310222745,
0.02511177957057953,
0.031505558639764786,
-0.05744999274611473,
0.06337343901395798,
0.06577889621257782,
-0.023783007636666298,
-0.05971337854862213,
0.011324193328619003,
-0.0311780646443367,
-0.06275614351034164,
0.11554525792598724,
0.03676668554544449,
-0.05370502546429634,
-0.02083488553762436,
-0.03858530893921852,
-0.038174837827682495,
-0.038815781474113464,
-0.07956275343894958,
-0.05952495336532593,
-0.024177366867661476,
-0.002220702590420842,
0.02571146935224533,
-0.01854715310037136,
0.03798685595393181,
-0.04975726082921028,
0.05318181589245796,
-0.03488146513700485,
-0.06693948805332184,
0.023707566782832146,
0.02476264350116253,
-0.011098586954176426,
0.010567431338131428,
-0.03548237308859825,
-0.03210430219769478,
-0.045690007507801056,
0.0013262967113405466,
-0.011854744516313076,
-0.042486660182476044,
0.07459210604429245,
-0.07307430356740952,
-0.07034169137477875,
-0.09704659134149551,
-0.043197739869356155,
0.003872727742418647,
0.10507664829492569,
-0.1128978580236435,
0.026621537283062935,
0.10381418466567993,
0.003494007745757699,
-0.06738688796758652,
-0.06935050338506699,
-0.02275189757347107,
-0.02668212354183197,
0.07172613590955734,
0.08104324340820312,
0.04144898056983948,
0.03896316513419151,
-0.03853828087449074,
0.007158889900892973,
-0.10375728458166122,
0.15851333737373352,
-0.014758859761059284,
-0.041054774075746536,
0.028530580922961235,
0.07381738722324371,
0.013095651753246784,
0.05807957798242569,
-0.052928369492292404,
0.056137509644031525,
0.02019299939274788,
-0.06613591313362122,
0.0602930448949337,
0.0992927998304367,
-0.006520893890410662,
-0.05117790773510933,
-0.04444913566112518,
0.03077089972794056,
-0.004001129884272814,
0.040879927575588226,
0.016794126480817795,
0.0570903904736042,
-0.01717349700629711,
-0.003179132705554366,
0.04760683327913284,
-0.1345631629228592,
0.009155209176242352,
0.10296084731817245,
-0.023630192503333092,
-0.06367306411266327,
0.0011483973357826471,
-0.014521258883178234,
-0.03398744389414787,
-1.4043408767373167e-33,
-0.004234574735164642,
0.07353182137012482,
-0.03635570779442787,
0.014998904429376125,
-0.015696825459599495,
-0.028036629781126976,
0.042727746069431305,
0.017812835052609444,
-0.06862343847751617,
-0.13135561347007751,
-0.013333585113286972,
-0.018720490857958794,
0.03828323259949684,
0.018831541761755943,
0.010993131436407566,
-0.02797858789563179,
-0.038365401327610016,
-0.020235635340213776,
-0.053517185151576996,
0.006759649142622948,
0.001057895366102457,
0.1465485394001007,
0.0011110244086012244,
-0.01644417829811573,
-0.06716207414865494,
0.023035245016217232,
-0.07747761160135269,
0.030473724007606506,
-0.00003100433241343126,
-0.06431195139884949,
-0.015591529197990894,
0.012887734919786453,
0.06224128231406212,
-0.07255218178033829,
0.022389667108654976,
0.0857996717095375,
0.020579315721988678,
-0.01800197921693325,
-0.021032825112342834,
0.013849761337041855,
0.09676520526409149,
-0.03059621900320053,
-0.04866451025009155,
0.0016979679930955172,
-0.06128014624118805,
-0.01336180791258812,
-0.06060943379998207,
0.030875708907842636,
-0.012939122505486012,
0.05642208829522133,
0.006847213488072157,
0.019728273153305054,
-0.1131332665681839,
-0.003363323165103793,
-0.011995826847851276,
0.01966528408229351,
-0.049508269876241684,
-0.038103651255369186,
0.03694169968366623,
-0.034199126064777374,
-0.008473635651171207,
-0.09233928471803665,
-0.02821126952767372,
-0.08357150107622147,
-0.03626352176070213,
-0.0385831817984581,
-0.011838087812066078,
-0.027851203456521034,
0.07180849462747574,
0.08674131333827972,
0.04439835995435715,
-0.05193346366286278,
0.018678732216358185,
0.04983890429139137,
-0.01071842573583126,
-0.03847266361117363,
0.056825388222932816,
-0.019578494131565094,
0.004732324276119471,
-0.006377772428095341,
-0.010578172281384468,
0.0670071542263031,
0.02906501106917858,
0.0012764312559738755,
0.09448793530464172,
0.07430745661258698,
0.05213846638798714,
0.024972109124064445,
0.08101626485586166,
0.06900513917207718,
-0.06063346937298775,
0.038344599306583405,
0.0038831522688269615,
0.04628737270832062,
-0.05650629475712776,
-3.8929638179752146e-8,
-0.029208946973085403,
0.04543878883123398,
-0.01687442511320114,
-0.04935304448008537,
-0.07215867191553116,
-0.06007453054189682,
0.03533558174967766,
0.06895710527896881,
0.01850890927016735,
0.04841288924217224,
0.0702475756406784,
-0.024337749928236008,
-0.060022562742233276,
-0.03156778961420059,
0.0386161282658577,
0.022943686693906784,
0.07967034727334976,
-0.043944790959358215,
0.01627226546406746,
-0.004729392006993294,
-0.018186131492257118,
0.03373093530535698,
-0.01295794453471899,
-0.027659155428409576,
0.010368067771196365,
-0.032836105674505234,
-0.04731954634189606,
0.06414439529180527,
0.03994103893637657,
-0.030590318143367767,
0.004781941417604685,
0.061750661581754684,
-0.01181801874190569,
-0.03546224161982536,
-0.015576859004795551,
0.08992700278759003,
-0.010516592301428318,
0.03852773830294609,
0.06871321052312851,
-0.014878610149025917,
0.015530284494161606,
0.01191117987036705,
-0.0187631007283926,
0.02166452631354332,
-0.06117917597293854,
-0.03936600685119629,
0.0015331238973885775,
-0.07649542391300201,
0.009106605313718319,
-0.045534275472164154,
-0.026706835255026817,
-0.011908021755516529,
-0.02097293548285961,
0.03772632032632828,
0.03495267778635025,
-0.02992781065404415,
0.07737812399864197,
-0.03053624927997589,
-0.005348639562726021,
0.07107967138290405,
0.009100114926695824,
0.02023809403181076,
-0.0635206326842308,
-0.0491013340651989
] |
kNN Attention Demystified: A Theoretical Exploration for Scalable Transformers | https://openreview.net/forum?id=49v8meXjHS | [
"Themistoklis Haris"
] | Poster | Despite their power, Transformers face challenges with long sequences due to the quadratic complexity of self-attention. To address this limitation, methods like $k$-Nearest-Neighbor ($k$NN) attention have been introduced [Roy et al., 2017], enabling each token to attend to only its $k$ closest tokens. While $k$NN attention has shown empirical success in making Transformers more efficient, its exact approximation guarantees have not been theoretically analyzed. In this work, we establish a theoretical framework for $k$NN attention, reformulating self-attention as expectations over softmax distributions and leveraging lazy Gumbel sampling [Mussmann et al., 2017] with $k$NN indices for efficient approximation. Building on this framework, we also propose novel sub-quadratic algorithms that approximate self-attention gradients by leveraging efficient sampling techniques, such as Markov Chain-based estimation. Finally, we demonstrate the practical effectiveness of these algorithms through empirical experiments, showcasing their benefits in both training and inference. | efficient transformers, self-attention mechanism, sublinear algorithms, sampling, k-nearest neighbors | This paper proposes a theoretical framework for kNN attention and develops novel algorithms for sub-quadratic attention gradient estimation in Transformers. | 932 | null | [
-0.08820484578609467,
-0.03734968602657318,
0.04100986197590828,
-0.011614923365414143,
0.02509305067360401,
0.023252010345458984,
0.039694108068943024,
-0.006646447815001011,
0.08127789199352264,
-0.0540165901184082,
-0.07300381362438202,
-0.004884117748588324,
0.0442945659160614,
0.00683362502604723,
-0.02599700726568699,
0.04444044455885887,
0.09314222633838654,
0.09569311141967773,
-0.05327165126800537,
-0.06296946108341217,
0.04379181191325188,
0.008596298284828663,
0.08741755038499832,
-0.057549476623535156,
0.04718963801860809,
-0.04885893315076828,
0.03253883123397827,
-0.06689808517694473,
-0.0036628900561481714,
0.025747578591108322,
0.003044772194698453,
-0.02381051518023014,
0.05138595402240753,
-0.029115203768014908,
-0.010842119343578815,
0.03215879946947098,
-0.04858734458684921,
0.04816540330648422,
0.024379350244998932,
0.017945585772395134,
0.050659939646720886,
-0.0013718982227146626,
0.008359265513718128,
0.05079979822039604,
0.06955519318580627,
-0.032736312597990036,
0.054955143481492996,
-0.03556259348988533,
0.005327512510120869,
-0.008579813875257969,
-0.012761141173541546,
-0.015877729281783104,
-0.049965545535087585,
0.10534252226352692,
0.033999428153038025,
-0.013457872904837132,
-0.04388541355729103,
0.009774243459105492,
0.06329100579023361,
0.06695932149887085,
0.009712531231343746,
-0.11133959144353867,
0.027866296470165253,
-0.08699379861354828,
-0.04258348047733307,
-0.020892826840281487,
0.09823188930749893,
0.024773575365543365,
-0.06032106280326843,
0.0300433486700058,
-0.026904435828328133,
0.056465040892362595,
-0.08461575955152512,
0.035058099776506424,
0.09483134001493454,
-0.004442921839654446,
0.03861716017127037,
-0.07978334277868271,
0.04090903699398041,
-0.06956330686807632,
-0.019199779257178307,
0.0025752182118594646,
0.05111057311296463,
-0.018619174137711525,
0.07932683825492859,
-0.01780594512820244,
0.01177431270480156,
0.04985042288899422,
0.041066210716962814,
-0.0015242639929056168,
-0.08224252611398697,
-0.07954004406929016,
-0.01670202985405922,
-0.048697881400585175,
0.019313646480441093,
-0.019313234835863113,
0.018384354189038277,
-0.018632078543305397,
-0.017964201048016548,
0.108999103307724,
0.032866768538951874,
0.045004308223724365,
0.03421499952673912,
-0.04984322935342789,
0.0046813529916107655,
-0.07511207461357117,
0.014008952304720879,
0.08886474370956421,
0.012430417351424694,
-0.08499716222286224,
0.0037268956657499075,
0.02676153928041458,
0.009223001077771187,
-0.01589152030646801,
0.060947444289922714,
-0.06215101107954979,
0.016917884349822998,
0.00717620924115181,
-0.033680811524391174,
0.038076601922512054,
-0.001920232200063765,
0.010280796326696873,
-0.005953875370323658,
0.05868733301758766,
0.09049887955188751,
0.026992792263627052,
-0.015549096278846264,
3.302559092816894e-33,
0.03780898451805115,
0.09090592712163925,
0.09750735014677048,
-0.031309597194194794,
-0.014289327897131443,
-0.02235565334558487,
0.08433494716882706,
0.027865689247846603,
-0.005010182503610849,
0.032429713755846024,
0.012017758563160896,
0.04401366412639618,
-0.021990342065691948,
0.04338519275188446,
0.01186017133295536,
0.015530180186033249,
-0.031101500615477562,
-0.003141158726066351,
-0.013985726982355118,
-0.013753358274698257,
0.05514262989163399,
-0.009867597371339798,
-0.003752142423763871,
-0.027474820613861084,
-0.03776833042502403,
-0.06419754773378372,
0.03810809180140495,
-0.06801562011241913,
0.027344102039933205,
0.0017627669731155038,
-0.10022199153900146,
0.0343981496989727,
-0.03336547315120697,
-0.030352802947163582,
0.021654285490512848,
-0.07644487917423248,
-0.004455152433365583,
-0.0238223634660244,
-0.013033518567681313,
-0.054757241159677505,
-0.0413084551692009,
0.08032747358083725,
0.01796332374215126,
-0.008454245515167713,
-0.10855529457330704,
-0.04706757515668869,
0.026332302019000053,
-0.0011455741478130221,
0.04449816420674324,
-0.08187281340360641,
0.03424287959933281,
-0.013452345505356789,
-0.04341737926006317,
0.008375147357583046,
0.03714967519044876,
0.015994491055607796,
0.05804520100355148,
0.07680001854896545,
0.01968751661479473,
0.07173530757427216,
0.036515362560749054,
-0.03630436584353447,
-0.033321138471364975,
0.08223170787096024,
0.021351968869566917,
0.06348402798175812,
-0.030724506825208664,
0.021790046244859695,
0.07838673144578934,
0.026185324415564537,
0.041560109704732895,
0.02696456015110016,
-0.06214408576488495,
-0.09453818947076797,
0.06514149904251099,
0.022105801850557327,
0.07589371502399445,
-0.0702923834323883,
-0.05379744991660118,
0.016715429723262787,
0.007170574273914099,
-0.02069879323244095,
-0.040346577763557434,
-0.07952351123094559,
-0.06346911936998367,
-0.03766036778688431,
0.045227982103824615,
-0.08576525747776031,
-0.04826795682311058,
-0.04333478957414627,
-0.053016144782304764,
-0.006785387173295021,
-0.07573758065700531,
-0.012216078117489815,
-0.0743572860956192,
-1.5442930300943104e-33,
0.018294503912329674,
0.019600164145231247,
0.0215609073638916,
0.042282357811927795,
0.015297478064894676,
-0.02077208086848259,
-0.03491359204053879,
0.002055979100987315,
-0.02710147015750408,
0.023911429569125175,
-0.07390174269676208,
-0.05658266320824623,
0.07526182383298874,
-0.03928592428565025,
0.06937851011753082,
0.04340445250272751,
-0.010844610631465912,
0.058037418872117996,
-0.07043605297803879,
-0.0177969578653574,
-0.03541411831974983,
0.1069377139210701,
-0.02418912947177887,
0.025517944246530533,
-0.06745205819606781,
0.03293658420443535,
-0.05756368115544319,
0.09058376401662827,
-0.010106390342116356,
-0.10232717543840408,
-0.11037669330835342,
-0.04339157044887543,
-0.06409510970115662,
0.008728585205972195,
-0.08724482357501984,
0.05260874703526497,
0.03287586569786072,
0.03639233484864235,
-0.07187783718109131,
0.0764048621058464,
0.047045934945344925,
-0.03321037068963051,
-0.030259719118475914,
0.05527496710419655,
-0.06793659180402756,
-0.06360438466072083,
-0.029278993606567383,
-0.026030341163277626,
-0.014198996126651764,
0.05421140417456627,
-0.011205587536096573,
0.04583260416984558,
-0.08135485649108887,
0.07751242816448212,
-0.0763663500547409,
-0.052444230765104294,
-0.07235309481620789,
0.06300382316112518,
0.1567627191543579,
-0.03143591061234474,
-0.0537283755838871,
-0.03389061242341995,
-0.006013740785419941,
-0.016457371413707733,
-0.03358105570077896,
-0.04570670425891876,
0.0017363724764436483,
0.0007469537667930126,
0.07970865070819855,
0.02756117656826973,
-0.014651124365627766,
0.02108979970216751,
0.06316273659467697,
-0.017710205167531967,
-0.05485650151968002,
0.008632263168692589,
0.028885385021567345,
-0.09692796319723129,
-0.091917484998703,
-0.07301650196313858,
0.037774305790662766,
0.027609823271632195,
-0.00445500249043107,
-0.01387881301343441,
0.030077386647462845,
0.07942673563957214,
0.19632503390312195,
0.014877762645483017,
0.04551343619823456,
-0.051115233451128006,
-0.020235590636730194,
0.007333656772971153,
-0.036257896572351456,
0.04670089855790138,
-0.08250928670167923,
-4.2193104121679426e-8,
-0.06700282543897629,
0.03492671623826027,
-0.0504499226808548,
0.033994536846876144,
0.11743772774934769,
-0.04664549231529236,
0.02931523136794567,
0.11606535315513611,
-0.08461444824934006,
-0.010942699387669563,
0.137392058968544,
-0.029460212215781212,
0.00908738188445568,
-0.0014535699738189578,
-0.004041749518364668,
0.016254933550953865,
0.037186842411756516,
-0.0015864629531279206,
-0.0036832517944276333,
-0.015102924779057503,
-0.037802498787641525,
0.0760543942451477,
0.043618060648441315,
0.013890697620809078,
0.00044964178232476115,
-0.010384761728346348,
-0.0022870637476444244,
0.017232593148946762,
0.06964815407991409,
-0.00590886315330863,
-0.02222432754933834,
0.04261072352528572,
0.0011779923224821687,
-0.10165131092071533,
-0.014624867588281631,
0.07839310169219971,
0.006154918111860752,
0.027768202126026154,
0.00783510971814394,
-0.013725506141781807,
0.0323861688375473,
-0.0031101927161216736,
-0.06274133920669556,
0.05622924119234085,
-0.0032986633013933897,
-0.03641698881983757,
-0.04411427304148674,
-0.08911771327257156,
0.06087854877114296,
-0.05650237575173378,
0.1009259894490242,
-0.08975627273321152,
-0.02089228853583336,
-0.019346311688423157,
0.012381317093968391,
0.006801795680075884,
-0.002183719305321574,
-0.026790158823132515,
0.00882377102971077,
0.06498473882675171,
0.0593516007065773,
0.031238308176398277,
-0.03660617396235466,
0.009911659173667431
] |
Diffusion Feedback Helps CLIP See Better | https://openreview.net/forum?id=tLFWU6izoA | [
"Wenxuan Wang",
"Quan Sun",
"Fan Zhang",
"Yepeng Tang",
"Jing Liu",
"Xinlong Wang"
] | Poster | Contrastive Language-Image Pre-training (CLIP), which excels at abstracting open-world representations across domains and modalities, has become a foundation for a variety of vision and multimodal tasks. However, recent studies reveal that CLIP has severe visual shortcomings, such as which can hardly distinguish orientation, quantity, color, structure, etc. These visual shortcomings also limit the perception capabilities of multimodal large language models (MLLMs) built on CLIP. The main reason could be that the image-text pairs used to train CLIP are inherently biased, due to the lack of the distinctiveness of the text and the diversity of images. In this work, we present a simple post-training approach for CLIP models, which largely overcomes its visual shortcomings via a self-supervised diffusion process. We introduce DIVA, which uses the DIffusion model as a Visual Assistant for CLIP. Specifically, DIVA leverages generative feedback from text-to-image diffusion models to optimize CLIP representations, with only images (without corresponding text). We demonstrate that DIVA improves CLIP's performance on the challenging MMVP-VLM benchmark which assesses fine-grained visual abilities to a large extent (e.g., 3-7%), and enhances the performance of MLLMs and vision models on multimodal understanding and segmentation tasks. Extensive evaluation on 29 image classification and retrieval benchmarks confirms that our framework preserves CLIP's strong zero-shot capabilities. The code is publicly available at https://github.com/baaivision/DIVA. | CLIP Model, Diffusion Model, Generative Feedback, Representation Learning | In this work, we present a simple post-training approach for CLIP models, which largely overcomes its visual shortcomings via a self-supervised diffusion process. | 918 | 2407.20171 | [
0.025305140763521194,
-0.1297871619462967,
0.027886206284165382,
0.02459181286394596,
0.09776086360216141,
-0.0012839330593124032,
-0.05419870838522911,
-0.02382645197212696,
0.08170722424983978,
-0.09380564838647842,
0.0005351709551177919,
-0.07725256681442261,
-0.016965514048933983,
0.0685863345861435,
-0.01607506535947323,
0.007562822662293911,
0.13795386254787445,
0.07656582444906235,
-0.10415125638246536,
-0.05624363198876381,
0.027422407642006874,
0.017646795138716698,
0.03793958202004433,
-0.004167996346950531,
0.064562126994133,
0.0667693093419075,
-0.019160542637109756,
-0.03856740519404411,
0.03152182698249817,
-0.03680167719721794,
0.01220647245645523,
-0.06750857830047607,
0.04104594141244888,
0.06522977352142334,
-0.0646110251545906,
0.08047714829444885,
-0.039329007267951965,
0.027920609340071678,
0.007041463162750006,
-0.058013711124658585,
-0.032410189509391785,
-0.01625966653227806,
0.020178627222776413,
-0.01967488043010235,
0.04982943460345268,
-0.030162449926137924,
-0.024206651374697685,
-0.007524378597736359,
0.023361172527074814,
0.020504239946603775,
-0.06620204448699951,
-0.05288396775722504,
-0.02994869090616703,
0.09641110897064209,
-0.0575965940952301,
0.0019525340758264065,
0.09566818922758102,
0.02310151793062687,
0.04108835756778717,
-0.0330689400434494,
0.0036839672829955816,
-0.03800354152917862,
-0.009957447648048401,
-0.004246285650879145,
0.001990271033719182,
-0.030767349526286125,
0.08059277385473251,
0.014459158293902874,
-0.07933016866445541,
-0.037198662757873535,
0.014751597307622433,
0.02673528715968132,
-0.027764318510890007,
0.054121777415275574,
0.03110555186867714,
0.058205436915159225,
0.033928707242012024,
-0.08819790184497833,
0.020020047202706337,
-0.07525672763586044,
0.09758508205413818,
0.005039996933192015,
0.10406497120857239,
0.009631912223994732,
0.07273383438587189,
0.006869758479297161,
-0.016006657853722572,
0.04942713677883148,
0.0007342497119680047,
-0.0035111214965581894,
-0.04864160716533661,
0.024531539529561996,
-0.020362921059131622,
-0.011424915865063667,
0.03650377318263054,
0.0018817696254700422,
-0.006548336707055569,
-0.06541173160076141,
0.027809370309114456,
0.10072144120931625,
0.02381264790892601,
0.03192152455449104,
-0.019803151488304138,
0.011774279177188873,
0.014134307391941547,
-0.021345917135477066,
0.08582235127687454,
-0.01942364312708378,
0.01820841245353222,
-0.04657837748527527,
0.09374188631772995,
0.01508354302495718,
-0.042869649827480316,
-0.01075368095189333,
0.07097071409225464,
-0.06157262250781059,
0.0052399528212845325,
-0.01782318577170372,
0.0444263257086277,
0.0333297923207283,
-0.028733940795063972,
-0.0047398959286510944,
-0.057314902544021606,
-0.008051947690546513,
0.0732860341668129,
-0.0033458282705396414,
0.004414339549839497,
-8.526389498971921e-34,
0.025281256064772606,
0.014058011583983898,
0.06153020262718201,
0.04670038819313049,
0.06567832082509995,
-0.03656234219670296,
-0.07298099994659424,
-0.04992872476577759,
-0.05739942193031311,
-0.10800402611494064,
0.00999507587403059,
0.015128032304346561,
-0.05936822667717934,
0.10041338950395584,
0.031147632747888565,
-0.01876259595155716,
-0.015971427783370018,
0.02231096476316452,
0.009552237577736378,
0.04079829528927803,
-0.004747480619698763,
0.01505809836089611,
0.008736994117498398,
0.024164635688066483,
0.034045372158288956,
-0.03536968678236008,
0.09917086362838745,
-0.09670361131429672,
-0.012754510156810284,
0.026785381138324738,
-0.15351347625255585,
-0.010924775153398514,
0.015476172789931297,
-0.018213072791695595,
0.04065686836838722,
-0.005896993912756443,
-0.03871741145849228,
-0.02103748545050621,
0.035360414534807205,
-0.01689075119793415,
-0.010831423103809357,
0.045552078634500504,
-0.025305725634098053,
-0.013189785182476044,
-0.13465386629104614,
-0.04014098271727562,
0.017100438475608826,
0.05294939503073692,
-0.10798481106758118,
-0.007020655553787947,
0.03740663826465607,
-0.03077594004571438,
-0.018028512597084045,
-0.07207343727350235,
-0.030906042084097862,
-0.029889632016420364,
0.0732564628124237,
0.010136016644537449,
0.06491624563932419,
-0.0014883201802149415,
0.020246082916855812,
0.10555350035429001,
-0.017925024032592773,
0.09845807403326035,
0.07229208946228027,
0.004584908951073885,
-0.052848875522613525,
0.0659029632806778,
0.02576509863138199,
-0.02219088189303875,
-0.11728318780660629,
0.040060438215732574,
-0.03589797765016556,
-0.0748615711927414,
0.045048948377370834,
-0.010411092080175877,
0.06805140525102615,
-0.09178891032934189,
0.04757171869277954,
0.04878690466284752,
0.003779756370931864,
-0.0660674124956131,
0.053244344890117645,
-0.13652385771274567,
-0.06953426450490952,
-0.0036684086080640554,
0.029542554169893265,
-0.12116722017526627,
0.01991480216383934,
-0.06205333396792412,
-0.049907587468624115,
-0.007758046966046095,
0.000979109900072217,
-0.005996992811560631,
0.059806495904922485,
1.15177654824254e-33,
-0.0010282339062541723,
0.07371200621128082,
-0.01628570444881916,
0.07977666705846786,
-0.0014417512575164437,
0.006810538005083799,
0.03252378851175308,
0.06027432903647423,
-0.0301783736795187,
-0.07387049496173859,
-0.021664995700120926,
0.0005117577966302633,
-0.033567510545253754,
0.013946201652288437,
0.011357401497662067,
-0.02857017144560814,
0.06131396442651749,
0.03911914303898811,
-0.051671162247657776,
0.026263361796736717,
0.01749577559530735,
0.06311869621276855,
-0.00317507260479033,
0.04470161348581314,
-0.11016283184289932,
0.06727790832519531,
-0.048564180731773376,
-0.005002210848033428,
-0.007115372456610203,
-0.028979580849409103,
0.02774919383227825,
0.04629466310143471,
0.01660795882344246,
-0.0476904958486557,
-0.08557194471359253,
0.08198525756597519,
-0.0009898929856717587,
-0.04853574559092522,
0.00464071286842227,
0.1177014634013176,
0.07328861951828003,
-0.07710827887058258,
-0.08987923711538315,
0.010156245902180672,
-0.0729537382721901,
-0.07792109251022339,
-0.059869807213544846,
-0.013785785995423794,
-0.004291133023798466,
0.07759635895490646,
-0.042870182543992996,
0.056055840104818344,
-0.09822160750627518,
0.03811310976743698,
-0.033174045383930206,
-0.12371791154146194,
-0.0499349869787693,
-0.010129015892744064,
0.056154750287532806,
-0.0030719649512320757,
-0.04449468106031418,
0.019990168511867523,
-0.0338737778365612,
-0.06950575858354568,
0.007174831815063953,
-0.08828876167535782,
-0.002102454425767064,
-0.06082475185394287,
0.06643234193325043,
-0.007003876380622387,
0.020387766882777214,
-0.006394596770405769,
0.015294534154236317,
-0.0076751415617764,
0.04750298708677292,
0.004755351692438126,
-0.003826865926384926,
0.03177325800061226,
-0.030239135026931763,
-0.024934707209467888,
0.03462987765669823,
-0.06532540917396545,
0.03916405513882637,
0.009206281043589115,
0.11376695334911346,
0.11940715461969376,
-0.020887674763798714,
-0.0018499605357646942,
0.03382520005106926,
0.04013669490814209,
-0.0054153925739228725,
-0.016883300617337227,
0.026820752769708633,
0.04594467207789421,
-0.03794608637690544,
-4.129088893023436e-8,
-0.07658912986516953,
0.009278430603444576,
0.011631165631115437,
0.018137171864509583,
-0.005918172653764486,
-0.06808271259069443,
0.0038619765546172857,
0.0333712212741375,
0.04856378585100174,
0.008668766357004642,
0.06953085213899612,
-0.013828100636601448,
-0.04898339509963989,
-0.06670169532299042,
-0.039136841893196106,
0.048580776900053024,
-0.004449835978448391,
0.038736920803785324,
0.023493174463510513,
0.030246427282691002,
0.014287127181887627,
-0.016667861491441727,
0.024811916053295135,
-0.05558453127741814,
0.03447902202606201,
-0.01642276905477047,
-0.0585966520011425,
0.03483027219772339,
0.0743388906121254,
-0.07546558976173401,
0.0509764701128006,
0.032989293336868286,
-0.02006976120173931,
-0.06606795638799667,
0.004606449045240879,
0.06681497395038605,
0.045498453080654144,
0.0036803644616156816,
-0.038083046674728394,
0.020900148898363113,
0.0387529581785202,
0.04351899400353432,
-0.012048599310219288,
0.015130067244172096,
-0.006093581207096577,
0.037653520703315735,
0.06114668771624565,
-0.08655497431755066,
-0.00030980631709098816,
-0.051399871706962585,
-0.019763657823204994,
-0.09328906983137131,
0.01135161891579628,
0.11810795217752457,
0.035221368074417114,
-0.05092942714691162,
0.019695205613970757,
0.03478120639920235,
0.06849085539579391,
0.11921554803848267,
-0.0131866205483675,
-0.006063527427613735,
-0.038398925215005875,
-0.012729959562420845
] |
Regulatory DNA Sequence Design with Reinforcement Learning | https://openreview.net/forum?id=F4IMiNhim1 | [
"Zhao Yang",
"Bing Su",
"Chuan Cao",
"Ji-Rong Wen"
] | Poster | $\textit{Cis}$-regulatory elements (CREs), such as promoters and enhancers, are relatively short DNA sequences that directly regulate gene expression. The fitness of CREs, measured by their ability to modulate gene expression, highly depends on the nucleotide sequences, especially specific motifs known as transcription factor binding sites (TFBSs). Designing high-fitness CREs is crucial for therapeutic and bioengineering applications. Current CRE design methods are limited by two major drawbacks: (1) they typically rely on iterative optimization strategies that modify existing sequences and are prone to local optima, and (2) they lack the guidance of biological prior knowledge in sequence optimization. In this paper, we address these limitations by proposing a generative approach that leverages reinforcement learning (RL) to fine-tune a pre-trained autoregressive (AR) model. Our method incorporates data-driven biological priors by deriving computational inference-based rewards that simulate the addition of activator TFBSs and removal of repressor TFBSs, which are then integrated into the RL process. We evaluate our method on promoter design tasks in two yeast media conditions and enhancer design tasks for three human cell types, demonstrating its ability to generate high-fitness CREs while maintaining sequence diversity. The code is available at https://github.com/yangzhao1230/TACO. | sequence optimize, generative models, ai4science, dna, rl | null | 897 | 2503.07981 | [
-0.0973711684346199,
-0.06569032371044159,
-0.002904100576415658,
-0.002943937201052904,
-0.028717823326587677,
0.05170504376292229,
-0.021923154592514038,
0.022542379796504974,
-0.0067076608538627625,
0.012450245209038258,
-0.027456412091851234,
-0.04549628123641014,
-0.05879875645041466,
0.02414589375257492,
-0.06008252128958702,
0.025061234831809998,
0.03351370245218277,
0.05992858111858368,
-0.02448631264269352,
-0.1115565374493599,
0.024926191195845604,
0.060182999819517136,
0.08349204808473587,
0.0007636926602572203,
-0.006778677925467491,
-0.0054230154491961,
-0.002249696059152484,
0.034450430423021317,
0.03915300592780113,
0.005910970736294985,
0.059512194246053696,
-0.03188092261552811,
0.07267887890338898,
-0.0837964192032814,
-0.06914497911930084,
0.0602554976940155,
-0.0695786327123642,
-0.005558904726058245,
0.010018267668783665,
0.019646912813186646,
0.051228541880846024,
0.000316659890813753,
0.002606435911729932,
0.08444435149431229,
0.02526979148387909,
-0.0016915848245844245,
0.03306564316153526,
-0.01833268627524376,
-0.013240552507340908,
-0.026143725961446762,
-0.04536135122179985,
-0.03516342490911484,
-0.010791237466037273,
-0.007678505498915911,
0.07331417500972748,
0.0637807846069336,
0.025691427290439606,
-0.06955495476722717,
0.04206662252545357,
-0.06309620290994644,
-0.005213883239775896,
-0.14662230014801025,
-0.014452468603849411,
-0.03299165517091751,
-0.03330065310001373,
-0.04120282083749771,
0.01651226170361042,
0.10113514214754105,
0.0018001459538936615,
0.02057527005672455,
-0.11108811944723129,
0.02875404991209507,
-0.09188181161880493,
0.04370918869972229,
0.03962627798318863,
0.11838255077600479,
-0.008679041638970375,
-0.05744122713804245,
0.04110034182667732,
-0.05432477965950966,
-0.007343686651438475,
-0.008185844868421555,
0.10232748836278915,
0.025850597769021988,
0.07328823953866959,
-0.028565270826220512,
0.01029506791383028,
0.04795004799962044,
0.12521561980247498,
0.09962554275989532,
0.05525858327746391,
-0.00923928152769804,
0.01570025645196438,
-0.0483291894197464,
-0.04145629331469536,
0.08009330928325653,
-0.024819888174533844,
-0.06637976318597794,
-0.02410239167511463,
0.018840616568922997,
0.031033622100949287,
0.07089763879776001,
-0.06415624171495438,
0.04932212829589844,
-0.016453636810183525,
-0.03307405859231949,
-0.024393826723098755,
0.06822281330823898,
0.04148946702480316,
0.029038291424512863,
-0.020348262041807175,
0.06407060474157333,
0.049178265035152435,
0.02853730320930481,
-0.01852254569530487,
0.018884366378188133,
0.007484195288270712,
-0.053332820534706116,
0.01957227662205696,
0.0439818911254406,
0.009577120654284954,
-0.053798168897628784,
-0.041608646512031555,
0.002066914224997163,
-0.035998884588479996,
-0.09847242385149002,
-0.11064509302377701,
1.0870868900648818e-33,
0.01948682963848114,
0.06405391544103622,
0.02024284191429615,
0.02131623402237892,
-0.021163059398531914,
-0.04533223435282707,
-0.03338536247611046,
0.011079559102654457,
-0.10467958450317383,
0.011600937694311142,
-0.023982293903827667,
-0.016153713688254356,
-0.12064497172832489,
0.10517662018537521,
0.0019800206646323204,
-0.03332774341106415,
-0.03858807310461998,
0.06475932896137238,
-0.0047325496561825275,
-0.03181004524230957,
-0.03779801353812218,
-0.04904021695256233,
-0.04236825555562973,
-0.025910040363669395,
0.054760806262493134,
0.020742051303386688,
0.07677222043275833,
0.06090833246707916,
-0.05025624483823776,
-0.020130880177021027,
-0.09332305938005447,
-0.03209759667515755,
0.024783151224255562,
0.0428922139108181,
0.04373420402407646,
-0.03598860651254654,
0.0658385306596756,
-0.023509640246629715,
0.0024937125854194164,
0.07328619062900543,
0.019231632351875305,
0.02290572226047516,
0.03716415539383888,
0.032513778656721115,
-0.05678003653883934,
-0.083877332508564,
0.0843362808227539,
-0.04048614203929901,
-0.04417620599269867,
0.00003745462527149357,
0.029413282871246338,
0.01824813149869442,
0.0647902563214302,
-0.05522726848721504,
-0.03407271206378937,
0.07609047740697861,
-0.07856647670269012,
0.03608935326337814,
-0.0018416889943182468,
0.045331940054893494,
-0.07635804265737534,
0.05204043164849281,
0.03031466342508793,
0.08627410978078842,
0.006083962041884661,
0.03468400612473488,
-0.0354500487446785,
-0.025997160002589226,
0.031900446861982346,
0.05281665921211243,
0.04870956763625145,
0.0041439286433160305,
0.03229114040732384,
-0.1073019728064537,
0.09501056373119354,
-0.05147437006235123,
-0.006834434811025858,
-0.03685537725687027,
-0.11523392796516418,
0.061261944472789764,
0.0032703657634556293,
0.07457009702920914,
-0.06439834088087082,
0.020045451819896698,
0.0009477593703195453,
-0.12221968919038773,
0.04817286878824234,
-0.019945764914155006,
-0.03228091448545456,
-0.0990934669971466,
-0.041077036410570145,
-0.044703610241413116,
0.007211475633084774,
0.056397754698991776,
0.08385240286588669,
-1.5842482667497523e-33,
0.06586689502000809,
-0.048160381615161896,
0.06726717948913574,
-0.046413566917181015,
0.0030328850261867046,
0.02870859205722809,
0.028088577091693878,
-0.06416159123182297,
0.030362650752067566,
-0.12771782279014587,
0.007409449201077223,
-0.012673781253397465,
-0.00910988450050354,
0.016323981806635857,
-0.043658580631017685,
0.0076047987677156925,
-0.0720120519399643,
0.04385770112276077,
-0.02690490335226059,
0.005313291680067778,
0.02666965313255787,
0.051708828657865524,
-0.10189241915941238,
0.07897310703992844,
-0.011952809058129787,
0.0006283064722083509,
-0.020365765318274498,
0.1265263706445694,
-0.02277647703886032,
0.04524480924010277,
-0.09386724978685379,
-0.019733965396881104,
-0.036338355392217636,
-0.012759564444422722,
-0.06411320716142654,
0.11793935298919678,
0.06341027468442917,
0.06748881191015244,
-0.024500854313373566,
0.04559481516480446,
0.06853576749563217,
0.038245175033807755,
-0.11103948205709457,
0.010559823364019394,
0.009174213744699955,
0.01846560649573803,
0.01289310585707426,
0.0007913429872132838,
0.057880256325006485,
0.02172604575753212,
-0.052626755088567734,
0.038720469921827316,
-0.032739803194999695,
0.0067072901874780655,
-0.02427639439702034,
-0.11791278421878815,
0.01355478260666132,
-0.024352628737688065,
0.1110868975520134,
-0.0008732572314329445,
-0.06669118255376816,
0.031004145741462708,
0.04653304070234299,
-0.030937697738409042,
-0.019564196467399597,
0.0002263213536934927,
0.00849827565252781,
0.024352794513106346,
0.03029036335647106,
-0.0414353646337986,
0.0011159437708556652,
0.0525282621383667,
0.07022759318351746,
0.036493562161922455,
0.07520409673452377,
-0.06391099840402603,
-0.026924671605229378,
-0.01485508680343628,
-0.04056774824857712,
0.028131620958447456,
0.05064119026064873,
-0.020176120102405548,
-0.00300754071213305,
0.024626338854432106,
0.012929977849125862,
0.04809725284576416,
0.031166188418865204,
0.09848205000162125,
0.045330338180065155,
-0.01710417866706848,
-0.02305726148188114,
0.006656637415289879,
0.008892879821360111,
0.03426064923405647,
-0.04582864046096802,
-4.538457787361949e-8,
-0.040153004229068756,
-0.06313608586788177,
0.08542618155479431,
-0.001680761924944818,
0.07578817754983902,
0.04637685790657997,
-0.07255079597234726,
0.013869185000658035,
0.03912392258644104,
-0.10773124545812607,
0.029449496418237686,
0.027707410976290703,
0.016243448480963707,
-0.02750592678785324,
0.02181997150182724,
0.013799858279526234,
0.008903855457901955,
-0.0086669297888875,
-0.0054082246497273445,
0.014143488369882107,
-0.014322258532047272,
0.04182364419102669,
-0.006643857806921005,
-0.0679708793759346,
0.07412301003932953,
-0.125014990568161,
-0.0021182855125516653,
-0.017627011984586716,
0.08049742132425308,
0.0016459458274766803,
0.04682886227965355,
-0.010776495561003685,
0.037987612187862396,
0.00291707762517035,
-0.01907929964363575,
0.05735423415899277,
0.04427216947078705,
-0.0840902253985405,
-0.057539600878953934,
-0.03605451062321663,
-0.03327944874763489,
0.08098454028367996,
-0.04422933980822563,
-0.05121273174881935,
-0.05911939963698387,
0.009133102372288704,
-0.00725986622273922,
-0.05939042568206787,
0.058518536388874054,
-0.039796892553567886,
0.08758531510829926,
-0.003260201308876276,
0.01876021921634674,
0.01943378709256649,
0.060224223881959915,
-0.03727845847606659,
0.04242868721485138,
-0.028817901387810707,
0.031162062659859657,
0.005666919983923435,
-0.0019010624382644892,
-0.028200069442391396,
-0.047660984098911285,
-0.057022180408239365
] |
In-Context Editing: Learning Knowledge from Self-Induced Distributions | https://openreview.net/forum?id=w6rHCuN3YG | [
"Siyuan Qi",
"Bangcheng Yang",
"Kailin Jiang",
"Xiaobo Wang",
"Jiaqi Li",
"Yifan Zhong",
"Yaodong Yang",
"Zilong Zheng"
] | Poster | In scenarios where language models must incorporate new information efficiently without extensive retraining, traditional fine-tuning methods are prone to overfitting, degraded generalization, and unnatural language generation. To address these limitations, we introduce Consistent In-Context Editing (ICE), a novel approach leveraging the model's in-context learning capability to optimize towards a contextual distribution rather than a one-hot target. ICE introduces a simple yet effective optimization framework for the model to internalize new knowledge by aligning its output distributions with and without additional context. This method enhances the robustness and effectiveness of gradient-based tuning methods, preventing overfitting and preserving the model's integrity. We analyze ICE across four critical aspects of knowledge editing: accuracy, locality, generalization, and linguistic quality, demonstrating its advantages. Experimental results confirm the effectiveness of ICE and demonstrate its potential for continual editing, ensuring that the integrity of the model is preserved while updating information. | Knowledge Editing, In-Context Learning, Language Models | We propose Consistent In-Context Editing, an approach for tuning language models through contextual distributions, overcoming the limitations of traditional fine-tuning methods that learn towards one-hot targets. | 896 | 2406.11194 | [
-0.007609253283590078,
-0.037514738738536835,
-0.0008910340839065611,
0.10606794059276581,
0.035241782665252686,
-0.03146706894040108,
-0.03135347366333008,
-0.02114214561879635,
0.022034918889403343,
0.01325471606105566,
-0.051866307854652405,
-0.01692172884941101,
0.06503306329250336,
0.005668018478900194,
0.0003748090530280024,
-0.00997739378362894,
0.10834301263093948,
0.05671745911240578,
-0.14649763703346252,
-0.06391638517379761,
0.005645356606692076,
0.06171978637576103,
0.06671368330717087,
0.008167905732989311,
0.03147530555725098,
-0.02559530921280384,
-0.052421241998672485,
-0.020631026476621628,
0.06191742792725563,
-0.004562706686556339,
0.004315294791013002,
0.08297789841890335,
0.008901492692530155,
0.02024146541953087,
-0.025734852999448776,
0.09920646250247955,
-0.04145665839314461,
0.08696088939905167,
0.014181194826960564,
-0.059307586401700974,
-0.008590112440288067,
-0.014163493178784847,
0.009409663267433643,
0.085778146982193,
0.05687916651368141,
0.016341732814908028,
0.011560862883925438,
0.022274499759078026,
0.009746777825057507,
0.07021953165531158,
-0.11150410026311874,
-0.056526314467191696,
-0.06197641044855118,
-0.0014648212818428874,
-0.004340522922575474,
0.0429132841527462,
0.002480593044310808,
0.048535726964473724,
-0.035991083830595016,
-0.06566490232944489,
-0.0016385938506573439,
-0.06292106956243515,
-0.04518045857548714,
0.024095075204968452,
-0.02911403588950634,
-0.04679396376013756,
0.0005569453933276236,
0.05108122155070305,
-0.046997275203466415,
-0.015397319570183754,
0.02840891294181347,
0.0586940161883831,
-0.01858396828174591,
0.08205033093690872,
-0.008978491649031639,
0.057344239205121994,
-0.03780534118413925,
-0.032827235758304596,
-0.006697051227092743,
-0.013232138007879257,
0.07664474099874496,
-0.01806763745844364,
0.029314948245882988,
0.005078761838376522,
0.11031646281480789,
0.010826325975358486,
0.11527438461780548,
-0.0055024148896336555,
0.07811807096004486,
0.0629177838563919,
-0.039923783391714096,
-0.0448395311832428,
0.05566301569342613,
-0.06392382830381393,
0.00030876006348989904,
0.04876773804426193,
0.0652993693947792,
-0.058876581490039825,
-0.026170650497078896,
0.10669420659542084,
0.03590438514947891,
0.04016948863863945,
-0.05600708723068237,
-0.03469232842326164,
-0.02052171528339386,
-0.026616564020514488,
0.007626831065863371,
0.08081571757793427,
-0.011381281539797783,
-0.08276936411857605,
0.013084768317639828,
0.028789740055799484,
0.002597227692604065,
-0.06983713060617447,
0.017871731892228127,
0.007800839375704527,
0.06191533803939819,
0.014995166100561619,
-0.0286578182131052,
0.1113692969083786,
-0.03841159865260124,
0.01411623228341341,
-0.0018538245931267738,
0.04504935443401337,
-0.11138897389173508,
-0.03987080603837967,
-0.028537996113300323,
-1.2337200757673694e-33,
0.10033585131168365,
0.11135149002075195,
0.011969560757279396,
0.039939865469932556,
0.052776191383600235,
0.055823542177677155,
-0.058292776346206665,
-0.0014380604261532426,
-0.01846887916326523,
-0.0514150932431221,
0.10640579462051392,
-0.007062162272632122,
-0.024206971749663353,
0.02349948137998581,
0.06416299939155579,
0.04951488971710205,
-0.06310030072927475,
0.07312902808189392,
-0.031919222325086594,
0.020880021154880524,
0.016390375792980194,
-0.02311628870666027,
-0.06648317724466324,
-0.09327986091375351,
-0.056489232927560806,
0.1005544438958168,
0.10792074352502823,
-0.007274002768099308,
-0.024133911356329918,
-0.02076140232384205,
-0.0796782597899437,
-0.03189828619360924,
0.04282708093523979,
0.05909872055053711,
-0.013005705550312996,
-0.07090027630329132,
0.00032875535544008017,
-0.09051904827356339,
0.005130477249622345,
-0.029966995120048523,
-0.02603454515337944,
0.03262218087911606,
0.0031326175667345524,
0.018098905682563782,
-0.07692819088697433,
-0.04352784529328346,
-0.0025869037490338087,
0.022684408351778984,
-0.049693524837493896,
-0.04731982201337814,
0.06398999691009521,
0.10539065301418304,
0.007403652183711529,
-0.001598647446371615,
-0.03413252532482147,
0.05245110020041466,
0.07270751893520355,
-0.010673398151993752,
0.028000587597489357,
0.045618388801813126,
0.04013250768184662,
0.026891285553574562,
-0.010822169482707977,
0.10537134110927582,
0.0655558779835701,
0.021688656881451607,
0.0030433405190706253,
0.05991019308567047,
0.08986634016036987,
-0.09458178281784058,
-0.08290275186300278,
-0.035320572555065155,
-0.05115523561835289,
0.01238569151610136,
-0.010327563621103764,
-0.03853912651538849,
-0.0038571234326809645,
-0.04807361215353012,
-0.014262545853853226,
-0.019834373146295547,
0.03381289541721344,
-0.019330380484461784,
0.024728303775191307,
-0.004203198943287134,
-0.0639762282371521,
-0.054993368685245514,
-0.0106113376095891,
0.014058483764529228,
-0.007075731176882982,
-0.045013491064310074,
-0.01619834452867508,
-0.03038061037659645,
-0.009914425201714039,
-0.021146688610315323,
0.002497937763109803,
1.6292818226717387e-33,
-0.022791409865021706,
-0.0573836974799633,
-0.05363010615110397,
0.03269769623875618,
-0.008602937683463097,
-0.06847058981657028,
0.009923241101205349,
0.06437346339225769,
0.01964068040251732,
-0.0943499356508255,
-0.033641036599874496,
-0.08124829828739166,
0.01580839790403843,
-0.013883224688470364,
-0.011737477034330368,
0.01345375832170248,
-0.006295241415500641,
0.04232769086956978,
-0.03249288722872734,
0.019413838163018227,
0.0043165357783436775,
-0.028325796127319336,
-0.05522415414452553,
0.12620535492897034,
-0.0003508192312438041,
0.0076700737699866295,
-0.06128254160284996,
0.04522017762064934,
-0.03899894654750824,
0.0344284251332283,
-0.0623280294239521,
0.029149502515792847,
0.000957872427534312,
-0.02877834066748619,
-0.058333106338977814,
0.025284111499786377,
0.03762922063469887,
-0.07074771821498871,
-0.06254410743713379,
0.05767309293150902,
0.07745383679866791,
0.009065068326890469,
-0.0343625470995903,
-0.01352768950164318,
-0.002512184903025627,
-0.009360835887491703,
-0.13356764614582062,
-0.037647366523742676,
0.04076424241065979,
0.029850715771317482,
-0.07645627111196518,
-0.012443707324564457,
-0.10933523625135422,
-0.004452979192137718,
-0.00140185107011348,
-0.07850459218025208,
0.004227229859679937,
-0.047114454209804535,
0.11762501299381256,
-0.05376024544239044,
-0.07738882303237915,
-0.029687441885471344,
0.03431370481848717,
-0.05419985204935074,
0.017035242170095444,
-0.04672911390662193,
-0.038498200476169586,
0.039157141000032425,
0.039065171033144,
0.009683947078883648,
0.012802579440176487,
-0.015293602831661701,
-0.02675141766667366,
-0.04351797327399254,
0.020030589774250984,
-0.03382173925638199,
0.08231915533542633,
-0.10593658685684204,
-0.01178649254143238,
-0.09050985425710678,
0.011046884581446648,
0.015023365616798401,
0.046390876173973083,
0.049155980348587036,
0.04180039465427399,
0.08767546713352203,
0.03955892473459244,
0.12838952243328094,
0.015575769357383251,
0.02038329653441906,
-0.03981998562812805,
0.011473209597170353,
-0.041116490960121155,
0.07903742790222168,
-0.13915592432022095,
-3.941894988201966e-8,
-0.09358420222997665,
0.05868558958172798,
-0.04110546410083771,
0.12071835249662399,
0.030948448926210403,
-0.004826252348721027,
-0.017521781846880913,
0.08569827675819397,
0.054571591317653656,
-0.054864175617694855,
-0.04239386320114136,
0.015756169334053993,
-0.037742193788290024,
-0.06734742969274521,
-0.05995660275220871,
0.05812516063451767,
-0.00006841101276222616,
0.0965644046664238,
-0.003119662869721651,
-0.016598764806985855,
0.04337701573967934,
0.062257442623376846,
-0.006364829372614622,
-0.03732729330658913,
0.07237236946821213,
-0.05559489130973816,
-0.008204752579331398,
0.08976631611585617,
0.10781759023666382,
-0.0059826988726854324,
-0.03785359486937523,
-0.027766024693846703,
0.04165129363536835,
0.022752052173018456,
-0.03380545973777771,
0.06014957278966904,
0.0645035058259964,
-0.005128628108650446,
0.009992868639528751,
-0.043033696711063385,
0.04547851160168648,
0.055082324892282486,
-0.03639915585517883,
0.04401267319917679,
-0.0073006064631044865,
-0.002137629548087716,
0.017239084467291832,
-0.08517492562532425,
0.009578069671988487,
-0.033105116337537766,
0.03233315423130989,
-0.0022262665443122387,
0.010774124413728714,
-0.030751729384064674,
0.030114592984318733,
0.03479614853858948,
-0.053082775324583054,
0.032702360302209854,
0.03558799624443054,
0.032002516090869904,
-0.049233973026275635,
0.05442912131547928,
0.024013565853238106,
0.027802899479866028
] |
CL-DiffPhyCon: Closed-loop Diffusion Control of Complex Physical Systems | https://openreview.net/forum?id=PiHGrTTnvb | [
"Long Wei",
"Haodong Feng",
"Yuchen Yang",
"Ruiqi Feng",
"Peiyan Hu",
"Xiang Zheng",
"Tao Zhang",
"Dixia Fan",
"Tailin Wu"
] | Poster | The control problems of complex physical systems have broad applications in science and engineering. Previous studies have shown that generative control methods based on diffusion models offer significant advantages for solving these problems. However, existing generative control approaches face challenges in both performance and efficiency when extended to the closed-loop setting, which is essential for effective control. In this paper, we propose an efficient Closed-Loop Diffusion method for Physical systems Control (CL-DiffPhyCon). By employing an asynchronous denoising framework for different physical time steps, CL-DiffPhyCon generates control signals conditioned on real-time feedback from the system with significantly reduced computational cost during sampling. Additionally, the control process could be further accelerated by incorporating fast sampling techniques, such as DDIM. We evaluate CL-DiffPhyCon on two tasks: 1D Burgers' equation control and 2D incompressible fluid control. The results demonstrate that CL-DiffPhyCon achieves superior control performance with significant improvements in sampling efficiency. The code can be found at https://github.com/AI4Science-WestlakeU/CL_DiffPhyCon. | physical systems control, closed-loop control, PDE, physical simulation, generative models | We propose a diffusion method with an asynchronous denoising schedule for physical systems control tasks. It achieves closed-loop control with significant speedup of sampling efficiency. | 889 | null | [
-0.10262899845838547,
-0.04665166512131691,
0.004982672166079283,
0.03561193495988846,
-0.010540466755628586,
-0.10019098222255707,
-0.048662856221199036,
-0.017962375655770302,
0.010893505066633224,
0.017146874219179153,
-0.021898474544286728,
-0.04640622437000275,
0.01473059132695198,
-0.10469964891672134,
-0.02537740394473076,
-0.00006118804594734684,
0.03862105309963226,
0.04995448514819145,
-0.02858804725110531,
-0.018123427405953407,
-0.049370668828487396,
0.014407206326723099,
-0.008846430107951164,
0.006296002771705389,
-0.019230473786592484,
0.07589341700077057,
0.0017011959571391344,
-0.08905620872974396,
0.03194355592131615,
-0.06578908860683441,
0.021928168833255768,
-0.04869598150253296,
-0.05140646919608116,
-0.0660707950592041,
-0.04077988862991333,
-0.002726742997765541,
-0.002954735653474927,
0.0030809533782303333,
-0.035526879131793976,
0.010875111445784569,
0.06168872490525246,
0.10010220855474472,
0.05110899358987808,
-0.0005328416009433568,
0.05097031593322754,
-0.014356110244989395,
-0.07178723812103271,
0.012277317233383656,
0.03021140955388546,
0.04140200465917587,
0.011776992119848728,
0.009357146918773651,
0.04286535084247589,
-0.060996606945991516,
0.12349137663841248,
-0.0029782194178551435,
0.07823160290718079,
0.07977461069822311,
0.002999641699716449,
-0.06626236438751221,
-0.005667706951498985,
-0.028514550998806953,
-0.007007546257227659,
-0.013849765062332153,
-0.0979136973619461,
-0.05976211279630661,
0.05458778142929077,
0.08553222566843033,
0.07105637341737747,
-0.03200342133641243,
-0.031033825129270554,
0.07065992802381516,
-0.03341114521026611,
-0.02255994640290737,
0.054227784276008606,
-0.05877625197172165,
-0.01725713536143303,
-0.0007840130710974336,
-0.020275618880987167,
-0.01011089887470007,
0.07521939277648926,
0.04664062708616257,
-0.006005444563925266,
-0.08323542028665543,
-0.09001553058624268,
0.017106834799051285,
0.0031032925471663475,
0.0010644887806847692,
0.08636941015720367,
-0.07191482186317444,
-0.012832659296691418,
0.11879942566156387,
-0.08656398952007294,
-0.05812451243400574,
-0.040335435420274734,
0.0016186138382181525,
0.03363381326198578,
0.10454947501420975,
0.00034439010778442025,
0.03369014337658882,
0.003431773278862238,
0.041332315653562546,
-0.02079521119594574,
0.054155346006155014,
0.01816263049840927,
-0.08815870434045792,
0.1176794022321701,
0.008285025134682655,
-0.05970953404903412,
0.028276246041059494,
0.03727882727980614,
-0.042176272720098495,
-0.001337748602963984,
0.11787144094705582,
0.061732664704322815,
-0.054551009088754654,
0.09585335850715637,
-0.057583559304475784,
0.054144661873579025,
-0.05710450932383537,
-0.022817594930529594,
-0.08699063211679459,
-0.0706234723329544,
-0.004775156732648611,
0.05432237684726715,
0.02797965332865715,
0.03076358139514923,
3.750540081750764e-34,
-0.032330162823200226,
-0.022005189210176468,
-0.04114997759461403,
0.030307093635201454,
0.04971417784690857,
0.007051087915897369,
0.04106974974274635,
-0.020219173282384872,
-0.01614193059504032,
-0.011713490821421146,
-0.0012691299198195338,
-0.036698829382658005,
-0.1360522359609604,
0.03146643936634064,
0.04017407074570656,
-0.12218810617923737,
0.020360521972179413,
0.020586149767041206,
-0.013820602558553219,
0.01634538732469082,
0.11380434036254883,
0.02182108908891678,
-0.08022487163543701,
-0.03864893689751625,
-0.015950532630085945,
-0.01602826453745365,
-0.024716630578041077,
0.04647429287433624,
0.06104111671447754,
0.010818159207701683,
-0.04695235192775726,
-0.007985522039234638,
-0.06954433023929596,
-0.002183082513511181,
0.01832009293138981,
0.06029500067234039,
0.02496875822544098,
-0.015592320822179317,
-0.00012335882638581097,
0.017222708091139793,
-0.002401135629042983,
-0.032322514802217484,
-0.046440355479717255,
-0.013260982930660248,
-0.12951013445854187,
-0.03628624230623245,
0.02803174778819084,
0.015939442440867424,
-0.052673909813165665,
-0.0295224878937006,
-0.0366525799036026,
0.04591285064816475,
-0.011843925341963768,
-0.08036816865205765,
0.029326042160391808,
-0.022313909605145454,
0.05695285648107529,
-0.11809474229812622,
-0.005208524875342846,
0.14276306331157684,
-0.0336625799536705,
0.058366477489471436,
-0.04400443285703659,
0.018825778737664223,
0.07853838801383972,
0.00781578291207552,
-0.04570640251040459,
0.08298345655202866,
0.07175607234239578,
-0.039544571191072464,
-0.008687224239110947,
-0.04540591686964035,
-0.05248415097594261,
-0.05838174745440483,
0.12085359543561935,
-0.04502079263329506,
0.024286042898893356,
0.04016231372952461,
-0.047113172709941864,
-0.043968088924884796,
0.03792858123779297,
0.04804612696170807,
-0.07405433803796768,
0.016379596665501595,
0.03843887895345688,
-0.061942752450704575,
-0.06258180737495422,
0.06411341577768326,
-0.06655272841453552,
-0.11460103839635849,
-0.08523745089769363,
-0.012497439049184322,
0.07537346333265305,
-0.029930133372545242,
-0.035583145916461945,
-3.402837944491651e-33,
0.06154987961053848,
-0.012470691464841366,
0.041208166629076004,
0.052344910800457,
-0.02123681828379631,
0.05480192229151726,
0.009787611663341522,
-0.05371616408228874,
0.07288239151239395,
-0.04142490401864052,
-0.08611960709095001,
-0.017083484679460526,
-0.02783912979066372,
0.02499403990805149,
0.031486645340919495,
-0.01576118916273117,
-0.024184858426451683,
-0.08655743300914764,
-0.02028634026646614,
-0.03155776113271713,
-0.09546177089214325,
0.022103866562247276,
0.0323391854763031,
-0.006865758448839188,
-0.08772774785757065,
0.0035067307762801647,
-0.060226332396268845,
0.09631805121898651,
0.01497903373092413,
-0.03882826119661331,
-0.11623957753181458,
0.028158824890851974,
0.0525984950363636,
0.0317949503660202,
-0.0789489671587944,
0.06850384175777435,
0.03685378655791283,
0.05339067801833153,
0.0004898214247077703,
0.008488785475492477,
0.005580095574259758,
-0.11873485893011093,
-0.0463688038289547,
0.055755048990249634,
0.08235397189855576,
0.023208167403936386,
-0.06370367109775543,
-0.02523593232035637,
-0.052468664944171906,
0.05680172145366669,
-0.019078580662608147,
0.027937399223446846,
-0.008590159006416798,
0.035077109932899475,
0.013382346369326115,
-0.033409375697374344,
-0.003298970405012369,
0.038090553134679794,
0.0369243286550045,
-0.002552083460614085,
-0.06063539534807205,
0.0019662114791572094,
-0.035725709050893784,
0.01266024261713028,
0.002778483321890235,
0.01701444573700428,
0.03049049898982048,
0.03362879529595375,
0.09114784747362137,
0.001171173295006156,
-0.025717059150338173,
0.02904415689408779,
0.07865551114082336,
0.007675875909626484,
0.03516869619488716,
-0.10624121129512787,
-0.025815462693572044,
-0.05163626745343208,
0.016551513224840164,
0.07158207893371582,
-0.0001257870753761381,
0.04888522997498512,
-0.0027810416650027037,
-0.01964888721704483,
-0.00580335408449173,
-0.03143393620848656,
-0.05038909614086151,
-0.06031091511249542,
-0.04721246659755707,
-0.01588393747806549,
0.048318520188331604,
-0.02745002694427967,
0.0916711762547493,
0.05801494047045708,
-0.04422139748930931,
-4.762618033282706e-8,
-0.021237114444375038,
0.07748155295848846,
0.06990870088338852,
0.03297797963023186,
0.0012494894908741117,
-0.010969212278723717,
-0.016398195177316666,
-0.07976511120796204,
0.043736111372709274,
-0.038570236414670944,
0.05136724188923836,
-0.03587943688035011,
0.017799751833081245,
-0.025444649159908295,
0.013212257996201515,
0.0453755147755146,
0.04127451032400131,
-0.038910627365112305,
0.03520263358950615,
-0.000505079806316644,
0.00786611158400774,
0.012775584124028683,
-0.0920986607670784,
0.027858570218086243,
0.06503470987081528,
0.0076271663419902325,
0.001026649260893464,
-0.012222899124026299,
0.08764167875051498,
0.002974802628159523,
-0.02723848819732666,
-0.0005110197816975415,
-0.005041142925620079,
0.10496778786182404,
0.0879441648721695,
-0.03409675881266594,
0.022565003484487534,
0.01840457320213318,
0.005833418108522892,
-0.04605431109666824,
-0.013332665897905827,
0.09934023022651672,
-0.087089404463768,
-0.012493740767240524,
-0.016028622165322304,
0.04464928060770035,
-0.01654045470058918,
-0.050083234906196594,
0.018502451479434967,
0.09167855232954025,
-0.01606878638267517,
0.008522194810211658,
0.04983241483569145,
0.03750327229499817,
-0.006470910739153624,
0.0037937043234705925,
-0.03936892747879028,
-0.05936247482895851,
0.02563302405178547,
0.03952571377158165,
-0.0392569899559021,
0.0492849200963974,
-0.016679862514138222,
0.08732635527849197
] |
OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting | https://openreview.net/forum?id=rAcgDBdKnP | [
"Xing Hu",
"Yuan Cheng",
"Dawei Yang",
"Zhixuan Chen",
"Zukang Xu",
"JiangyongYu",
"XUCHEN",
"Zhihang Yuan",
"Zhe jiang",
"Sifan Zhou"
] | Poster | Post-training quantization (PTQ) has emerged as a widely adopted technique for compressing and accelerating Large Language Models (LLMs).
The major challenge in LLM quantization is that uneven and heavy-tailed data distributions can expand the quantization range, thereby reducing bit precision for most values.
Recent methods attempt to eliminate outliers and balance inter-channel differences by employing linear transformations; however, they remain heuristic and are often overlook optimizing the data distribution across the entire quantization space.
In this paper, we introduce Quantization Space Utilization Rate (QSUR), a novel metric that effectively assesses the quantizability of transformed data by measuring the space utilization of the data in the quantization space. We complement QSUR with mathematical derivations that examine the effects and limitations of various transformations, guiding our development of Orthogonal and Scaling Transformation-based Quantization (OSTQuant). OSTQuant employs a learnable equivalent transformation, consisting of an orthogonal transformation and a scaling transformation, to optimize the distributions of weights and activations across the entire quantization space. Futhermore, we propose the KL-Top loss function, designed to mitigate noise during optimization while retaining richer semantic information within the limited calibration data imposed by PTQ.
OSTQuant outperforms existing work on various LLMs and benchmarks. In the W4-only setting, it retains 99.5\% of the floating-point accuracy. In the more challenging W4A4KV4 configuration, OSTQuant reduces the performance gap by 32\% on the LLaMA-3-8B model compared to state-of-the-art methods. Code will be available. | Large Language Models, Quantization | null | 887 | 2501.13987 | [
-0.03816552460193634,
-0.06077573820948601,
0.026089562103152275,
0.011634312570095062,
-0.00838241633027792,
0.03425066918134689,
-0.006931570824235678,
-0.01759476587176323,
0.09092980623245239,
-0.05427205190062523,
0.05084286630153656,
-0.006494433153420687,
-0.03389842435717583,
0.03905448690056801,
-0.01605992019176483,
-0.044922783970832825,
0.04768291488289833,
0.09197386354207993,
-0.12139476090669632,
-0.09447332471609116,
0.068149633705616,
0.040435682982206345,
-0.018763910979032516,
-0.005961219314485788,
0.13962379097938538,
0.041475385427474976,
-0.04883410409092903,
0.006107374094426632,
0.030886327847838402,
-0.006951547227799892,
0.03766578808426857,
0.06046851724386215,
0.06485116481781006,
-0.03321182727813721,
-0.035699401050806046,
0.0006592792924493551,
0.020198827609419823,
0.004388720262795687,
0.02483988367021084,
-0.03372184932231903,
0.018129166215658188,
-0.02775285392999649,
0.007244987413287163,
0.04019002988934517,
0.016615377739071846,
-0.023952694609761238,
-0.004802406765520573,
0.022172866389155388,
-0.034874312579631805,
0.038102105259895325,
0.0011636210838332772,
0.029700221493840218,
-0.014264833182096481,
0.08869308978319168,
-0.035769518464803696,
-0.05607953295111656,
0.049166202545166016,
0.08783436566591263,
-0.01875215768814087,
-0.026068788021802902,
-0.08960914611816406,
-0.0582067146897316,
0.008763834834098816,
0.053800858557224274,
-0.026203393936157227,
-0.027686448767781258,
0.025703510269522667,
0.031031114980578423,
-0.05477527529001236,
0.03600537031888962,
-0.0845080092549324,
0.029415858909487724,
-0.07085172832012177,
0.11479151993989944,
0.010352207347750664,
0.06728430837392807,
0.014296318404376507,
0.007647235877811909,
0.07131166756153107,
-0.035207588225603104,
0.04526735469698906,
0.021443909034132957,
0.04299407824873924,
0.005466809496283531,
0.1345464587211609,
0.0005181206506676972,
-0.019933711737394333,
0.03137902542948723,
-0.014397179707884789,
-0.007068284321576357,
-0.08915674686431885,
0.03899785131216049,
0.013737981207668781,
-0.008754686452448368,
0.02746802382171154,
0.022379282861948013,
-0.019030939787626266,
-0.013997252099215984,
0.040520209819078445,
0.09581111371517181,
0.05307018756866455,
0.06347402185201645,
0.010187174193561077,
-0.017839081585407257,
0.003741803113371134,
-0.03867762163281441,
-0.0028058646712452173,
0.03575529903173447,
0.047501713037490845,
-0.08188115060329437,
-0.007817642763257027,
0.03675907477736473,
-0.03576807677745819,
-0.0130958566442132,
0.06303201615810394,
0.007608390413224697,
0.046587832272052765,
0.014919674955308437,
-0.017560811713337898,
-0.009197504259645939,
-0.06613335758447647,
-0.0027417324017733335,
-0.06783653795719147,
0.05320711433887482,
0.007090097758919001,
0.006417765282094479,
-0.09158842265605927,
4.619166397667612e-33,
0.06941918283700943,
0.08397280424833298,
0.046255771070718765,
-0.04360732436180115,
0.03949090465903282,
-0.02483549900352955,
0.04132363200187683,
0.04505433887243271,
-0.012684804387390614,
0.018145861104130745,
0.010118646547198296,
0.022180387750267982,
-0.03525600954890251,
-0.03279011696577072,
0.05004826560616493,
0.014371194876730442,
-0.06383524090051651,
0.06732586771249771,
0.005462316330522299,
-0.08484074473381042,
0.09122255444526672,
-0.1185048520565033,
-0.011584400199353695,
0.029774688184261322,
0.008329370990395546,
-0.009322074241936207,
0.05305614322423935,
-0.015984654426574707,
-0.09551054984331131,
0.001730737159959972,
-0.10515357553958893,
-0.06769408285617828,
0.012373783625662327,
-0.08566597104072571,
0.050790805369615555,
-0.10698512196540833,
0.010566188022494316,
-0.04486358165740967,
0.0030021164566278458,
-0.03505928814411163,
-0.015273948200047016,
-0.019914638251066208,
0.03337010368704796,
-0.0579754002392292,
-0.050982557237148285,
0.040287911891937256,
0.00910233985632658,
-0.06069783866405487,
0.004364561755210161,
-0.1079082190990448,
0.07760483026504517,
-0.029863571748137474,
-0.03219030797481537,
0.049575649201869965,
0.06914617866277695,
0.02761050872504711,
0.0747046172618866,
0.006294028367847204,
0.0005213135154917836,
0.09400654584169388,
-0.03557489439845085,
-0.012402704916894436,
-0.0016075727762654424,
0.06627831608057022,
0.04411281645298004,
-0.07243438810110092,
0.025391485542058945,
0.06963595747947693,
-0.008640170097351074,
0.11479158699512482,
-0.049309760332107544,
0.00007613177876919508,
-0.0242890827357769,
0.014125928282737732,
0.07607176899909973,
-0.00508160749450326,
0.07371807098388672,
-0.056948766112327576,
-0.031776975840330124,
-0.021634802222251892,
-0.006147863809019327,
0.004525153432041407,
0.05924999713897705,
-0.05891045555472374,
-0.02228006161749363,
-0.104620061814785,
0.021960439160466194,
-0.05612146481871605,
-0.03643174469470978,
0.006242418196052313,
-0.038078226149082184,
-0.028743501752614975,
-0.1008082702755928,
-0.010477610863745213,
-0.05453941971063614,
-4.058739547679502e-33,
-0.1226675808429718,
0.07875845581293106,
-0.012193418107926846,
0.06963406503200531,
-0.08477578312158585,
-0.10192093998193741,
0.015097642317414284,
0.09177220612764359,
0.05506140738725662,
-0.09448558837175369,
-0.0823550745844841,
-0.08563605695962906,
0.050041571259498596,
-0.02937382273375988,
0.016722125932574272,
0.03826387971639633,
0.00644806819036603,
0.06844951957464218,
-0.06188426911830902,
0.1000804677605629,
-0.008003082126379013,
0.04173033684492111,
-0.046798884868621826,
0.010353682562708855,
0.005994101520627737,
0.002822425216436386,
-0.03362404927611351,
0.06533043086528778,
-0.00008413070463575423,
0.009574800729751587,
-0.05453451722860336,
-0.062039054930210114,
-0.06228981167078018,
-0.11303652822971344,
-0.09445086121559143,
0.00716240331530571,
0.030171843245625496,
0.04582485929131508,
-0.013380400836467743,
0.10314270108938217,
0.07229553908109665,
-0.05868958681821823,
-0.08705008774995804,
-0.00871705450117588,
0.020010842010378838,
0.015008918009698391,
-0.048212990164756775,
-0.03163982927799225,
-0.041500624269247055,
0.0239285696297884,
0.041403379291296005,
0.05929340794682503,
0.027949629351496696,
0.13653923571109772,
-0.03488747403025627,
-0.040635932236909866,
-0.031826138496398926,
-0.04119578003883362,
0.025139683857560158,
-0.02821645699441433,
-0.018601465970277786,
-0.029237186536192894,
0.021203335374593735,
-0.05542014166712761,
-0.013415000401437283,
-0.09843604266643524,
0.020021827891469002,
0.021408474072813988,
-0.03913440927863121,
0.024539414793252945,
0.01169243361800909,
-0.08017144352197647,
-0.017063429579138756,
0.016129959374666214,
-0.048855263739824295,
-0.03291330114006996,
0.03273310139775276,
-0.05147204175591469,
-0.020342500880360603,
0.04092792421579361,
0.028181271627545357,
0.07122080028057098,
0.027072008699178696,
-0.010070054791867733,
0.05918285995721817,
0.14594189822673798,
0.06356044858694077,
0.08626680821180344,
0.042460087686777115,
0.020093504339456558,
-0.041979752480983734,
0.04304647818207741,
-0.017688939347863197,
0.037436310201883316,
-0.05761191248893738,
-4.7308958528446965e-8,
-0.10505179315805435,
0.028117602691054344,
-0.058513566851615906,
-0.019326122477650642,
0.05216163024306297,
-0.0059114499017596245,
-0.023571118712425232,
0.13624650239944458,
0.002536810701712966,
-0.026040297001600266,
0.015073426067829132,
-0.017319224774837494,
-0.06937522441148758,
-0.04196254163980484,
-0.03828341141343117,
0.07499567419290543,
-0.06878683716058731,
-0.007094113621860743,
0.0035552599001675844,
-0.051204364746809006,
-0.040077630430459976,
0.1186584085226059,
-0.007848519831895828,
0.014740651473402977,
0.02677113190293312,
-0.035169996321201324,
0.007720056921243668,
0.08810580521821976,
0.13598129153251648,
-0.028374087065458298,
0.010135330259799957,
0.0007697888650000095,
-0.0036702563520520926,
-0.0012161441845819354,
-0.008409220725297928,
0.0504491925239563,
-0.04221614450216293,
-0.027389459311962128,
0.006631996016949415,
-0.012866740114986897,
0.0012627635151147842,
0.00747679965570569,
-0.05523304641246796,
0.07391911745071411,
0.03570225089788437,
0.021969184279441833,
-0.021959848701953888,
0.057346753776073456,
0.025769144296646118,
-0.09159044921398163,
0.0422133207321167,
-0.03636173903942108,
-0.02326780930161476,
-0.012263407930731773,
0.006915200501680374,
0.01331443339586258,
-0.01171028334647417,
-0.02115819603204727,
0.01491394080221653,
-0.010672164149582386,
-0.009576933458447456,
-0.051441896706819534,
-0.03987501189112663,
0.001829200773499906
] |
Distribution Backtracking Builds A Faster Convergence Trajectory for Diffusion Distillation | https://openreview.net/forum?id=2ySt3cdGfJ | [
"Shengyuan Zhang",
"Ling Yang",
"Zejian Li",
"An Zhao",
"Chenye Meng",
"Changyuan Yang",
"Guang Yang",
"Zhiyuan Yang",
"Lingyun Sun"
] | Poster | Accelerating the sampling speed of diffusion models remains a significant challenge. Recent score distillation methods distill a heavy teacher model into a student generator to achieve one-step generation, which is optimized by calculating the difference between two score functions on the samples generated by the student model.
However, there is a score mismatch issue in the early stage of the score distillation process, since existing methods mainly focus on using the endpoint of pre-trained diffusion models as teacher models, overlooking the importance of the convergence trajectory between the student generator and the teacher model.
To address this issue, we extend the score distillation process by introducing the entire convergence trajectory of the teacher model and propose $\textbf{Dis}$tribution $\textbf{Back}$tracking Distillation ($\textbf{DisBack}$). DisBask is composed of two stages: $\textit{Degradation Recording}$ and $\textit{Distribution Backtracking}$.
$\textit{Degradation Recording}$ is designed to obtain the convergence trajectory by recording the degradation path from the pre-trained teacher model to the untrained student generator.
The degradation path implicitly represents the intermediate distributions between the teacher and the student, and its reverse can be viewed as the convergence trajectory from the student generator to the teacher model.
Then $\textit{Distribution Backtracking}$ trains the student generator to backtrack the intermediate distributions along the path to approximate the convergence trajectory of the teacher model.
Extensive experiments show that DisBack achieves faster and better convergence than the existing distillation method and achieves comparable or better generation performance, with an FID score of 1.38 on the ImageNet 64$\times$64 dataset.
DisBack is easy to implement and can be generalized to existing distillation methods to boost performance. | Diffusion Model, Diffusion Distillation, One-step Generation | This paper proposes an efficient and fast distillation method for diffusion models by introducing the convergence trajectory. | 886 | 2408.15991 | [
-0.07998634874820709,
-0.07337725907564163,
0.03126566484570503,
0.027672497555613518,
0.02598286233842373,
-0.09028409421443939,
-0.04594532772898674,
-0.014460730366408825,
0.03303596004843712,
-0.0191635899245739,
0.013462424278259277,
0.04151470214128494,
0.03251872956752777,
0.0056010643020272255,
-0.04354298114776611,
-0.06487508118152618,
0.004217790439724922,
0.08975530415773392,
-0.06088458374142647,
-0.045106127858161926,
-0.012538902461528778,
0.013358519412577152,
-0.06131565570831299,
0.05141808092594147,
0.029318921267986298,
0.003235273528844118,
0.023170631378889084,
-0.03330241143703461,
0.017178701236844063,
-0.11456862092018127,
0.07839389890432358,
-0.04648956283926964,
-0.05791667476296425,
0.019835766404867172,
-0.08492095023393631,
0.05522899702191353,
-0.03430673107504845,
0.03219907730817795,
0.033943478018045425,
0.05127975344657898,
0.04871779680252075,
0.1116754487156868,
0.0004678419791162014,
0.055952221155166626,
0.10259297490119934,
-0.06055973470211029,
-0.032552070915699005,
-0.03886791318655014,
0.01899445429444313,
0.03100370429456234,
0.01949833333492279,
0.007416516076773405,
-0.028092702850699425,
0.045936647802591324,
-0.009186403825879097,
-0.010696890763938427,
0.09082230925559998,
0.03162755072116852,
-0.05348828062415123,
-0.040988702327013016,
-0.0648387148976326,
-0.03213178366422653,
-0.10127165168523788,
-0.0367322638630867,
0.01562664471566677,
-0.05542704835534096,
0.02175002731382847,
0.07090623676776886,
0.07282710820436478,
0.04824996739625931,
-0.07365771383047104,
0.07418309897184372,
-0.021595392376184464,
0.010370585136115551,
0.044644225388765335,
0.0400594063103199,
-0.0029673296958208084,
0.06433109939098358,
-0.02855067141354084,
-0.09656322747468948,
0.048976484686136246,
0.0007278705597855151,
0.08844880759716034,
-0.09925688803195953,
0.05390395596623421,
0.014568106271326542,
0.011620808392763138,
0.010012186132371426,
0.11196137219667435,
-0.057273514568805695,
-0.01624481752514839,
0.09574249386787415,
-0.0539921335875988,
-0.010499464347958565,
0.01416747085750103,
0.07116816192865372,
0.011651190929114819,
0.03027098812162876,
0.06217809394001961,
0.07312210649251938,
-0.022258874028921127,
0.06631051748991013,
0.036908674985170364,
-0.04850330948829651,
0.01717163622379303,
-0.08548218756914139,
0.1143922284245491,
0.05567312613129616,
0.018886884674429893,
-0.03606144338846207,
0.034881722182035446,
0.024756919592618942,
0.06646406650543213,
0.0849909856915474,
0.038233380764722824,
-0.04607456177473068,
0.031615350395441055,
-0.015694240108132362,
-0.022343449294567108,
0.03188827261328697,
-0.01225342694669962,
-0.0106884203851223,
-0.05843553692102432,
-0.02881757542490959,
-0.04245857894420624,
-0.025947246700525284,
-0.013463258743286133,
3.892583421190771e-33,
-0.03205033764243126,
0.012436884455382824,
0.007333165034651756,
-0.03321164473891258,
0.005115061067044735,
0.0023805764503777027,
-0.001799767604097724,
0.008253680542111397,
-0.023429347202181816,
0.030476700514554977,
-0.010927504859864712,
0.013607868924736977,
-0.1385660469532013,
0.05065908655524254,
0.04156649857759476,
-0.06588759273290634,
-0.009431065991520882,
0.06599873304367065,
-0.007791046518832445,
0.031090401113033295,
0.08788750320672989,
-0.08977598696947098,
-0.04939265549182892,
-0.0675244927406311,
0.013863340951502323,
0.10326787829399109,
0.055411890149116516,
0.020485684275627136,
-0.03138856962323189,
0.02724786475300789,
-0.0053507802076637745,
-0.011403625831007957,
-0.07050177454948425,
-0.023912327364087105,
-0.02170586585998535,
-0.036103423684835434,
0.012957410886883736,
0.0026081795804202557,
0.03150677680969238,
-0.05579572916030884,
-0.017640892416238785,
0.04239970073103905,
-0.000505701289512217,
-0.005446215625852346,
-0.09424281865358353,
-0.03565209358930588,
0.05998443439602852,
0.003095017047598958,
0.04178139567375183,
-0.05293820798397064,
-0.018469665199518204,
-0.047766152769327164,
-0.03558855503797531,
-0.1288243979215622,
-0.08429961651563644,
0.009446321986615658,
0.08007083833217621,
-0.020696669816970825,
-0.009903376922011375,
0.03256921097636223,
0.0196858998388052,
0.073015958070755,
-0.08129018545150757,
0.058509502559900284,
0.08456451445817947,
-0.04108722507953644,
-0.058102209120988846,
0.04803783819079399,
0.11438765376806259,
-0.005954382009804249,
-0.07245395332574844,
-0.05317692458629608,
-0.07428989559412003,
-0.09207797795534134,
0.04556256905198097,
-0.080889992415905,
0.04933274909853935,
-0.0347568541765213,
0.03141520917415619,
-0.08066963404417038,
-0.060956936329603195,
-0.08774694055318832,
-0.1046203151345253,
-0.12814679741859436,
-0.024797873571515083,
0.029969584196805954,
0.004011152777820826,
-0.05013264715671539,
-0.04940502718091011,
-0.07634717971086502,
-0.0024749867152422667,
0.048201680183410645,
0.006958845537155867,
0.0565764494240284,
0.007035525981336832,
-2.269665614731789e-33,
-0.020561018958687782,
0.07246411591768265,
0.014935645274817944,
0.12796835601329803,
0.06583797931671143,
0.015125882811844349,
-0.011276103556156158,
0.04615016654133797,
0.06070072203874588,
-0.04374166205525398,
-0.045214567333459854,
-0.06548747420310974,
-0.06957339495420456,
0.04288171976804733,
0.026854382827878,
0.032900575548410416,
0.022907765582203865,
0.04105468466877937,
-0.014160657301545143,
-0.04357322305440903,
-0.020910611376166344,
0.06305459886789322,
-0.029178915545344353,
-0.03945509344339371,
-0.06265690177679062,
-0.017587557435035706,
0.0018840074772015214,
0.09577849507331848,
0.04300446808338165,
-0.015901615843176842,
-0.037565406411886215,
0.00855196826159954,
0.05632216855883598,
0.006542164832353592,
-0.1351783573627472,
0.028166648000478745,
-0.035421621054410934,
-0.0008050297619774938,
-0.04208172485232353,
0.09605119377374649,
0.0625254288315773,
-0.029569663107395172,
-0.06065734103322029,
-0.020390935242176056,
-0.009126764722168446,
0.05802873894572258,
-0.021266277879476547,
-0.004157182294875383,
0.06254169344902039,
0.05888637155294418,
0.08638864755630493,
0.01867269165813923,
-0.0534440241754055,
0.096901535987854,
-0.014696549624204636,
0.01541929878294468,
0.018048567697405815,
-0.04454423114657402,
-0.020969295874238014,
0.034642256796360016,
-0.04052257910370827,
0.010178405791521072,
0.0721798837184906,
0.007469130679965019,
-0.03383493050932884,
-0.028356021270155907,
-0.04880763217806816,
0.12116752564907074,
0.007165681570768356,
0.07478392124176025,
-0.010623768903315067,
0.069541797041893,
0.07180121541023254,
-0.020671874284744263,
-0.04117898270487785,
-0.06877169758081436,
-0.03416433930397034,
-0.045674458146095276,
-0.0124291917309165,
-0.03561893105506897,
-0.042441174387931824,
0.018807528540492058,
-0.0297488272190094,
-0.003379018744453788,
0.05161468684673309,
-0.08060882240533829,
0.057590633630752563,
0.026919536292552948,
0.00901747215539217,
0.05120968446135521,
0.04320799931883812,
0.012030497193336487,
0.09885258227586746,
0.015148009173572063,
-0.011086712591350079,
-4.463457869974263e-8,
-0.11960702389478683,
0.03490825369954109,
0.016575295478105545,
0.07023953646421432,
0.06325028091669083,
0.025141021236777306,
0.030435191467404366,
0.07408962398767471,
-0.03566264361143112,
-0.04344360902905464,
0.07758540660142899,
-0.04884156584739685,
-0.021694347262382507,
-0.023903436958789825,
0.0405888631939888,
0.10410096496343613,
0.027778515592217445,
0.09717871993780136,
-0.013789171352982521,
-0.0026857678312808275,
0.0379948653280735,
0.003500343533232808,
0.01667526736855507,
-0.0009508149232715368,
0.021733472123742104,
-0.00475010322406888,
-0.009825794026255608,
0.05010693892836571,
0.05221261829137802,
-0.002870069583877921,
0.0017448071157559752,
0.009282410144805908,
0.03336399421095848,
-0.00938551314175129,
-0.018285807222127914,
0.09080597758293152,
-0.020571302622556686,
-0.0003348594473209232,
0.0032472044695168734,
0.05306866392493248,
-0.01632186397910118,
0.04823969677090645,
-0.07060261070728302,
0.01934201270341873,
-0.01664557494223118,
-0.052597496658563614,
-0.04259900376200676,
0.003878102172166109,
0.05902986228466034,
0.04197998344898224,
-0.0029461353551596403,
-0.04440796375274658,
-0.01745772548019886,
-0.06906769424676895,
0.05243567377328873,
0.011826744303107262,
-0.06962420046329498,
-0.0380282923579216,
-0.03601007163524628,
0.007222071290016174,
0.06527534872293472,
-0.010599149391055107,
-0.004478353541344404,
0.013668161816895008
] |
Pursuing Better Decision Boundaries for Long-Tailed Object Detection via Category Information Amount | https://openreview.net/forum?id=LW55JrLYPg | [
"Yanbiao Ma",
"Wei Dai",
"Jiayi Chen"
] | Poster | In object detection, the number of instances is commonly used to determine whether a dataset follows a long-tailed distribution, implicitly assuming that the model will perform poorly on categories with fewer instances. This assumption has led to extensive research on category bias in datasets with imbalanced instance distributions. However, even in datasets with relatively balanced instance counts, models still exhibit bias toward certain categories, indicating that instance count alone cannot explain this phenomenon. In this work, we first introduce the concept and measurement of category informativeness. We observe a significant negative correlation between a category’s informativeness and its accuracy, suggesting that informativeness more accurately reflects the learning difficulty of a category. Based on this observation, we propose the Informativeness-Guided Angular Margin Loss (IGAM Loss), which dynamically adjusts the decision space of categories according to their informativeness, thereby mitigating category bias in long-tailed datasets. IGAM Loss not only achieves superior performance on long-tailed benchmark datasets such as LVIS v1.0 and COCO-LT but also demonstrates significant improvements for underrepresented categories in non-long-tailed datasets like Pascal VOC. Extensive experiments confirm the potential of category informativeness as a tool and the generalizability of our proposed method. | Long-tailed recognition, Class Imbalanced, Image processing | null | 885 | 2502.03852 | [
0.024733951315283775,
-0.037428244948387146,
0.014896953478455544,
0.050368532538414,
0.14330102503299713,
-0.021583668887615204,
-0.038967564702034,
-0.014313007704913616,
0.0679488256573677,
-0.010536324232816696,
-0.06382561475038528,
-0.08804729580879211,
0.0036710717249661684,
0.017528565600514412,
0.034329190850257874,
-0.04690637066960335,
0.13166797161102295,
0.0838674008846283,
-0.081666000187397,
0.023596862331032753,
0.04400084540247917,
0.05610135570168495,
0.1082935631275177,
0.004215124994516373,
-0.0003374400257598609,
-0.05485100299119949,
-0.024292394518852234,
-0.06137270852923393,
-0.014723665080964565,
-0.10928861796855927,
-0.01689625158905983,
0.039246898144483566,
0.01135294884443283,
0.04535549506545067,
-0.11715187877416611,
0.020249344408512115,
-0.00537196546792984,
0.04516689479351044,
0.035234663635492325,
-0.009122177958488464,
0.044695187360048294,
0.05051626265048981,
0.04368892312049866,
0.0044122845865786076,
0.005853360518813133,
0.010525519028306007,
-0.043859366327524185,
-0.034305136650800705,
0.0331304594874382,
0.032825905829668045,
-0.03007776476442814,
0.0018575539579614997,
-0.0227675661444664,
0.08142084628343582,
-0.12069565802812576,
-0.012181232683360577,
0.059908632189035416,
-0.005927340127527714,
-0.023142686113715172,
0.05214628577232361,
0.031040720641613007,
-0.08026541024446487,
0.012140406295657158,
0.0037497791927307844,
-0.019681546837091446,
-0.053845539689064026,
-0.017111919820308685,
-0.008001615293323994,
-0.034195687621831894,
0.054804325103759766,
0.03976766765117645,
0.027823399752378464,
-0.06205971911549568,
0.10046103596687317,
0.03680838644504547,
0.03273340314626694,
0.08345955610275269,
0.00739889033138752,
0.03194756060838699,
-0.06735251843929291,
-0.020309731364250183,
0.0543217733502388,
0.00046922205365262926,
-0.013064097613096237,
0.12333092093467712,
-0.014483972452580929,
-0.02081296220421791,
0.02924218960106373,
-0.10396043956279755,
-0.018785765394568443,
-0.016552556306123734,
-0.04980910196900368,
0.07997597008943558,
0.04067719355225563,
0.05719069764018059,
-0.03522021323442459,
0.035729389637708664,
-0.026448212563991547,
-0.0164470337331295,
0.09545594453811646,
-0.02879207395017147,
0.00676728505641222,
-0.027943797409534454,
-0.04437423124909401,
-0.013851454481482506,
-0.04644203558564186,
0.005354462657123804,
-0.004006624687463045,
0.057403672486543655,
-0.0750502422451973,
0.07980253547430038,
0.03286657854914665,
-0.035629600286483765,
-0.04361049830913544,
-0.07169514149427414,
-0.015334156341850758,
-0.027468673884868622,
0.030019398778676987,
-0.017130181193351746,
0.034128107130527496,
0.004509646911174059,
-0.009480058215558529,
0.001308418344706297,
-0.00589994341135025,
0.05585363507270813,
-0.002323730615898967,
-0.07650815695524216,
2.499672757647263e-33,
0.0002080076519632712,
-0.05189109221100807,
0.0028672688640654087,
-0.05277179554104805,
0.028592154383659363,
0.018226061016321182,
-0.07590540498495102,
0.05138809606432915,
-0.0328100323677063,
0.07927335053682327,
-0.004703422077000141,
0.06163935735821724,
-0.059931255877017975,
-0.011536828242242336,
0.08800847828388214,
0.018138086423277855,
-0.04327308386564255,
0.09830441325902939,
-0.06163569539785385,
-0.0032469260040670633,
-0.04125463217496872,
-0.05536710098385811,
-0.01953522302210331,
-0.06326939910650253,
0.020058773458003998,
0.05355910211801529,
0.013830582611262798,
-0.03282355144619942,
-0.060458600521087646,
0.04873320460319519,
-0.06958091259002686,
0.020589817315340042,
0.024767084047198296,
-0.05300924926996231,
0.03289206326007843,
-0.10115011036396027,
-0.0611947663128376,
0.026029732078313828,
0.04512627050280571,
-0.028591327369213104,
-0.010752195492386818,
0.01553382072597742,
-0.04988469183444977,
0.009251240640878677,
-0.06127896159887314,
-0.022650083526968956,
0.09346508979797363,
-0.000056070010032271966,
-0.017037833109498024,
0.04611840471625328,
0.07908500730991364,
-0.03696019947528839,
-0.00804293341934681,
0.05434795841574669,
0.0022544600069522858,
0.04402897134423256,
-0.010861711576581001,
0.031479984521865845,
-0.004130642395466566,
0.00529318256303668,
-0.037876542657613754,
0.07920446991920471,
0.042142048478126526,
0.041750360280275345,
-0.05070040747523308,
0.029633350670337677,
-0.008915171958506107,
0.02262171357870102,
0.001104153343476355,
0.0639837235212326,
0.008189604617655277,
0.028369536623358727,
-0.05437110364437103,
-0.009977670386433601,
0.05171998217701912,
0.005171139724552631,
0.01631953939795494,
-0.004050915129482746,
-0.009916278533637524,
0.024501493200659752,
0.043761543929576874,
-0.003545250277966261,
0.025475680828094482,
-0.08828029036521912,
-0.080005943775177,
0.030183663591742516,
0.05122905969619751,
-0.0729844868183136,
-0.021843886002898216,
0.010742821730673313,
-0.06532131880521774,
0.0034355781972408295,
-0.08322958648204803,
0.0016852020053192973,
0.02759166806936264,
-3.1526075232962336e-33,
-0.059030964970588684,
0.004697656258940697,
-0.04979170486330986,
0.020005252212285995,
-0.020144999027252197,
-0.02990899607539177,
-0.027433719485998154,
0.046885471791028976,
-0.08752851188182831,
-0.06397727876901627,
0.011627400293946266,
0.0028562441002577543,
-0.06796283274888992,
0.009449386037886143,
0.03895650431513786,
0.020223239436745644,
-0.012522987090051174,
0.03254520893096924,
-0.027157027274370193,
-0.034721847623586655,
0.07826977968215942,
0.06622684001922607,
-0.11101046204566956,
0.0002632835239637643,
-0.018493719398975372,
0.07597516477108002,
-0.08375108987092972,
0.01898595504462719,
0.036214906722307205,
-0.0850377008318901,
-0.08188515901565552,
0.010375052690505981,
-0.04253992438316345,
-0.06238217651844025,
-0.08084139227867126,
0.005437291692942381,
0.06936238706111908,
-0.050902754068374634,
-0.0459621362388134,
0.12306801974773407,
0.056295931339263916,
-0.000441197887994349,
-0.05628158524632454,
-0.004578575491905212,
-0.0373685248196125,
-0.13944752514362335,
0.02293211594223976,
0.04455944523215294,
0.0913749486207962,
0.08638586848974228,
-0.048054859042167664,
0.0039664991199970245,
-0.04411302134394646,
0.12793533504009247,
-0.013065428473055363,
-0.05194295570254326,
-0.09578486531972885,
-0.020349452272057533,
0.003844879101961851,
0.04300183430314064,
-0.09866167604923248,
0.024984486401081085,
-0.047950297594070435,
-0.017054207623004913,
-0.014811407774686813,
-0.08939239382743835,
-0.006173545960336924,
-0.09451623260974884,
0.10809775441884995,
0.02251524105668068,
0.021432261914014816,
0.04802959784865379,
-0.042472954839468,
-0.026026833802461624,
-0.016166554763913155,
0.10109139233827591,
0.0713646411895752,
-0.03762253746390343,
0.028816724196076393,
-0.01053198054432869,
0.050610724836587906,
-0.05067184194922447,
0.03070509247481823,
0.045002929866313934,
0.08373044431209564,
0.0865602046251297,
0.000036011679185321555,
0.0020621931180357933,
0.011658832430839539,
0.0753248780965805,
-0.016802025958895683,
-0.038431186228990555,
0.02490774542093277,
0.00908665731549263,
-0.039822377264499664,
-4.2855376136685663e-8,
-0.028355058282613754,
0.03402438014745712,
-0.006004161201417446,
0.01006222702562809,
0.025874292477965355,
0.05173264071345329,
-0.09803630411624908,
0.12804079055786133,
-0.05870667099952698,
-0.030482495203614235,
0.014059069566428661,
0.03873346000909805,
-0.046463992446660995,
0.044940270483493805,
0.015417330898344517,
0.0388113297522068,
-0.01812353916466236,
0.03535329923033714,
0.018038971349596977,
0.08118999749422073,
0.049380719661712646,
-0.03777441382408142,
0.020536184310913086,
-0.043147049844264984,
-0.05980129539966583,
-0.11561271548271179,
0.020519258454442024,
0.0899285301566124,
-0.004577175714075565,
-0.019202573224902153,
-0.07937115430831909,
0.050366126000881195,
-0.06620453298091888,
0.03193574398756027,
0.11175334453582764,
0.04797864705324173,
-0.04545903578400612,
-0.019894426688551903,
0.008881242014467716,
0.002127626910805702,
0.019551517441868782,
0.08792050182819366,
-0.039269525557756424,
0.061956729739904404,
-0.009969806298613548,
0.015611435286700726,
0.06582088023424149,
-0.01559485960751772,
-0.04026249051094055,
-0.047032490372657776,
-0.018969180062413216,
-0.04660611227154732,
-0.0645761787891388,
0.06341668218374252,
-0.01623821258544922,
0.02308613993227482,
0.007302162237465382,
-0.06724460422992706,
-0.016836943104863167,
0.08362841606140137,
0.04061659425497055,
-0.07505204528570175,
-0.07638133317232132,
-0.044920872896909714
] |
ShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents | https://openreview.net/forum?id=kKILfPkhSz | [
"Haiyang SHEN",
"Yue Li",
"Desong Meng",
"Dongqi Cai",
"Sheng Qi",
"Li Zhang",
"Mengwei Xu",
"Yun Ma"
] | Poster | Recent advancements in integrating large language models (LLMs) with application programming interfaces (APIs) have gained significant interest in both academia and industry. Recent work demonstrates that these API-based agents exhibit relatively strong autonomy and planning capabilities. However, their ability to handle multi-dimensional difficulty levels, diverse task types, and real-world demands remains unknown.
In this paper, we introduce ShortcutsBench, a large-scale benchmark for the comprehensive evaluation of API-based agents in solving real-world complex tasks. ShortcutsBench includes a wealth of real APIs from Apple Inc., refined user queries, human-annotated
high-quality action sequences, detailed parameter filling values, and parameters requesting necessary input from the system or user. We put in significant effort in collecting and processing the data. We revealed how existing benchmarks / datasets struggle to accommodate the advanced reasoning capabilities of existing more intelligent LLMs. Moreover, our extensive evaluation of agents built with 5 leading open-source (size >= 57B) and 5 closed-source LLMs (e.g. Gemini-1.5-Pro and GPT-4o-mini) reveals significant limitations of existing API-based agents in the whole process of handling complex queries related to API selection, parameter filling, and requesting necessary input from the system and the user. These findings highlight the great challenges that API-based agents face in effectively fulfilling real and complex user queries. All datasets, code, experimental logs, and results are available at \url{https://anonymous.4open.science/r/ShortcutsBench}. | Benchmark, Agent, LLM, Shortcuts | null | 880 | 2407.00132 | [
-0.03741154819726944,
-0.05409716069698334,
-0.005570347420871258,
0.020401841029524803,
-0.01778588630259037,
-0.040909141302108765,
-0.016943421214818954,
0.04213995113968849,
-0.017942985519766808,
-0.022929292172193527,
-0.05754370614886284,
-0.12636691331863403,
0.04652860388159752,
0.012688059359788895,
0.07929672300815582,
0.008718598634004593,
0.08354175835847855,
-0.036558978259563446,
-0.08318539708852768,
-0.07799428701400757,
-0.003396935062482953,
0.05306624621152878,
-0.006362648215144873,
-0.0007789008668623865,
0.060481179505586624,
-0.014347842894494534,
0.01893738843500614,
-0.0700330138206482,
-0.013895494863390923,
0.02888130024075508,
0.02532907947897911,
0.002453079679980874,
-0.01162506453692913,
0.06604914367198944,
-0.01758795790374279,
0.039089299738407135,
-0.054370101541280746,
-0.03461813926696777,
-0.03171869367361069,
-0.07195570319890976,
-0.10378910601139069,
-0.015373812057077885,
-0.018313750624656677,
0.02228408306837082,
0.02550310082733631,
-0.01346508413553238,
-0.027733491733670235,
-0.028939129784703255,
0.013527422212064266,
0.07473404705524445,
-0.12803632020950317,
-0.056099265813827515,
0.0440392903983593,
-0.003153058933094144,
-0.021333619952201843,
0.000655381300020963,
-0.010048799216747284,
0.018247060477733612,
0.06395497173070908,
-0.0048004137352108955,
0.026347626000642776,
-0.13928283751010895,
-0.00606951117515564,
-0.008561295457184315,
-0.05687762051820755,
-0.021401742473244667,
0.016886664554476738,
-0.02578236162662506,
0.08162706345319748,
-0.052426133304834366,
-0.04927017167210579,
-0.00838755164295435,
0.008926667273044586,
0.025089098140597343,
-0.02676057629287243,
-0.041466839611530304,
0.0029689378570765257,
-0.025762666016817093,
0.04386896640062332,
-0.04506648704409599,
-0.03040221706032753,
-0.04390975460410118,
-0.05363350734114647,
0.15643273293972015,
0.08759085088968277,
0.023466283455491066,
0.037519462406635284,
0.07644560933113098,
0.03371340408921242,
0.06615689396858215,
0.027910331264138222,
-0.09143482148647308,
0.07747530192136765,
0.0024465948808938265,
0.06973878294229507,
0.07526685297489166,
-0.029402457177639008,
-0.09894371777772903,
-0.09371667355298996,
0.10249990224838257,
-0.06196603178977966,
0.07742635905742645,
-0.04578610882163048,
-0.060697898268699646,
-0.0011209241347387433,
0.008727877400815487,
0.04880782216787338,
-0.028561541810631752,
-0.009258914738893509,
-0.026750318706035614,
0.009180556051433086,
-0.0201454795897007,
0.08479365706443787,
0.008503281511366367,
0.05680150166153908,
0.0218804944306612,
-0.002758912043645978,
0.013867701403796673,
0.02981509454548359,
0.009649061597883701,
0.020337557420134544,
0.09330444782972336,
-0.005782932974398136,
0.03703931346535683,
0.0513126403093338,
0.0422525517642498,
0.0242107305675745,
1.9131675654861598e-33,
0.08415979892015457,
0.043869342654943466,
0.06983786076307297,
0.04994315654039383,
0.03117680363357067,
-0.06237639859318733,
0.07147550582885742,
0.11541478335857391,
-0.0524645633995533,
-0.027275731787085533,
0.023336539044976234,
0.021316776052117348,
-0.06181859225034714,
0.05236952379345894,
0.026219751685857773,
-0.04107561707496643,
-0.03865604102611542,
0.05045761913061142,
-0.08392365276813507,
-0.016042396426200867,
0.024929018691182137,
-0.012967921793460846,
0.03547309711575508,
-0.01658027246594429,
0.06990795582532883,
0.06269742548465729,
0.05903531610965729,
-0.034721750766038895,
0.01343825738877058,
-0.014408781193196774,
-0.09370551258325577,
-0.031459320336580276,
-0.020935259759426117,
0.022842658683657646,
0.012494731694459915,
0.0000837010593386367,
-0.03158469870686531,
-0.052083712071180344,
0.017190665006637573,
0.0585201233625412,
-0.08334431797266006,
0.04342176765203476,
0.018948795273900032,
-0.029798466712236404,
-0.07341591268777847,
-0.0656147450208664,
-0.04817942902445793,
0.025581780821084976,
-0.016280915588140488,
0.005807910114526749,
-0.014444501139223576,
0.06096265837550163,
0.05533711612224579,
0.01241424959152937,
-0.014620082452893257,
-0.03238682448863983,
0.028781728819012642,
0.03854571282863617,
0.007995327934622765,
0.09978697448968887,
-0.041981618851423264,
0.034854184836149216,
0.010356992483139038,
0.13664355874061584,
-0.012956654652953148,
0.0480867438018322,
-0.02620924450457096,
0.12760476768016815,
0.0775618776679039,
0.02550383470952511,
-0.013992074877023697,
0.01524755172431469,
0.06993944942951202,
-0.04103255644440651,
0.03761804848909378,
-0.005896619521081448,
0.040963515639305115,
-0.1525132805109024,
0.006866344716399908,
-0.043027717620134354,
0.0018584838835522532,
0.013827469199895859,
-0.013676367700099945,
-0.01273062638938427,
0.011368779465556145,
-0.02972269058227539,
-0.011934859678149223,
-0.08784302324056625,
-0.03300836309790611,
0.020674392580986023,
-0.0894913449883461,
0.027119675651192665,
-0.10389389842748642,
0.03416609764099121,
-0.02889917604625225,
-1.6285137105818733e-33,
0.021246779710054398,
-0.13312307000160217,
-0.0683690533041954,
0.0619574673473835,
-0.01294278260320425,
-0.09358890354633331,
-0.029597170650959015,
-0.0013257015962153673,
0.005499010439962149,
-0.03193219006061554,
-0.06373566389083862,
0.006682842504233122,
0.04327905550599098,
-0.02234775573015213,
0.051302872598171234,
-0.03682452440261841,
-0.05027639865875244,
-0.05719828978180885,
0.024449892342090607,
0.021860938519239426,
-0.07725020498037338,
0.08177872002124786,
-0.07507939636707306,
-0.025944389402866364,
-0.06470692902803421,
0.008859953843057156,
-0.017605112865567207,
-0.006619595922529697,
-0.010610589757561684,
0.023048289120197296,
0.005675518419593573,
-0.004426281433552504,
-0.07806995511054993,
-0.05417685583233833,
-0.021162236109375954,
0.06648208200931549,
0.009500687010586262,
0.06872352212667465,
-0.04779792204499245,
0.04047997295856476,
0.0768703743815422,
-0.08572718501091003,
-0.04675034061074257,
-0.009768146090209484,
0.006490995176136494,
0.13609546422958374,
-0.12421559542417526,
0.0704173892736435,
-0.03792740777134895,
-0.027631565928459167,
0.0013647412415593863,
0.004857284482568502,
-0.04923249036073685,
-0.01890200935304165,
-0.00721860071644187,
-0.057454027235507965,
-0.0750647559762001,
-0.05012742802500725,
-0.01377094630151987,
-0.012120820581912994,
-0.09187620878219604,
0.008200895972549915,
0.07328153401613235,
0.016328314319252968,
0.007462520617991686,
-0.03212480992078781,
-0.026891184970736504,
-0.01514028012752533,
-0.03304724767804146,
0.0018506990745663643,
0.10781838744878769,
-0.08856719732284546,
-0.0028241050895303488,
0.01691206358373165,
0.03737993910908699,
0.05081995949149132,
0.013820667751133442,
-0.026769401505589485,
-0.005294309463351965,
0.010071779601275921,
0.060758862644433975,
0.08039756119251251,
0.08497762680053711,
0.05076472461223602,
0.00026771757984533906,
0.09370281547307968,
0.028535019606351852,
0.11774648725986481,
-0.013618320226669312,
0.0882425382733345,
-0.047014132142066956,
0.049199532717466354,
0.002039004350081086,
0.01702493615448475,
-0.10268426686525345,
-4.1719154353359045e-8,
-0.020060013979673386,
0.06541900336742401,
-0.011591736227273941,
0.01226078812032938,
-0.03372129052877426,
-0.020047206431627274,
-0.05186167731881142,
0.11854663491249084,
0.030078710988163948,
0.06420346349477768,
0.06635352969169617,
-0.04687373712658882,
-0.046901408582925797,
0.01725180260837078,
-0.019066978245973587,
0.04348175972700119,
0.001095149083994329,
0.04272490367293358,
-0.016700657084584236,
-0.058372125029563904,
0.027225645259022713,
0.0642622709274292,
-0.018909713253378868,
-0.029522407799959183,
0.018102647736668587,
-0.052684009075164795,
-0.04742630571126938,
0.11584274470806122,
0.04445812106132507,
0.033231958746910095,
0.011862008832395077,
-0.007013606373220682,
0.05687265843153,
-0.021183498203754425,
0.05669287592172623,
0.003932194318622351,
-0.01424072403460741,
0.0051101986318826675,
0.020207472145557404,
-0.03348846733570099,
-0.007307705003768206,
0.022743579000234604,
-0.024848897010087967,
-0.022307369858026505,
-0.006298433989286423,
0.002291648415848613,
-0.051595643162727356,
-0.06234794110059738,
0.05925703048706055,
-0.05623345822095871,
-0.04077376425266266,
-0.04432637244462967,
0.0035459487698972225,
0.0605485737323761,
0.022070325911045074,
0.07897541671991348,
-0.045211032032966614,
-0.05505739897489548,
0.07109920680522919,
0.022175859659910202,
0.03368493914604187,
0.0053592403419315815,
-0.028630778193473816,
0.030951356515288353
] |
Sitcom-Crafter: A Plot-Driven Human Motion Generation System in 3D Scenes | https://openreview.net/forum?id=FvIASa0tau | [
"Jianqi Chen",
"Panwen Hu",
"Xiaojun Chang",
"Zhenwei Shi",
"Michael Kampffmeyer",
"Xiaodan Liang"
] | Poster | Recent advancements in human motion synthesis have focused on specific types of motions, such as human-scene interaction, locomotion or human-human interaction, however, there is a lack of a unified system capable of generating a diverse combination of motion types. In response, we introduce *Sitcom-Crafter*, a comprehensive and extendable system for human motion generation in 3D space, which can be guided by extensive plot contexts to enhance workflow efficiency for anime and game designers. The system is comprised of eight modules, three of which are dedicated to motion generation, while the remaining five are augmentation modules that ensure consistent fusion of motion sequences and system functionality. Central to the generation modules is our novel 3D scene-aware human-human interaction module, which addresses collision issues by synthesizing implicit 3D Signed Distance Function (SDF) points around motion spaces, thereby minimizing human-scene collisions without additional data collection costs. Complementing this, our locomotion and human-scene interaction modules leverage existing methods to enrich the system's motion generation capabilities. Augmentation modules encompass plot comprehension for command generation, motion synchronization for seamless integration of different motion types, hand pose retrieval to enhance motion realism, motion collision revision to prevent human collisions, and 3D retargeting to ensure visual fidelity. Experimental evaluations validate the system's ability to generate high-quality, diverse, and physically realistic motions, underscoring its potential for advancing creative workflows. Code and demonstration videos can be found in the supplementary files. | Human Motion Synthesis, Human-Human Interaction, Physically Compliant Motion, Creative Workflow Automation | Sitcom-Crafter is a system for generating diverse, physically compliant 3D human motions across multiple motion types, guided by plot contexts to streamline creative workflows in anime and game design. | 870 | null | [
-0.03671484440565109,
-0.09748546779155731,
-0.0047392528504133224,
-0.08179821074008942,
0.03821412846446037,
0.02398017980158329,
0.013489860109984875,
-0.021792631596326828,
0.04968396946787834,
-0.030212590470910072,
0.020413286983966827,
-0.10204780101776123,
-0.06540029495954514,
0.006370420102030039,
0.05330174043774605,
-0.024998683482408524,
0.062186360359191895,
0.07806617021560669,
-0.07859964668750763,
0.06753641366958618,
0.036620669066905975,
-0.011868567205965519,
0.015603647567331791,
-0.024584006518125534,
0.002285712631419301,
-0.02945392206311226,
0.0636550560593605,
-0.09266328066587448,
-0.014411680400371552,
-0.043492235243320465,
0.0853869691491127,
0.023057540878653526,
0.055197279900312424,
-0.03539162129163742,
-0.011407246813178062,
0.06222879886627197,
0.06035520136356354,
-0.004780142568051815,
-0.07742661237716675,
-0.03984517976641655,
-0.09252583980560303,
0.017023881897330284,
0.06293845921754837,
0.04595049470663071,
0.05485767126083374,
0.09289930760860443,
-0.061153534799814224,
-0.08229966461658478,
0.04303926229476929,
0.08851085603237152,
-0.07869633287191391,
0.005139156710356474,
0.029778914526104927,
0.030013712123036385,
0.060101594775915146,
0.05503444746136665,
0.07367946207523346,
-0.018481262028217316,
0.08458533138036728,
-0.06314214318990707,
0.0030528760980814695,
-0.029672468081116676,
0.058032505214214325,
-0.00705692870542407,
-0.035398930311203,
-0.11736945062875748,
0.04583961144089699,
-0.01583877205848694,
0.003924912307411432,
0.03959638997912407,
-0.09916367381811142,
-0.049396831542253494,
0.02780666947364807,
0.00962358620017767,
-0.033251117914915085,
0.05568021535873413,
-0.00107092026155442,
-0.11180004477500916,
-0.028900984674692154,
-0.05149029195308685,
0.02362094447016716,
-0.061225421726703644,
-0.01220554206520319,
0.0461253896355629,
-0.015136618167161942,
0.06363601237535477,
-0.00946817547082901,
0.09050045907497406,
-0.015862489119172096,
0.07135457545518875,
-0.09888963401317596,
-0.0004362619947642088,
0.04787125438451767,
-0.05261141434311867,
-0.0029841172508895397,
-0.011615375056862831,
-0.029126029461622238,
-0.05664770305156708,
0.03408113121986389,
0.01086998451501131,
-0.024778662249445915,
-0.03807636350393295,
0.04136144742369652,
0.009863313287496567,
0.06307648122310638,
0.03807522729039192,
0.039070338010787964,
-0.017350997775793076,
0.00831367913633585,
0.07244086265563965,
0.0027044746093451977,
-0.08490446954965591,
0.010651582852005959,
0.04661744087934494,
0.0420573465526104,
0.019315119832754135,
-0.012464410625398159,
0.010341216810047626,
0.05295908451080322,
0.02595466747879982,
0.09029953926801682,
0.015220776200294495,
0.0053032273426651955,
0.028290193527936935,
0.02569224126636982,
-0.07737376540899277,
0.0017482356633991003,
1.3275830012151666e-34,
-0.0022244679275900126,
0.08441360294818878,
0.03631690889596939,
0.07254572957754135,
0.03064369410276413,
-0.019367694854736328,
-0.03309419006109238,
0.007450244855135679,
-0.03896397724747658,
0.03490477427840233,
-0.05246717482805252,
-0.0012458923738449812,
-0.05707123875617981,
0.11232439428567886,
0.02373524382710457,
-0.03414028510451317,
0.025800181552767754,
-0.0497858040034771,
-0.05020660534501076,
0.10510902851819992,
0.03419676050543785,
-0.05414728820323944,
-0.03727466240525246,
-0.014565753750503063,
-0.0537889190018177,
0.03386465460062027,
0.05599576234817505,
0.003132184734568,
0.04223696514964104,
-0.0442475751042366,
-0.10823161154985428,
0.026760825887322426,
0.014822850935161114,
0.010220807045698166,
0.04333178326487541,
-0.030904246494174004,
-0.03556405007839203,
-0.08311257511377335,
-0.0314403735101223,
0.08155113458633423,
-0.03496578335762024,
-0.0004056443285662681,
-0.08792033791542053,
-0.025202911347150803,
-0.03670172765851021,
-0.02897258847951889,
0.1304539442062378,
0.12754735350608826,
-0.053996749222278595,
0.03771290183067322,
0.0174656193703413,
0.04852121323347092,
0.027315637096762657,
-0.08588511496782303,
-0.027498649433255196,
-0.09698686748743057,
0.0023388515692204237,
-0.07667406648397446,
0.04832754284143448,
0.09364700317382812,
-0.08776281028985977,
0.024849096313118935,
0.039034463465213776,
0.01424339972436428,
0.02591761201620102,
-0.006368633825331926,
0.060736022889614105,
0.07650112360715866,
0.06854274868965149,
0.004145907238125801,
-0.06420556455850601,
0.0008262234623543918,
0.044619105756282806,
0.02105719782412052,
0.004064809065312147,
-0.020195525139570236,
-0.032103635370731354,
-0.05347824469208717,
-0.015672853216528893,
-0.005251302849501371,
-0.10941066592931747,
-0.0307230893522501,
-0.04420079290866852,
0.01827191933989525,
0.023048657923936844,
-0.08985264599323273,
0.0008920435211621225,
0.040235232561826706,
-0.0772756040096283,
-0.09247664362192154,
0.0037306915037333965,
-0.12792609632015228,
-0.045836109668016434,
0.053024958819150925,
0.017913207411766052,
-1.2471357723881134e-33,
0.08208932727575302,
0.01878206804394722,
-0.08136545121669769,
0.01152089424431324,
0.030445098876953125,
-0.020306698977947235,
0.0058714719489216805,
-0.057121168822050095,
0.009106592275202274,
-0.018919048830866814,
-0.053086478263139725,
-0.00608163932338357,
0.04609069228172302,
-0.029116826131939888,
0.02401021122932434,
0.00925882626324892,
0.010107187554240227,
-0.038516778498888016,
-0.06256595253944397,
-0.02318168617784977,
0.07993859797716141,
0.03986164927482605,
-0.02882765419781208,
-0.027436379343271255,
-0.009107799269258976,
0.06496619433164597,
0.011371327564120293,
-0.01362501922994852,
-0.00425252178683877,
0.07802136987447739,
0.002609116956591606,
-0.013912963680922985,
0.015188432298600674,
-0.06650684773921967,
0.021649101749062538,
0.015601018443703651,
0.06522876769304276,
0.023958029225468636,
-0.024321306496858597,
-0.05198878049850464,
0.02565689943730831,
-0.005386870354413986,
-0.048145268112421036,
0.10692001134157181,
-0.025039318948984146,
0.09205131232738495,
-0.06567542254924774,
0.026972666382789612,
-0.041630569845438004,
0.026614343747496605,
-0.09093990921974182,
0.026157716289162636,
-0.05545917898416519,
-0.14943276345729828,
-0.008985289372503757,
-0.06921124458312988,
-0.0017280393512919545,
-0.03463555872440338,
0.05683165416121483,
-0.05956222489476204,
-0.008025087416172028,
-0.005789526272565126,
0.030261706560850143,
0.02624722756445408,
-0.014921454712748528,
0.07347757369279861,
-0.004034470766782761,
-0.024235155433416367,
0.008304058574140072,
0.04736853763461113,
0.03866700828075409,
0.027143850922584534,
0.05348072946071625,
0.010204510763287544,
0.016759393736720085,
-0.050991881638765335,
0.02818557620048523,
-0.0433635488152504,
0.1110881119966507,
-0.053076788783073425,
0.007996316067874432,
-0.05177963152527809,
0.0689174011349678,
0.08974958211183548,
-0.03972373530268669,
0.047363847494125366,
-0.10135477036237717,
0.14509505033493042,
0.00022772887314204127,
0.042204599827528,
0.0034904885105788708,
0.08742275834083557,
-0.017922289669513702,
0.05439193919301033,
-0.07541938871145248,
-3.9342996416280585e-8,
-0.10432488471269608,
0.01743929274380207,
-0.007267980370670557,
-0.0397656075656414,
-0.05573274940252304,
0.038483455777168274,
0.0073314751498401165,
0.05496492236852646,
0.029664570465683937,
-0.041349537670612335,
0.03689468279480934,
-0.007472238037735224,
0.07496324926614761,
-0.004230122547596693,
0.03528235852718353,
0.027329573407769203,
0.01244437787681818,
0.039015818387269974,
-0.03672868013381958,
0.056175295263528824,
0.00005180754669709131,
0.04558683931827545,
-0.017922120168805122,
0.029496412724256516,
-0.039035044610500336,
-0.013701211661100388,
-0.04038558900356293,
0.028920728713274002,
0.03162739425897598,
-0.034707825630903244,
0.056947845965623856,
-0.043073609471321106,
0.0007478497573174536,
-0.021949222311377525,
-0.026474257931113243,
0.012427960522472858,
-0.03594324737787247,
-0.011373783461749554,
0.0030167021322995424,
-0.0025520429480820894,
-0.02267681434750557,
0.08223123848438263,
-0.01973666064441204,
0.013622370548546314,
0.013981329277157784,
-0.022166773676872253,
-0.020019272342324257,
-0.1287522166967392,
-0.029135745018720627,
-0.07798994332551956,
-0.051359448581933975,
-0.06699036061763763,
-0.005885032005608082,
0.042655881494283676,
0.06176065281033516,
0.10245447605848312,
-0.03003758005797863,
0.030465058982372284,
0.0872388556599617,
0.009344116784632206,
-0.0522289052605629,
-0.011485817842185497,
-0.019256753847002983,
-0.04481805860996246
] |
Gap Preserving Distillation by Building Bidirectional Mappings with A Dynamic Teacher | https://openreview.net/forum?id=PnfghHD4Pi | [
"Yong Guo",
"Shulian Zhang",
"Haolin Pan",
"Jing Liu",
"Yulun Zhang",
"Jian Chen"
] | Poster | Knowledge distillation aims to transfer knowledge from a large teacher model to a compact student counterpart, often coming with a significant performance gap between them. Interestingly, we find that a too-large performance gap can hamper the training process.
To alleviate this, we propose a **Gap Preserving Distillation (GPD)** method that trains an additional dynamic teacher model from scratch along with the student to maintain a reasonable performance gap. To further strengthen distillation, we develop a hard strategy by enforcing both models to share parameters. Besides, we also build the soft bidirectional mappings between them through ***Inverse Reparameterization (IR)*** and ***Channel-Branch Reparameterization (CBR)***.
IR initializes a larger dynamic teacher with approximately the same accuracy as the student to avoid a too large gap in early stage of training. CBR enables direct extraction of an effective student model from the dynamic teacher without post-training.
In experiments, GPD significantly outperforms existing distillation methods on top of both CNNs and transformers, achieving up to 1.58\% accuracy improvement.
Interestingly, GPD also generalizes well to the scenarios without a pre-trained teacher, including training from scratch and fine-tuning, yielding a large improvement of 1.80\% and 0.89\% on ResNet18, respectively. | Knowledge Distillation; Model Expansion; Reparameterization | null | 869 | 2410.04140 | [
-0.05834616348147392,
-0.11585227400064468,
0.05873619019985199,
0.02368765138089657,
-0.08686280250549316,
-0.06878849118947983,
-0.04307736083865166,
-0.0451468862593174,
-0.002572326920926571,
-0.05593336001038551,
-0.01144809927791357,
0.03732986003160477,
0.013277396559715271,
0.055932365357875824,
-0.051533911377191544,
0.041790880262851715,
0.016131946817040443,
0.1541801244020462,
-0.1283172070980072,
-0.05109616741538048,
0.01869867369532585,
0.029910564422607422,
-0.05397414788603783,
0.027468547224998474,
0.06290168315172195,
0.016113698482513428,
0.038696642965078354,
-0.054020725190639496,
0.03874826431274414,
-0.08176003396511078,
-0.001582263852469623,
-0.08288700133562088,
-0.10032659024000168,
0.010705770924687386,
-0.05804930999875069,
0.07247056066989899,
0.023499777540564537,
0.1082373633980751,
-0.01788819395005703,
-0.004618445411324501,
-0.004417737480252981,
0.04479387402534485,
0.02678314782679081,
0.054550595581531525,
0.07527483999729156,
0.03598138689994812,
0.020902803167700768,
-0.03031359612941742,
0.018306439742445946,
-0.01127718947827816,
-0.0334300771355629,
-0.020413324236869812,
-0.017794765532016754,
0.10155998170375824,
-0.07179778814315796,
0.07513093203306198,
0.061573952436447144,
0.0022143356036394835,
-0.00964782852679491,
0.00858782883733511,
-0.07413995265960693,
0.0026140843983739614,
-0.02736341580748558,
-0.03430498018860817,
-0.07070372998714447,
-0.015602121129631996,
0.016467705368995667,
0.11074068397283554,
0.07512657344341278,
0.005362272262573242,
-0.04253033548593521,
0.08380720764398575,
-0.010377686470746994,
-0.03382744640111923,
0.11777974665164948,
-0.012077132239937782,
0.03599397465586662,
0.035379212349653244,
0.10165949910879135,
-0.06606091558933258,
0.07992812246084213,
-0.04966539517045021,
0.04391428828239441,
-0.05575380474328995,
0.07573360949754715,
0.02792564406991005,
0.01087818294763565,
-0.009341899305582047,
0.043343134224414825,
-0.038901813328266144,
-0.0722053274512291,
0.007544044405221939,
-0.01729780063033104,
-0.007763467729091644,
0.04625595733523369,
-0.03344469517469406,
0.019963989034295082,
-0.024334164336323738,
-0.004612835124135017,
0.0943768322467804,
-0.045606065541505814,
0.05073878541588783,
0.02130468375980854,
-0.05056152492761612,
-0.03462185338139534,
-0.05582138150930405,
0.06670434027910233,
0.0453680194914341,
0.011773948557674885,
-0.05563441291451454,
0.08168316632509232,
0.04097752645611763,
0.004905196372419596,
0.03195095434784889,
0.009494850412011147,
0.03146427869796753,
0.0027871334459632635,
-0.05279993265867233,
-0.029859555885195732,
0.012042446993291378,
-0.05149513855576515,
-0.013409835286438465,
-0.043434347957372665,
0.0007808554219081998,
-0.00153900939039886,
-0.05736203491687775,
-0.07048311829566956,
1.2810870798112343e-33,
0.057208351790905,
0.08124832808971405,
-0.02030576579272747,
0.031716614961624146,
-0.04841795191168785,
0.028883211314678192,
0.003572270041331649,
0.0205141082406044,
-0.028070420026779175,
0.01604873687028885,
-0.0839531347155571,
-0.003575650043785572,
-0.06668645888566971,
0.039672207087278366,
0.034716393798589706,
-0.009944540448486805,
-0.0574227049946785,
0.05061783269047737,
-0.09726814925670624,
-0.005922326352447271,
0.020963018760085106,
0.007612182758748531,
-0.04704337194561958,
-0.07036761939525604,
-0.039489779621362686,
0.009800623171031475,
0.048388466238975525,
0.04636646807193756,
-0.026158617809414864,
0.010661142878234386,
-0.03235536441206932,
-0.02092066779732704,
-0.03351776674389839,
0.009563594125211239,
0.006053800228983164,
-0.04665391892194748,
0.040413811802864075,
-0.03856774792075157,
0.0323886014521122,
-0.02996518276631832,
-0.023525463417172432,
0.03263022005558014,
0.012904378585517406,
-0.034137528389692307,
-0.06477488577365875,
-0.07724355906248093,
0.08294273912906647,
0.013533540070056915,
0.01140572689473629,
0.0141545245423913,
0.0051892888732254505,
0.03472364693880081,
-0.11061488091945648,
-0.10999510437250137,
-0.07857323437929153,
-0.052363909780979156,
0.0318242646753788,
0.08438880741596222,
0.03565793111920357,
0.04017816483974457,
-0.05700778588652611,
0.05465652421116829,
-0.041363898664712906,
0.09723929315805435,
0.028116120025515556,
0.00480178976431489,
-0.03011486306786537,
0.048419855535030365,
0.0676216259598732,
-0.04660777747631073,
-0.11009200662374496,
-0.10635798424482346,
-0.037533022463321686,
-0.10434965044260025,
0.06619801372289658,
-0.06380246579647064,
0.05955728143453598,
-0.057093776762485504,
0.03468651697039604,
-0.017741579562425613,
0.004585970658808947,
-0.023193340748548508,
-0.05571949854493141,
-0.027927827090024948,
-0.060078214854002,
-0.005478948820382357,
0.008483339101076126,
-0.055582135915756226,
-0.004937003832310438,
-0.021269937977194786,
-0.03747214004397392,
0.016946708783507347,
-0.03456377238035202,
0.05420827865600586,
0.08898813277482986,
-1.061405369070788e-33,
0.044173263013362885,
0.04861908033490181,
-0.014500931836664677,
0.12857472896575928,
0.08307784795761108,
-0.04362967982888222,
0.0538746602833271,
0.05652693286538124,
-0.027581071481108665,
-0.08982115238904953,
-0.0003599377232603729,
-0.04682837426662445,
-0.060488101094961166,
0.03414777293801308,
-0.015994422137737274,
0.014469072222709656,
-0.041644491255283356,
0.02378835156559944,
0.03874388337135315,
0.03776527941226959,
0.04656529426574707,
0.10802307724952698,
-0.07352424412965775,
0.025146078318357468,
-0.06227570027112961,
0.025671256706118584,
-0.024275997653603554,
0.041182052344083786,
0.025126850232481956,
0.0388156957924366,
-0.03523887321352959,
-0.030650105327367783,
0.02229539304971695,
-0.057629190385341644,
-0.11394266039133072,
0.0533599928021431,
-0.04575889930129051,
-0.028980853036046028,
-0.03221346065402031,
0.09223229438066483,
0.007734929211437702,
-0.04150122404098511,
-0.05442090705037117,
-0.03600363805890083,
-0.014297747984528542,
0.05828240141272545,
-0.009461370296776295,
-0.018124504014849663,
-0.034951064735651016,
0.09250061959028244,
0.05062326416373253,
0.040420182049274445,
-0.10374713689088821,
-0.033183760941028595,
0.007118872832506895,
-0.0537700280547142,
0.025146780535578728,
-0.002832278609275818,
-0.013952179811894894,
0.06063777208328247,
-0.052335627377033234,
-0.09252754598855972,
0.11363960802555084,
-0.05702121555805206,
0.009268918074667454,
0.011522111482918262,
0.011071373708546162,
0.07718541473150253,
0.03501269221305847,
0.06495855748653412,
-0.020825406536459923,
0.04518447071313858,
0.05643777549266815,
-0.013279435224831104,
0.007459812797605991,
-0.02301189675927162,
-0.040495213121175766,
-0.050170931965112686,
-0.023747652769088745,
0.009892051108181477,
-0.061060983687639236,
0.006054517813026905,
-0.0060647097416222095,
0.07434918731451035,
0.0841713473200798,
0.04026423394680023,
0.02001028321683407,
0.020519891753792763,
0.03584183752536774,
0.03250931575894356,
0.008355927653610706,
0.042045194655656815,
0.08855980634689331,
-0.0030850162729620934,
0.016597304493188858,
-4.452085278217055e-8,
-0.0960046648979187,
0.10443704575300217,
0.0318877175450325,
0.11115313321352005,
-0.026977546513080597,
-0.06345553696155548,
0.041211389005184174,
0.09284597635269165,
-0.022822508588433266,
-0.003354755463078618,
0.08147536218166351,
-0.023121224716305733,
0.0055390591733157635,
-0.04485712945461273,
0.061601679772138596,
0.0782136619091034,
0.030031727626919746,
0.08331376314163208,
-0.043684616684913635,
0.004789529833942652,
0.03436410427093506,
-0.02872343920171261,
0.0009780284017324448,
0.0021526652853935957,
0.04293563961982727,
-0.09173427522182465,
-0.0027508093044161797,
0.12362110614776611,
0.016121825203299522,
-0.03254598751664162,
0.042876213788986206,
-0.006394130177795887,
0.03529351204633713,
-0.041595879942178726,
-0.05564344301819801,
0.07310086488723755,
0.010310537181794643,
0.010667783208191395,
-0.06051564961671829,
0.004734744783490896,
-0.0018717217026278377,
0.02255481481552124,
-0.03399123623967171,
0.032417308539152145,
-0.004996407777070999,
0.012994090095162392,
0.009359053336083889,
-0.036466486752033234,
0.05447578802704811,
0.05825019255280495,
0.03461245447397232,
-0.033131688833236694,
-0.006473204120993614,
-0.05067262426018715,
0.048764634877443314,
0.03881623595952988,
-0.03672019764780998,
-0.05328800901770592,
-0.03476051613688469,
-0.016117755323648453,
-0.018590521067380905,
0.00901743583381176,
-0.06395112723112106,
-0.01450278889387846
] |
Bridging Information Asymmetry in Text-video Retrieval: A Data-centric Approach | https://openreview.net/forum?id=Tn6lrFbiP4 | [
"Zechen Bai",
"Tianjun Xiao",
"Tong He",
"Pichao WANG",
"Zheng Zhang",
"Thomas Brox",
"Mike Zheng Shou"
] | Poster | As online video content rapidly grows, the task of text-video retrieval (TVR) becomes increasingly important. A key challenge in TVR is the information asymmetry between video and text: videos are inherently richer in information, while their textual descriptions often capture only fragments of this complexity. This paper introduces a novel, data-centric framework to bridge this gap by enriching textual representations to better match the richness of video content. During training, videos are segmented into event-level clips and captioned to ensure comprehensive coverage. During retrieval, a large language model (LLM) generates semantically diverse queries to capture a broader range of possible matches. To enhance retrieval efficiency, we propose a query selection mechanism that identifies the most relevant and diverse queries, reducing computational cost while improving accuracy. Our method achieves state-of-the-art results across multiple benchmarks, demonstrating the power of data-centric approaches in addressing information asymmetry in TVR. This work paves the way for new research focused on leveraging data to improve cross-modal retrieval. | Text-video Retrieval, Vision-Language Model, Multimodal | We introduce a data-centric approach that bridges the inherent information asymmetry in the text-video video task and achieves state-of-the-art performance. | 862 | 2408.07249 | [
-0.009169511497020721,
-0.12437310069799423,
0.024458741769194603,
0.0075361160561442375,
0.12341796606779099,
0.061111971735954285,
-0.04209289699792862,
0.021955888718366623,
0.06390959024429321,
-0.12954267859458923,
-0.03882749006152153,
-0.05224896967411041,
-0.007587673608213663,
0.08202018588781357,
0.01140843890607357,
-0.006891760043799877,
0.1280931830406189,
0.03694862499833107,
-0.1296113133430481,
-0.05638457462191582,
0.0411800816655159,
0.010508286766707897,
0.07439857721328735,
-0.008859728462994099,
0.039649225771427155,
-0.008709115907549858,
0.01139124482870102,
0.022156469523906708,
0.02937043085694313,
-0.0443093404173851,
-0.002829760080203414,
0.05125775188207626,
-0.05431762710213661,
0.04973556101322174,
-0.04048829525709152,
0.0004618493840098381,
-0.09882368892431259,
-0.014442911371588707,
0.010417811572551727,
-0.033172715455293655,
0.022622214630246162,
0.007192269433289766,
0.0716513991355896,
0.01630569063127041,
-0.0044744121842086315,
0.03239365667104721,
-0.0763244777917862,
-0.05908951535820961,
0.05876662954688072,
0.04651226848363876,
-0.11272264271974564,
0.045324597507715225,
-0.006890347693115473,
0.05192290246486664,
-0.09608576446771622,
-0.007616715505719185,
0.00680009787902236,
-0.022860849276185036,
0.026087598875164986,
-0.03815795108675957,
0.0028990996070206165,
-0.11139664798974991,
0.012664413079619408,
0.046801116317510605,
-0.0017476645298302174,
-0.058172792196273804,
0.05660894140601158,
-0.021840112283825874,
-0.03311578184366226,
-0.008743318729102612,
-0.06597241759300232,
0.043347764760255814,
-0.005402096547186375,
0.13341937959194183,
-0.016634531319141388,
0.030063826590776443,
0.016739903017878532,
-0.07252030074596405,
0.00399717828258872,
-0.03450648859143257,
0.022390002384781837,
0.008388697169721127,
0.10764625668525696,
0.0397978238761425,
0.15822139382362366,
-0.05539485812187195,
0.021049922332167625,
-0.08201643824577332,
0.004951469600200653,
0.01881088688969612,
-0.14039810001850128,
-0.0008040990214794874,
0.09077982604503632,
-0.032628078013658524,
0.04648630693554878,
0.059124305844306946,
0.0018575226422399282,
-0.059858083724975586,
0.02857021614909172,
0.0521034337580204,
0.03803305700421333,
0.04012232646346092,
0.0014508296735584736,
0.008812769316136837,
-0.0037365297321230173,
0.0023004242684692144,
0.0093238465487957,
0.05365019291639328,
0.014910340309143066,
-0.0859571248292923,
0.027117973193526268,
0.05162837356328964,
0.033062588423490524,
-0.05598137900233269,
0.054039131850004196,
0.024594655260443687,
-0.02862001210451126,
-0.025633549317717552,
0.01449333131313324,
0.03833047300577164,
0.02748475968837738,
0.024121029302477837,
-0.052549656480550766,
-0.055931493639945984,
0.01649629697203636,
-0.02063959278166294,
0.0187844131141901,
3.5368935243113324e-34,
0.05312729254364967,
-0.01140106376260519,
0.0592348650097847,
0.044892068952322006,
0.013187260366976261,
-0.027135977521538734,
-0.025817938148975372,
0.11839024722576141,
-0.07143654674291611,
-0.03268832713365555,
-0.0011146299075335264,
-0.006733136251568794,
0.02819824405014515,
-0.015434416010975838,
-0.004144883714616299,
-0.01014056708663702,
-0.13059358298778534,
0.0043502263724803925,
-0.01929217204451561,
0.024923471733927727,
0.0327613539993763,
0.011238482780754566,
0.05356982722878456,
0.043174441903829575,
0.007298358250409365,
-0.03922944515943527,
0.06350209563970566,
-0.09674923121929169,
-0.02999212220311165,
-0.021615084260702133,
-0.07561896741390228,
-0.0302291139960289,
0.08413441479206085,
-0.045262813568115234,
0.12160096317529678,
-0.005556765478104353,
-0.05850223824381828,
-0.03974199667572975,
-0.0011780356289818883,
0.018039004877209663,
-0.03336220607161522,
-0.0012711237650364637,
-0.07146702706813812,
-0.06244805082678795,
-0.06945089995861053,
-0.08454632014036179,
-0.026491427794098854,
0.010693922638893127,
-0.062002357095479965,
-0.03271952643990517,
0.10441737622022629,
0.046774402260780334,
-0.01393209770321846,
-0.032876841723918915,
0.06241697445511818,
-0.004881440661847591,
0.03973502293229103,
0.022260690107941628,
0.017152873799204826,
0.010187724605202675,
0.002559021348133683,
-0.015549608506262302,
0.04942876100540161,
0.07301858067512512,
0.036926478147506714,
0.012348543852567673,
0.044318534433841705,
0.06104562431573868,
0.07339410483837128,
0.009926719591021538,
-0.009027834981679916,
0.002658929442986846,
0.018343500792980194,
-0.053755804896354675,
0.03520644083619118,
0.050144363194704056,
0.0028349892236292362,
-0.06505173444747925,
0.057112790644168854,
0.025911089032888412,
0.013371598906815052,
-0.07108794897794724,
0.05671541392803192,
-0.08327902853488922,
-0.10171204060316086,
-0.06366932392120361,
0.050315260887145996,
-0.12032284587621689,
0.0440329946577549,
-0.022400593385100365,
0.00008081280248006806,
0.009246822446584702,
-0.06105348840355873,
-0.03867616131901741,
0.06955791264772415,
6.868255993978715e-34,
0.022636884823441505,
-0.027560211718082428,
-0.008586299605667591,
0.030362574383616447,
0.0338469035923481,
-0.042724207043647766,
0.04406522959470749,
-0.0000333446660079062,
0.0037412475794553757,
-0.0504460446536541,
0.005992972292006016,
-0.061296675354242325,
-0.010311647318303585,
-0.07176509499549866,
0.011588080786168575,
0.054623913019895554,
0.05242839455604553,
0.014639987610280514,
-0.11912035197019577,
0.07407248020172119,
0.011958617717027664,
0.07203923910856247,
-0.060134731233119965,
0.027499381452798843,
0.0019071806455031037,
0.03703523799777031,
-0.03197740763425827,
0.011967110447585583,
0.027456233277916908,
-0.00119278475176543,
-0.0009987547527998686,
-0.03911307826638222,
-0.029150117188692093,
-0.06154773756861687,
-0.09867032617330551,
0.08366305381059647,
0.08202633261680603,
-0.05147306248545647,
-0.060343045741319656,
0.08758922666311264,
0.06125238537788391,
-0.011862345971167088,
-0.05558266118168831,
0.002190383616834879,
-0.024045724421739578,
-0.07397127151489258,
-0.14358489215373993,
0.0003709729644469917,
-0.0016059945337474346,
-0.02316012978553772,
-0.0018443804001435637,
-0.0006269526784308255,
-0.04684169217944145,
0.03306373953819275,
0.012049606069922447,
-0.072628453373909,
-0.08600551635026932,
0.01611204631626606,
0.03175605461001396,
0.03317905217409134,
-0.025888727977871895,
0.004096090793609619,
-0.011685235425829887,
-0.037296146154403687,
0.018726332113146782,
-0.08659786731004715,
-0.007949464954435825,
-0.01712828502058983,
-0.037862710654735565,
-0.01419894304126501,
0.005412176251411438,
-0.03781682625412941,
0.03398243710398674,
-0.020701663568615913,
0.03667253255844116,
0.04628365486860275,
0.09059921652078629,
0.006402386352419853,
-0.005738178733736277,
-0.004471281077712774,
-0.05532457306981087,
-0.055197637528181076,
0.029924733564257622,
0.05295691639184952,
0.05502908676862717,
0.15550975501537323,
0.017888303846120834,
0.09108611196279526,
0.0045173270627856255,
-0.007728535681962967,
0.0069877575151622295,
-0.05737040936946869,
0.018557926639914513,
0.03009842149913311,
0.043610647320747375,
-3.53992710699913e-8,
-0.07053105533123016,
0.03149759769439697,
-0.0949372947216034,
0.03442046791315079,
-0.04803112894296646,
-0.008814400993287563,
-0.029231172055006027,
0.05294005945324898,
-0.009665877558290958,
-0.04650789871811867,
0.03203556686639786,
0.024182960391044617,
0.029803162440657616,
0.01461219321936369,
-0.024293947964906693,
0.022604113444685936,
-0.011470185592770576,
0.05793006345629692,
0.022980213165283203,
0.03866361826658249,
-0.010349007323384285,
0.027797814458608627,
0.028996257111430168,
-0.026370691135525703,
-0.006440415512770414,
0.05842909589409828,
-0.035794947296381,
0.07103154808282852,
0.123886838555336,
-0.03062654659152031,
0.004058527760207653,
-0.016391700133681297,
0.022781910374760628,
-0.11632555723190308,
0.07025215774774551,
0.022063152864575386,
0.05464053153991699,
-0.04932465776801109,
-0.037637826055288315,
0.008455006405711174,
0.059584882110357285,
0.017672842368483543,
0.024113761261105537,
0.06995351612567902,
0.012778239324688911,
0.0535435676574707,
0.015221000649034977,
-0.04156913980841637,
0.04743592441082001,
-0.04609493911266327,
-0.044526565819978714,
-0.11691762506961823,
-0.004960780031979084,
0.07644391804933548,
0.08083762973546982,
-0.011464487761259079,
-0.0017537707462906837,
0.030300231650471687,
0.05938917398452759,
-0.006150278262794018,
-0.004042502027004957,
0.007584603037685156,
-0.020824408158659935,
0.0064989845268428326
] |
Streaming Video Question-Answering with In-context Video KV-Cache Retrieval | https://openreview.net/forum?id=8g9fs6mdEG | [
"Shangzhe Di",
"Zhelun Yu",
"Guanghao Zhang",
"Haoyuan Li",
"TaoZhong",
"Hao Cheng",
"Bolin Li",
"Wanggui He",
"Fangxun Shu",
"Hao Jiang"
] | Poster | We propose ReKV, a novel training-free approach that enables efficient streaming video question-answering (StreamingVQA), by seamlessly integrating with existing Video Large Language Models (Video-LLMs). Traditional VideoQA systems struggle with long videos, as they must process entire videos before responding to queries, and repeat this process for each new question. In contrast, our approach analyzes long videos in a streaming manner, allowing for prompt responses as soon as user queries are received. Building on a common Video-LLM, we first incorporate a sliding-window attention mechanism, ensuring that input frames attend to a limited number of preceding frames, thereby reducing computational overhead. To prevent information loss, we store processed video key-value caches (KV-Caches) in RAM and disk, reloading them into GPU memory as needed. Additionally, we introduce a retrieval method that leverages an external retriever or the parameters within Video-LLMs to retrieve only query-relevant KV-Caches, ensuring both efficiency and accuracy in question answering. ReKV enables the separation of video analyzing and question-answering across different processes and GPUs, significantly enhancing the efficiency of StreamingVQA. Through comprehensive experimentation, we validate the efficacy and practicality of our approach, which significantly boosts efficiency and enhances applicability over existing VideoQA models. | Video Understanding, Multimodal Large Language Models, Streaming Video Question-answering | We propose a training-free approach that integrates with existing Video-LLMs for streaming video question-answering. | 861 | 2503.00540 | [
-0.009897290728986263,
-0.07020080089569092,
-0.044172897934913635,
-0.004554864950478077,
0.06180182099342346,
0.03371722623705864,
0.017288140952587128,
0.038400329649448395,
0.05543527007102966,
-0.05898783355951309,
-0.06411337852478027,
-0.03566257655620575,
-0.03165081888437271,
0.04494310915470123,
-0.04956050217151642,
-0.044910021126270294,
0.10248471796512604,
0.043888308107852936,
-0.10864982008934021,
-0.05017373338341713,
0.017435919493436813,
-0.07747288793325424,
0.06961648166179657,
0.026592785492539406,
0.008752478286623955,
0.007813729345798492,
0.006209100596606731,
-0.012888377532362938,
0.06504692882299423,
0.02453687973320484,
0.0737462192773819,
0.007316731847822666,
-0.03057882748544216,
0.07619012147188187,
-0.04045077785849571,
0.0235242061316967,
-0.08837249130010605,
-0.01145453192293644,
-0.11169421672821045,
-0.10219361633062363,
0.003262909594923258,
-0.008291741833090782,
0.002319433493539691,
0.04698671028017998,
0.07253917306661606,
0.02912556193768978,
0.008965998888015747,
-0.023454848676919937,
0.04521017149090767,
-0.0161625724285841,
-0.08194628357887268,
-0.0023628140334039927,
-0.056954722851514816,
0.061895813792943954,
-0.06665178388357162,
0.06660263240337372,
-0.004185594618320465,
0.00172284001018852,
-0.015008226037025452,
0.01094723679125309,
0.0011694724671542645,
-0.12008446455001831,
0.0022897333838045597,
0.05778452754020691,
-0.07610998302698135,
-0.04237149655818939,
0.09240975975990295,
-0.007816107012331486,
0.010976987890899181,
-0.00967561174184084,
-0.07228846102952957,
0.08731310069561005,
-0.025041522458195686,
0.05315210297703743,
-0.09651175141334534,
0.021208908408880234,
0.009818428196012974,
-0.07210466265678406,
0.031936462968587875,
-0.040511880069971085,
0.07873611152172089,
-0.0699237585067749,
0.05843253806233406,
0.03506606072187424,
0.13977937400341034,
-0.011157909408211708,
0.0806851014494896,
-0.023298189043998718,
-0.014098557643592358,
-0.004695453681051731,
-0.097525954246521,
-0.0045811948366463184,
0.09318820387125015,
-0.05205873027443886,
0.07049050182104111,
0.0297681987285614,
0.034459188580513,
-0.05292128026485443,
-0.053404852747917175,
0.06742069870233536,
0.04097122326493263,
0.05479799211025238,
0.0007673698710277677,
-0.022623511031270027,
-0.03029312565922737,
0.04024772346019745,
-0.022998757660388947,
0.1268211305141449,
-0.006854058708995581,
-0.02058948390185833,
0.0348651297390461,
0.08770681917667389,
0.04079212620854378,
-0.024291876703500748,
0.06223253905773163,
0.034219902008771896,
-0.01047872006893158,
-0.0033456350211054087,
-0.0012190580600872636,
0.045760974287986755,
0.021378042176365852,
0.005676602944731712,
-0.07847973704338074,
-0.04290282726287842,
0.049997273832559586,
-0.09514130651950836,
0.008647753857076168,
3.0474029829495466e-34,
0.051539741456508636,
-0.008888408541679382,
0.006196807138621807,
0.005294515751302242,
0.00649790046736598,
0.04096873477101326,
0.052190572023391724,
0.12040645629167557,
-0.010217956267297268,
-0.019246911630034447,
0.016250694170594215,
-0.021613014861941338,
-0.011132053099572659,
0.03117867186665535,
0.028476033359766006,
0.0066342102363705635,
-0.0991583839058876,
0.02428770624101162,
-0.05985886603593826,
-0.010821088217198849,
0.05916254594922066,
0.0013405585195869207,
0.05571229010820389,
-0.005774593912065029,
0.037462037056684494,
-0.0531381331384182,
0.023443477228283882,
-0.08657655119895935,
0.006015843246132135,
-0.015111135318875313,
-0.10701409727334976,
-0.0911860242486,
-0.05389440804719925,
-0.027414552867412567,
0.05499985069036484,
-0.036265816539525986,
-0.02091147005558014,
-0.06529060751199722,
-0.0030767738353461027,
-0.03593486547470093,
0.006339479237794876,
0.00821293331682682,
0.02399652637541294,
-0.062258269637823105,
-0.12872833013534546,
-0.05740412697196007,
0.015036683529615402,
0.07161629945039749,
-0.08356228470802307,
0.019096218049526215,
0.06965983659029007,
0.03175538405776024,
-0.05839608609676361,
-0.06208468973636627,
-0.022219261154532433,
-0.024390030652284622,
0.05725046992301941,
0.06632078438997269,
-0.009776857681572437,
0.07442422956228256,
-0.02286987379193306,
0.026891646906733513,
-0.010946962982416153,
0.09034928679466248,
-0.011035267263650894,
-0.016385354101657867,
0.04439963027834892,
0.13489656150341034,
0.034504394978284836,
0.017242152243852615,
-0.0337478369474411,
-0.024765565991401672,
0.05402760952711105,
-0.06218620389699936,
0.029706209897994995,
0.06937531381845474,
-0.0326160192489624,
-0.051326051354408264,
0.008583961986005306,
0.01000848226249218,
0.05247310921549797,
-0.03190738707780838,
0.012078589759767056,
-0.014065033756196499,
-0.028179829940199852,
-0.04801886901259422,
0.05788761004805565,
-0.09495298564434052,
0.0031197017524391413,
-0.03512189909815788,
-0.04804730787873268,
-0.023295937106013298,
0.0743720531463623,
-0.01854744181036949,
-0.007594168186187744,
1.300887822973355e-33,
0.048875343054533005,
-0.04779291898012161,
-0.04101673141121864,
0.08941377699375153,
0.03691539540886879,
-0.004590305499732494,
0.006331689190119505,
0.015756482258439064,
-0.018306352198123932,
-0.12622365355491638,
-0.06680093705654144,
-0.03507574275135994,
0.02247542142868042,
0.04666436091065407,
0.011390646919608116,
-0.020054208114743233,
-0.02283606119453907,
-0.05292497202754021,
-0.12008523941040039,
0.026524648070335388,
-0.07179145514965057,
0.0723743885755539,
-0.02784401923418045,
-0.03802066296339035,
-0.04469483718276024,
0.047225382179021835,
0.019172314554452896,
0.05528629571199417,
-0.014229494147002697,
0.003660864196717739,
-0.00897945649921894,
-0.010626130737364292,
-0.053421664983034134,
-0.06619434058666229,
-0.025550518184900284,
0.06664108484983444,
0.10971751064062119,
-0.06409861892461777,
-0.06940999627113342,
0.13137871026992798,
0.1265772432088852,
-0.008198420517146587,
-0.007908597588539124,
0.04866073280572891,
-0.004558546934276819,
0.036615584045648575,
-0.13000325858592987,
-0.019534092396497726,
-0.045000918209552765,
0.01113541703671217,
0.005383907817304134,
-0.00891978107392788,
-0.05493193492293358,
-0.004802523646503687,
0.0022995739709585905,
-0.05962517485022545,
-0.10680031776428223,
0.08233234286308289,
0.038010016083717346,
-0.018040692433714867,
0.047773413360118866,
-0.05014597624540329,
-0.022468123584985733,
-0.05774841457605362,
0.0566219799220562,
-0.015215365216135979,
0.013470948673784733,
-0.036454081535339355,
-0.035979002714157104,
-0.05168405547738075,
0.017845341935753822,
-0.06667570024728775,
0.01579260267317295,
-0.004186323378235102,
0.023996546864509583,
0.05100976675748825,
0.05643242970108986,
-0.09461197257041931,
0.012277071364223957,
-0.02122260443866253,
-0.03385106101632118,
0.0012711239978671074,
0.031634628772735596,
0.043613411486148834,
0.050456687808036804,
0.04414309561252594,
0.04582999274134636,
0.08073889464139938,
0.02471321076154709,
0.004736858420073986,
-0.005487451329827309,
0.058484386652708054,
0.013739668764173985,
0.027334637939929962,
0.025708531960844994,
-3.320813490859109e-8,
-0.06536702811717987,
0.054404497146606445,
-0.051444102078676224,
0.085493303835392,
-0.018186982721090317,
-0.027266068384051323,
0.0061804247088730335,
0.07826828211545944,
0.045926615595817566,
-0.049108099192380905,
0.0651402547955513,
-0.04303663223981857,
0.00925291609019041,
-0.011491862125694752,
0.0002935305528808385,
0.0623360238969326,
0.010149925947189331,
-0.02953278087079525,
-0.016279814764857292,
-0.06127920374274254,
0.022777384147047997,
0.040992118418216705,
0.03153490275144577,
0.03356272354722023,
-0.04590025916695595,
0.038926757872104645,
0.05041163042187691,
0.09471738338470459,
0.11488769948482513,
-0.02666492573916912,
0.062442243099212646,
-0.02142399176955223,
0.024631880223751068,
-0.08219712227582932,
0.05408601462841034,
0.010499413125216961,
0.01774493232369423,
0.02676774002611637,
0.07989458739757538,
0.008725383318960667,
0.009436912834644318,
0.006269523873925209,
0.05605850741267204,
0.06301581859588623,
-0.03870617225766182,
0.03290744498372078,
-0.05401213839650154,
-0.033540092408657074,
0.02681068331003189,
0.01158059947192669,
-0.06188465282320976,
-0.08857489377260208,
-0.00467082392424345,
0.05646021291613579,
0.037662044167518616,
-0.039294976741075516,
0.011702642776072025,
0.011589118279516697,
0.03266984596848488,
0.023816021159291267,
0.011512737721204758,
0.026322439312934875,
-0.037132807075977325,
0.022095616906881332
] |
PuzzleFusion++: Auto-agglomerative 3D Fracture Assembly by Denoise and Verify | https://openreview.net/forum?id=7E7v5mJnfl | [
"Zhengqing Wang",
"Jiacheng Chen",
"Yasutaka Furukawa"
] | Poster | This paper proposes a novel “auto-agglomerative” 3D fracture assembly method, PuzzleFusion++, resembling how humans solve challenging spatial puzzles. Starting from individual fragments, the approach 1) aligns and merges fragments into larger groups akin to agglomerative clustering and 2) repeats the process iteratively in completing the assembly akin to auto-regressive methods. Concretely, a diffusion model denoises the 6-DoF alignment parameters of the fragments simultaneously,
and a transformer model verifies and merges pairwise alignments into larger ones, whose process repeats iteratively. Extensive experiments on the Breaking Bad dataset show that PuzzleFusion++ outperforms all other state-of-the-art techniques by significant margins across all metrics In particular by over 10% in part accuracy and 50% in Chamfer distance. We will release code and model. | 3D fracture assembly | null | 855 | 2406.00259 | [
-0.10740963369607925,
-0.10899326950311661,
0.06934977322816849,
-0.05554819479584694,
-0.06281881034374237,
-0.09167762100696564,
-0.08174319565296173,
-0.04534250497817993,
-0.0037502816412597895,
0.013636843301355839,
-0.0649271234869957,
0.0002635866403579712,
-0.08015581220388412,
0.06461308896541595,
-0.09119335561990738,
-0.03319097310304642,
-0.056042857468128204,
0.029749836772680283,
-0.08332792669534683,
-0.028154214844107628,
-0.05822064355015755,
-0.03757934644818306,
-0.06612571328878403,
0.005830289330333471,
0.12044976651668549,
0.0463707260787487,
-0.009050680324435234,
-0.030491281300783157,
0.029964342713356018,
-0.059286631643772125,
0.09329083561897278,
0.0708642527461052,
0.02287549525499344,
0.029360361397266388,
0.04269209876656532,
0.03686228394508362,
-0.12928272783756256,
0.018081296235322952,
-0.04627351462841034,
-0.0607001967728138,
-0.04125256836414337,
0.07932781428098679,
0.03469676524400711,
-0.0019916833844035864,
0.08041588217020035,
0.05017572268843651,
-0.011731117963790894,
-0.027706598863005638,
0.029350943863391876,
-0.026113741099834442,
-0.07983125746250153,
0.021379884332418442,
-0.038603782653808594,
0.010834136977791786,
0.003699076361954212,
0.018947482109069824,
0.06842313706874847,
-0.06758745014667511,
0.020140543580055237,
-0.005235267803072929,
0.04846462979912758,
0.013591747730970383,
-0.02936871536076069,
0.03652646020054817,
0.07415754348039627,
0.005702473688870668,
0.09276503324508667,
-0.016085239127278328,
-0.04454704001545906,
0.04489782080054283,
-0.029309116303920746,
0.06454388797283173,
0.009956720285117626,
0.02815554291009903,
0.08346506953239441,
0.04011368378996849,
-0.042961254715919495,
-0.07867315411567688,
0.0003375050437171012,
-0.00776947196573019,
0.02736547775566578,
0.08246233314275742,
0.07572543621063232,
0.035410795360803604,
0.0031672471668571234,
-0.015916258096694946,
-0.018998660147190094,
0.11554905772209167,
0.027740640565752983,
-0.027687812224030495,
0.015227984637022018,
0.010156460106372833,
-0.05250055715441704,
-0.013806378468871117,
0.06948577612638474,
0.08768981695175171,
0.043862391263246536,
0.013947774656116962,
0.018301187083125114,
0.05243111029267311,
-0.05114994943141937,
0.005830391775816679,
0.007860712707042694,
-0.047641921788454056,
0.047692108899354935,
-0.0540679469704628,
0.01938944309949875,
0.06903983652591705,
0.05493084713816643,
0.019855914637446404,
0.046209096908569336,
-0.010494658723473549,
0.04156048595905304,
0.05705305188894272,
-0.05143796652555466,
-0.10295043140649796,
0.025341913104057312,
-0.017130181193351746,
0.03871051222085953,
-0.0037254116032272577,
0.010447677224874496,
-0.023239657282829285,
-0.06417366117238998,
-0.00020343702635727823,
0.029211457818746567,
-0.008424244821071625,
-0.056825969368219376,
1.7654667003053394e-33,
-0.0017055588541552424,
0.013671591877937317,
0.024578498676419258,
0.07749256491661072,
0.02799897827208042,
-0.042027875781059265,
-0.0632593184709549,
0.015882408246397972,
-0.012290597893297672,
0.09854674339294434,
-0.08446168899536133,
0.0020651023369282484,
-0.02055010013282299,
0.05545266345143318,
-0.07914145290851593,
-0.018191371113061905,
-0.025184867903590202,
0.07757408916950226,
-0.18352890014648438,
0.048611920326948166,
0.03520137816667557,
-0.054560400545597076,
-0.0004499726928770542,
-0.0012207438703626394,
-0.036000825464725494,
0.06605900079011917,
-0.008706871420145035,
-0.03735686466097832,
-0.026875637471675873,
0.0205702967941761,
-0.11001060903072357,
-0.026502367109060287,
-0.07897475361824036,
0.04150863364338875,
0.014512810856103897,
-0.024211449548602104,
-0.03325946256518364,
-0.09751163423061371,
-0.05388350039720535,
0.07533953338861465,
0.039877455681562424,
0.05228223651647568,
-0.022576954215765,
-0.05142547935247421,
0.03199773654341698,
0.009676137939095497,
0.023820584639906883,
0.04032038897275925,
-0.021529780700802803,
0.02758529782295227,
0.025846175849437714,
0.05679843947291374,
-0.015988366678357124,
-0.04453399032354355,
-0.0540161058306694,
0.034634996205568314,
0.02646266110241413,
-0.10934998840093613,
0.039387669414281845,
0.08500156551599503,
0.011753499507904053,
-0.010433604009449482,
-0.07997731119394302,
0.017900167033076286,
0.09901677817106247,
0.002692842623218894,
-0.028797229751944542,
0.07039842754602432,
0.04124997928738594,
0.0049269129522144794,
-0.03740121051669121,
-0.058580655604600906,
-0.0046570682898163795,
0.04563136771321297,
0.007720897439867258,
-0.014551633968949318,
0.10212048888206482,
0.061551179736852646,
-0.035741254687309265,
-0.061502717435359955,
-0.058007583022117615,
0.029176604002714157,
-0.027310382574796677,
-0.1360565572977066,
-0.12925559282302856,
-0.01865367777645588,
0.043593525886535645,
-0.04630713909864426,
-0.05413179472088814,
-0.06519944965839386,
-0.025047725066542625,
-0.09408275038003922,
0.017991622909903526,
-0.04016147926449776,
0.0010806757491081953,
-2.722014942711406e-33,
-0.012204754166305065,
-0.058053240180015564,
-0.005272998008877039,
-0.010020040906965733,
-0.004220602102577686,
-0.04427061602473259,
0.001705995062366128,
-0.0015408442122861743,
0.051712989807128906,
-0.012644416652619839,
-0.04132410138845444,
0.05188777670264244,
0.019462311640381813,
0.006436372175812721,
0.010854987427592278,
-0.041783809661865234,
0.03999900445342064,
-0.009939895011484623,
0.01822005584836006,
0.09313952922821045,
0.03195273503661156,
0.06707651913166046,
-0.03543483093380928,
-0.03170505911111832,
0.025192784145474434,
0.029179643839597702,
0.03492286801338196,
-0.043852947652339935,
-0.005260035861283541,
0.0862705335021019,
-0.03603001683950424,
-0.0034245208371430635,
0.032541558146476746,
0.013205679133534431,
-0.048997361212968826,
0.05197741836309433,
-0.047441381961107254,
-0.02881445176899433,
-0.0030313394963741302,
-0.04627815634012222,
-0.026152033358812332,
0.06318648159503937,
-0.1320452094078064,
0.07926683127880096,
0.028614137321710587,
0.0088471919298172,
-0.037852488458156586,
0.0585206113755703,
-0.03858964517712593,
0.05054722726345062,
-0.004998167511075735,
0.07627524435520172,
-0.04013444110751152,
-0.07185418903827667,
0.029975082725286484,
0.024106767028570175,
-0.0787360817193985,
-0.10249658674001694,
-0.010395176708698273,
0.0019293169025331736,
-0.05098579451441765,
0.0299538504332304,
0.06591631472110748,
0.020527027547359467,
0.0427737757563591,
0.052280012518167496,
-0.009004324674606323,
0.026024997234344482,
-0.0838819220662117,
0.05715295299887657,
0.026760578155517578,
-0.04448044300079346,
0.03133123740553856,
0.015120120719075203,
0.024895332753658295,
-0.0667852833867073,
-0.02829851396381855,
0.00010655893856892362,
0.0014794983435422182,
-0.008708872832357883,
-0.034984927624464035,
-0.07256463170051575,
0.02449127659201622,
0.1309245079755783,
-0.02153703197836876,
0.019444122910499573,
0.02348146215081215,
0.055708639323711395,
-0.025517398491501808,
0.010185894556343555,
-0.06802593171596527,
0.017993440851569176,
0.10775502026081085,
0.03756244480609894,
-0.02329074591398239,
-4.5085773336950297e-8,
0.01575201004743576,
0.06723592430353165,
-0.03519894927740097,
-0.03483288735151291,
-0.020708700641989708,
-0.01011718437075615,
-0.024999449029564857,
0.11651114374399185,
-0.04083123803138733,
-0.0023213333915919065,
0.07137180864810944,
-0.05202804133296013,
-0.07499223947525024,
0.005054517183452845,
-0.0022877464070916176,
0.04854898527264595,
0.033625874668359756,
0.03799182549118996,
-0.11564256995916367,
-0.006680120248347521,
-0.02445363439619541,
0.05225926637649536,
0.05737318843603134,
0.01664401963353157,
-0.03338988125324249,
0.04505836218595505,
-0.07753951847553253,
-0.01614389754831791,
0.0008587702759541571,
-0.004025756381452084,
-0.018311692401766777,
0.00479266420006752,
0.09367057681083679,
0.05774698406457901,
-0.024453457444906235,
0.022565754130482674,
0.023051248863339424,
0.021311122924089432,
0.00300203962251544,
-0.059661999344825745,
-0.002077721292153001,
0.078225277364254,
0.032685425132513046,
0.009001505561172962,
0.016767898574471474,
0.06980547308921814,
0.0015600973274558783,
-0.049932681024074554,
-0.013439983129501343,
-0.012672173790633678,
0.04053304344415665,
0.08252407610416412,
-0.10999894887208939,
0.05979029834270477,
0.08429808169603348,
0.04088183492422104,
-0.0013825867790728807,
-0.026850685477256775,
0.08924537897109985,
0.017907025292515755,
-0.023501506075263023,
-0.0001930584985530004,
-0.004452441353350878,
-0.03309791907668114
] |
Ready-to-React: Online Reaction Policy for Two-Character Interaction Generation | https://openreview.net/forum?id=mm0cqJ2O3f | [
"Zhi Cen",
"Huaijin Pi",
"Sida Peng",
"Qing Shuai",
"Yujun Shen",
"Hujun Bao",
"Xiaowei Zhou",
"Ruizhen Hu"
] | Poster | This paper addresses the task of generating two-character online interactions. Previously, two main settings existed for two-character interaction generation: (1) generating one's motions based on the counterpart's complete motion sequence, and (2) jointly generating two-character motions based on specific conditions. We argue that these settings fail to model the process of real-life two-character interactions, where humans will react to their counterparts in real time and act as independent individuals. In contrast, we propose an online reaction policy, called Ready-to-React, to generate the next character pose based on past observed motions. Each character has its own reaction policy as its ``brain'', enabling them to interact like real humans in a streaming manner. Our policy is implemented by incorporating a diffusion head into an auto-regressive model, which can dynamically respond to the counterpart's motions while effectively mitigating the error accumulation throughout the generation process. We conduct comprehensive experiments using the challenging boxing task. Experimental results demonstrate that our method outperforms existing baselines and can generate extended motion sequences. Additionally, we show that our approach can be controlled by sparse signals, making it well-suited for VR and other online interactive environments. Code and data will be made publicly available. | Human Interaction Generation, Reactive Motion Generation | null | 854 | null | [
-0.10291273146867752,
-0.10639875382184982,
0.01115572638809681,
-0.018603872507810593,
0.0034956003073602915,
0.05578907951712608,
-0.047397010028362274,
-0.053902678191661835,
0.0796569362282753,
-0.02789129503071308,
0.006167857442051172,
-0.07231169939041138,
-0.034078363329172134,
-0.04128728806972504,
-0.005480872467160225,
-0.03367644175887108,
0.05115143582224846,
0.03362730145454407,
-0.028268862515687943,
0.035152748227119446,
0.0053916447795927525,
-0.052168432623147964,
0.0512574166059494,
-0.015945499762892723,
0.014533170498907566,
0.0007024827064014971,
0.027265839278697968,
-0.05828334391117096,
0.025080518797039986,
-0.02812432311475277,
0.0933515653014183,
-0.00419261772185564,
-0.009160126559436321,
-0.020764734596014023,
-0.15723355114459991,
0.0684361606836319,
-0.030982518568634987,
0.05181840807199478,
-0.06771358102560043,
-0.06637831777334213,
-0.014269973151385784,
0.01219544094055891,
0.03897484019398689,
0.0630839541554451,
0.0734800174832344,
0.06372445076704025,
0.0028445650823414326,
-0.016073020175099373,
-0.003225774271413684,
-0.015462039038538933,
-0.09541231393814087,
0.00921100564301014,
0.06673148274421692,
0.05271606892347336,
0.03421610966324806,
0.048099808394908905,
0.04457129165530205,
0.028795968741178513,
0.06136033311486244,
-0.03849157691001892,
-0.027138199657201767,
-0.0490366667509079,
0.04987163096666336,
-0.013096265494823456,
-0.030071720480918884,
0.018442628905177116,
0.0559215173125267,
0.04983612522482872,
0.0013678240356966853,
0.09775909036397934,
-0.08411580324172974,
-0.04000402241945267,
-0.013476650230586529,
-0.001552107511088252,
0.00851605273783207,
-0.023626018315553665,
0.013574140146374702,
-0.08741451054811478,
0.030262209475040436,
0.02995188906788826,
0.026069054380059242,
-0.10573229193687439,
0.025516390800476074,
-0.010205360129475594,
-0.037385355681180954,
0.03846735507249832,
-0.038209982216358185,
0.053788766264915466,
0.021749943494796753,
0.03723223879933357,
-0.13604219257831573,
0.007648207247257233,
0.09856308996677399,
-0.03757396712899208,
-0.023080389946699142,
0.01185153890401125,
-0.001687259180471301,
-0.03846266493201256,
0.04602697119116783,
0.05895976349711418,
0.008884130045771599,
0.014843009412288666,
-0.015156562440097332,
0.017108749598264694,
0.02667098119854927,
-0.02666647918522358,
0.05892764404416084,
0.012774227187037468,
0.013673569075763226,
-0.0020050429739058018,
0.041876666247844696,
-0.05644918605685234,
-0.0036761898081749678,
0.009416534565389156,
0.043150514364242554,
0.042263489216566086,
-0.010072875767946243,
-0.02359589748084545,
0.042966313660144806,
0.00254032202064991,
0.1394791156053543,
-0.0695006251335144,
-0.007487230468541384,
0.006448941770941019,
0.06591594964265823,
-0.07070799171924591,
-0.024183854460716248,
-1.849320957163232e-33,
0.060888953506946564,
0.048548392951488495,
0.02695278450846672,
0.053342223167419434,
-0.00016070804849732667,
0.024788586422801018,
-0.011173420585691929,
-0.024908168241381645,
-0.027131639420986176,
-0.014975398778915405,
-0.02017502672970295,
-0.012917917221784592,
-0.025550594553351402,
0.1499701887369156,
-0.004454950336366892,
-0.028441790491342545,
-0.06524500995874405,
0.05346625670790672,
0.011494377627968788,
0.017048656940460205,
-0.007934853434562683,
0.024168390780687332,
-0.07024702429771423,
-0.05634387210011482,
-0.05375978723168373,
0.042987506836652756,
0.017210081219673157,
-0.028722651302814484,
0.040002696216106415,
-0.035025835037231445,
-0.04608352482318878,
0.0038629481568932533,
-0.035917505621910095,
-0.013346116058528423,
0.05476949363946915,
-0.04656302183866501,
-0.04026534780859947,
-0.0395364984869957,
-0.01440843939781189,
0.04900231212377548,
0.02650310844182968,
0.06131575256586075,
-0.03583420068025589,
-0.02952859364449978,
-0.07877214252948761,
-0.06332600116729736,
0.05847479775547981,
0.1069873720407486,
-0.0996788963675499,
-0.008264456875622272,
0.02333207242190838,
0.09129439294338226,
-0.020259350538253784,
-0.048509541898965836,
-0.03311185538768768,
0.014351477846503258,
-0.03743879869580269,
-0.04186859726905823,
0.01749350316822529,
0.06569536030292511,
0.0071039954200387,
0.05337018892168999,
0.024129332974553108,
0.006271879654377699,
0.0456894151866436,
0.015042301267385483,
-0.037596818059682846,
0.07798679172992706,
0.038788050413131714,
-0.0011403380194678903,
-0.005498439073562622,
-0.01528941746801138,
-0.02065346948802471,
-0.11331142485141754,
0.03248753771185875,
0.01705385558307171,
0.08434396982192993,
-0.04158646985888481,
-0.08748257905244827,
0.06406354159116745,
0.017560895532369614,
-0.029780449345707893,
-0.055496715009212494,
-0.012587185949087143,
0.009442076086997986,
-0.04120545834302902,
-0.03807297721505165,
-0.056574635207653046,
-0.14288227260112762,
-0.013486642390489578,
-0.050706636160612106,
-0.06550701707601547,
0.014327291399240494,
0.017849963158369064,
0.024359887465834618,
-9.11013144140969e-36,
0.03304615989327431,
0.006266691256314516,
-0.043847378343343735,
0.04503040015697479,
0.038771212100982666,
-0.022406702861189842,
0.03833472728729248,
-0.043820519000291824,
0.048459526151418686,
-0.01821615919470787,
-0.07321695983409882,
-0.02591106668114662,
0.03583313152194023,
0.005814907141029835,
0.03974917158484459,
-0.07980728149414062,
0.006301714573055506,
-0.012831557542085648,
-0.01011078990995884,
-0.01670845039188862,
0.08913961052894592,
0.053386151790618896,
-0.05322832614183426,
-0.012340077199041843,
-0.025205038487911224,
0.06544896215200424,
0.031006287783384323,
0.03827643394470215,
-0.042509544640779495,
0.032721422612667084,
-0.034166716039180756,
-0.0016182322287932038,
-0.03522500395774841,
-0.0008695476572029293,
-0.06784040480852127,
0.11380349844694138,
0.021158475428819656,
0.048029281198978424,
0.0027492197696119547,
0.05401976406574249,
0.052960023283958435,
-0.02367192506790161,
-0.02428165078163147,
0.05863625556230545,
-0.021065667271614075,
0.027766242623329163,
-0.11851172149181366,
0.05352821201086044,
-0.08805989474058151,
0.08131839334964752,
-0.07072429358959198,
0.11470771580934525,
-0.13908234238624573,
-0.12508827447891235,
-0.0871332511305809,
-0.10338112711906433,
-0.006214246153831482,
0.008888434618711472,
0.08496812731027603,
-0.025103075429797173,
-0.033905550837516785,
-0.062308698892593384,
-0.06461618840694427,
-0.023299913853406906,
0.013790139928460121,
0.032225556671619415,
-0.03308025747537613,
0.021809181198477745,
0.08155011385679245,
0.021562380716204643,
0.0916978195309639,
0.020042559131979942,
0.026299888268113136,
0.03306882828474045,
0.0935843214392662,
-0.04873303323984146,
-0.012637478299438953,
-0.023312296718358994,
0.05880064517259598,
-0.002579920459538698,
-0.06482095271348953,
0.046809226274490356,
0.09675467759370804,
0.04974360391497612,
-0.006299557164311409,
0.09445512294769287,
-0.0715826228260994,
0.1043153628706932,
-0.025611063465476036,
0.02415098436176777,
-0.027316447347402573,
0.06033032014966011,
0.021761296316981316,
0.04058266431093216,
-0.03201404958963394,
-4.7215838350211925e-8,
-0.08213815093040466,
-0.005421539302915335,
0.0778464525938034,
0.06153617426753044,
-0.013165437616407871,
0.05108451098203659,
0.026465527713298798,
-0.09541960805654526,
-0.007460801396518946,
-0.04485306888818741,
0.02724887803196907,
0.037277910858392715,
0.048535265028476715,
-0.015406064689159393,
-0.0032585018780082464,
0.0030539403669536114,
-0.09887398779392242,
-0.05738459527492523,
-0.09286507964134216,
0.029874149709939957,
0.046586476266384125,
0.021587589755654335,
-0.052150480449199677,
0.01603464037179947,
-0.0055326190777122974,
0.01524445042014122,
0.0218387171626091,
0.019981037825345993,
-0.004626456182450056,
-0.008478137664496899,
-0.010717672295868397,
0.007526670582592487,
-0.006348605267703533,
-0.07839648425579071,
0.0102566322311759,
-0.03396674990653992,
0.02770080976188183,
-0.0040836213156580925,
0.010465942323207855,
-0.04250781238079071,
-0.09018437564373016,
0.07252815365791321,
-0.022180194035172462,
0.0015923988539725542,
0.007380759809166193,
0.0078126210719347,
-0.01802358217537403,
-0.13095460832118988,
0.044802382588386536,
-0.048356976360082626,
0.00571444584056735,
-0.083407923579216,
0.012209217064082623,
0.07564391940832138,
0.08098023384809494,
0.045853909105062485,
-0.007376804947853088,
0.02894464321434498,
0.10071723163127899,
0.083061583340168,
0.005059343297034502,
0.08450382202863693,
-0.07938330620527267,
-0.017892390489578247
] |
Tight Time Complexities in Parallel Stochastic Optimization with Arbitrary Computation Dynamics | https://openreview.net/forum?id=cUN8lJB4rD | [
"Alexander Tyurin"
] | Poster | In distributed stochastic optimization, where parallel and asynchronous methods are employed, we establish optimal time complexities under virtually any computation behavior of workers/devices/CPUs/GPUs, capturing potential disconnections due to hardware and network delays, time-varying computation powers, and any possible fluctuations and trends of computation speeds. These real-world scenarios are formalized by our new universal computation model. Leveraging this model and new proof techniques, we discover tight lower bounds that apply to virtually all synchronous and asynchronous methods, including Minibatch SGD, Asynchronous SGD (Recht et al., 2011), and Picky SGD (Cohen et al., 2021). We show that these lower bounds, up to constant factors, are matched by the optimal Rennala SGD and Malenia SGD methods (Tyurin & Richtárik, 2023). | nonconvex optimization, lower bounds, parallel methods, asynchronous methods, convex optimization | null | 852 | 2408.04929 | [
-0.1290624588727951,
0.007908483035862446,
-0.02066492848098278,
0.009051821194589138,
0.012316365726292133,
-0.0841926708817482,
-0.05693717673420906,
-0.04462893307209015,
0.029458923265337944,
-0.011974942870438099,
-0.027258113026618958,
0.03299252316355705,
-0.02865917608141899,
-0.052177779376506805,
0.01864694058895111,
-0.045438677072525024,
0.07805819809436798,
0.042549945414066315,
-0.08068521320819855,
-0.04479491338133812,
-0.009278499521315098,
-0.04278574883937836,
-0.00951319932937622,
-0.026421889662742615,
0.047441739588975906,
-0.08024345338344574,
0.024488452821969986,
-0.09532265365123749,
0.03204250708222389,
0.025572316721081734,
0.037296805530786514,
-0.04632013291120529,
-0.047042325139045715,
-0.013321064412593842,
0.020253736525774002,
0.08901097625494003,
-0.04020754247903824,
-0.02233145199716091,
-0.0197139959782362,
-0.00001104000421037199,
0.021688144654035568,
-0.04124622046947479,
-0.03901803866028786,
0.05898653343319893,
-0.004148381762206554,
-0.030593840405344963,
0.036482106894254684,
-0.029569610953330994,
-0.009322666563093662,
-0.04579361528158188,
0.016952306032180786,
-0.00044225368765182793,
-0.03956393525004387,
0.020323317497968674,
-0.003184140892699361,
0.009653668850660324,
0.07476036250591278,
-0.03983930125832558,
0.0002948276815004647,
0.03064987249672413,
0.06406062096357346,
-0.14091944694519043,
-0.05198955908417702,
0.001777611905708909,
0.03584212809801102,
0.02194955386221409,
0.10379926860332489,
-0.020562561228871346,
0.019521748647093773,
0.06375039368867874,
0.021869542077183723,
0.01821674406528473,
-0.06848359107971191,
0.05399126932024956,
0.005312354303896427,
-0.0050575449131429195,
0.057319868355989456,
-0.019670339301228523,
0.019935963675379753,
-0.04925340786576271,
0.01794082298874855,
0.04745275527238846,
-0.003412055317312479,
-0.03202402964234352,
0.041496288031339645,
-0.02952839806675911,
-0.012071666307747364,
0.03531736135482788,
0.13328295946121216,
-0.06716465950012207,
0.00376737117767334,
0.03396429494023323,
-0.06559447199106216,
-0.01647201180458069,
-0.0291993897408247,
-0.027207234874367714,
0.04540077969431877,
0.01829082891345024,
-0.033054519444704056,
0.07472693175077438,
-0.002459292532876134,
-0.012052544392645359,
0.04546095058321953,
0.006770213134586811,
0.04583262279629707,
0.00830657035112381,
0.005200119689106941,
0.02536841481924057,
-0.04469393938779831,
-0.030469492077827454,
0.06330082565546036,
0.09973234683275223,
0.04141802340745926,
0.09051434695720673,
-0.026645395904779434,
-0.016297241672873497,
0.008137090131640434,
-0.04943358525633812,
0.05835141986608505,
0.04597342386841774,
-0.08966375142335892,
0.03632017597556114,
0.030095120891928673,
-0.019790133461356163,
0.016972258687019348,
0.07388933002948761,
-0.06419797986745834,
3.229908399148292e-33,
-0.003941905684769154,
-0.014287213794887066,
0.057409822940826416,
-0.14869649708271027,
0.04921218007802963,
0.023588648065924644,
0.010229903273284435,
-0.00965455174446106,
-0.017557119950652122,
0.0006937171565368772,
-0.02939772419631481,
-0.047669049352407455,
-0.024944424629211426,
-0.013623952865600586,
0.04365546256303787,
-0.03685035929083824,
-0.00700352992862463,
-0.010016207583248615,
0.060597099363803864,
-0.07095933705568314,
0.10411686450242996,
-0.026014693081378937,
-0.06983737647533417,
0.015462581999599934,
0.018571000546216965,
-0.08095809072256088,
0.017704861238598824,
-0.04966961592435837,
0.04204104095697403,
0.07552958279848099,
-0.09196960180997849,
0.04145973175764084,
0.01800699718296528,
0.08346389979124069,
-0.0010388633236289024,
0.04040912538766861,
-0.04100611433386803,
-0.08363759517669678,
-0.004703083541244268,
-0.03273357078433037,
-0.029894346371293068,
0.015670068562030792,
0.03266255557537079,
-0.08694949746131897,
-0.13955944776535034,
-0.1293545961380005,
0.022669222205877304,
0.05094393715262413,
0.032002322375774384,
-0.06469785422086716,
0.0031941081397235394,
0.04102453589439392,
-0.04145429655909538,
-0.04525759443640709,
0.03643054515123367,
-0.002532230457291007,
0.043548960238695145,
0.004988526459783316,
0.07202646881341934,
0.17175261676311493,
-0.010867932811379433,
0.0044492073357105255,
0.006259463727474213,
0.06480172276496887,
-0.024337928742170334,
0.05430849269032478,
-0.032391976565122604,
0.047092244029045105,
0.04521334543824196,
-0.003445851616561413,
-0.013265501707792282,
-0.028575193136930466,
0.01112399436533451,
-0.046150799840688705,
0.07611969858407974,
0.014729277230799198,
0.07858296483755112,
-0.04207639396190643,
-0.08398952335119247,
-0.0116658303886652,
-0.06401381641626358,
0.07275312393903732,
-0.07107952982187271,
-0.000878905993886292,
-0.04478868469595909,
-0.06472945213317871,
-0.018105071038007736,
0.008241593837738037,
-0.048864394426345825,
-0.029555561020970345,
-0.14244131743907928,
0.029293706640601158,
0.09627635776996613,
0.007556541822850704,
-0.050985924899578094,
-4.306181475869578e-33,
-0.06561771035194397,
-0.08046365529298782,
0.027490846812725067,
0.11441180109977722,
0.14221642911434174,
-0.037836119532585144,
-0.009416988119482994,
-0.10504535585641861,
-0.05773051828145981,
-0.042753666639328,
-0.035536959767341614,
0.017851896584033966,
-0.0324363075196743,
-0.0069642080925405025,
0.0514904223382473,
-0.05083594471216202,
0.07064606249332428,
0.07151143997907639,
0.029528984799981117,
0.0018800231628119946,
0.018998343497514725,
0.058325380086898804,
-0.023059962317347527,
0.03626210615038872,
0.08685373514890671,
-0.028076522052288055,
-0.05661817267537117,
0.02409990318119526,
-0.045536722987890244,
-0.02613060548901558,
-0.02634853683412075,
-0.01133597269654274,
-0.10140693187713623,
0.04377211257815361,
-0.0009234788594767451,
0.02695351466536522,
-0.0017467525321990252,
0.061585042625665665,
0.02161564491689205,
0.04496971517801285,
0.10582173615694046,
-0.03585173562169075,
-0.04701342433691025,
0.04228801280260086,
0.08147057145833969,
0.0337228998541832,
-0.10318988561630249,
0.007250691764056683,
-0.048261623829603195,
-0.020294122397899628,
0.04309754818677902,
0.019981050863862038,
0.023764507845044136,
0.0583031065762043,
0.0023790630511939526,
-0.030730437487363815,
0.01533125713467598,
-0.05130403861403465,
0.06674860417842865,
0.0687696784734726,
-0.010094108060002327,
-0.07788872718811035,
-0.02877129800617695,
0.04528966173529625,
0.014574692584574223,
-0.05341002717614174,
0.024966154247522354,
-0.015629157423973083,
0.04362686350941658,
0.033354196697473526,
0.03056156449019909,
0.06507554650306702,
-0.09412042796611786,
0.06105222925543785,
-0.0766100063920021,
0.09045957773923874,
0.08748864382505417,
-0.01675243116915226,
0.001081869937479496,
-0.0007277206750586629,
0.06544904410839081,
0.08888613432645798,
0.011631356552243233,
-0.006689924281090498,
-0.03392105922102928,
0.024138225242495537,
0.08845341205596924,
-0.00546668004244566,
0.010024255141615868,
-0.08104248344898224,
-0.09880314022302628,
0.037438832223415375,
-0.04008536785840988,
0.0009213478770107031,
-0.049160148948431015,
-4.249153207069867e-8,
0.07167962193489075,
0.011076662689447403,
0.06325068324804306,
0.019236553460359573,
0.11128043383359909,
-0.06843660771846771,
-0.019408047199249268,
-0.050455838441848755,
0.05199262872338295,
0.03166493400931358,
0.04518377408385277,
-0.051632415503263474,
0.06543350964784622,
-0.06400544196367264,
-0.033079832792282104,
0.044379740953445435,
-0.041401755064725876,
-0.05524405837059021,
-0.016650283709168434,
-0.042124006897211075,
-0.012396883219480515,
0.04619317501783371,
0.00786350853741169,
-0.010171432048082352,
-0.004521167371422052,
-0.022187231108546257,
0.09137682616710663,
0.041724659502506256,
0.05486089363694191,
0.017637383192777634,
-0.0771370530128479,
0.009046204388141632,
-0.02652737684547901,
-0.005161104258149862,
0.003974096849560738,
0.017310762777924538,
0.07518723607063293,
0.04035035893321037,
0.09052003920078278,
-0.0019521733047440648,
-0.021422894671559334,
-0.007950276136398315,
0.038244809955358505,
0.02141471952199936,
-0.030671490356326103,
-0.025307299569249153,
-0.021197684109210968,
-0.011057869531214237,
-0.053630415350198746,
0.049954332411289215,
-0.004675345029681921,
-0.0016457405872642994,
-0.001330471015535295,
0.052592650055885315,
0.053387243300676346,
-0.011807845905423164,
0.004486177582293749,
-0.10873576253652573,
-0.014023588970303535,
0.029681449756026268,
0.05949268117547035,
-0.009619665332138538,
-0.09631780534982681,
-0.0037839901633560658
] |
MEGA-Bench: Scaling Multimodal Evaluation to over 500 Real-World Tasks | https://openreview.net/forum?id=2rWbKbmOuM | [
"Jiacheng Chen",
"Tianhao Liang",
"Sherman Siu",
"Zhengqing Wang",
"Kai Wang",
"Yubo Wang",
"Yuansheng Ni",
"Ziyan Jiang",
"Wang Zhu",
"Bohan Lyu",
"Dongfu Jiang",
"Xuan He",
"Yuan Liu",
"Hexiang Hu",
"Xiang Yue",
"Wenhu Chen"
] | Poster | We present MEGA-Bench, an evaluation suite that scales multimodal evaluation to over 500 real-world tasks, to address the highly heterogeneous daily use cases of end users.
Our objective is to optimize for a set of high-quality data samples that cover a highly diverse and rich set of multimodal tasks, while enabling cost-effective and accurate model evaluation.
In particular, we collected 505 realistic tasks encompassing over 8,000 samples from 16 expert annotators to extensively cover the multimodal task space. Instead of unifying these problems into standard multi-choice questions (like MMMU, MM-Bench, and MMT-Bench), we embrace a wide range of output formats like numbers, phrases, code, \LaTeX, coordinates, JSON, free-form, etc. To accommodate these formats, we developed over 40 metrics to evaluate these tasks.
Unlike existing benchmarks, MEGA-Bench offers a fine-grained capability report across multiple dimensions (e.g., application, input type, output format, skill), allowing users to interact with and visualize model capabilities in depth. We evaluate a wide variety of frontier vision-language models on MEGA-Bench to understand their capabilities across these dimensions. | evaluation of multimodal large language models | null | 851 | null | [
-0.027166150510311127,
-0.06965027004480362,
-0.0036274450831115246,
0.020944654941558838,
0.09519010782241821,
-0.0517340824007988,
-0.08654973655939102,
0.08104851841926575,
-0.0386909656226635,
-0.015880774706602097,
-0.09351437538862228,
-0.10109654068946838,
0.03629268333315849,
0.08671806007623672,
0.03280148282647133,
0.015390532091259956,
0.03674742951989174,
0.024558259174227715,
-0.06811311095952988,
-0.0009257887722924352,
0.06300216168165207,
0.015846583992242813,
0.02665434591472149,
-0.05780380964279175,
0.07398232072591782,
-0.0019195623463019729,
-0.018657146021723747,
-0.028549715876579285,
-0.004138719290494919,
-0.029946977272629738,
-0.060680728405714035,
0.04666732996702194,
0.009113556705415249,
0.02748917229473591,
0.013036178424954414,
0.059412069618701935,
-0.0244757030159235,
-0.036077845841646194,
-0.02043524943292141,
-0.07375829666852951,
-0.02204114943742752,
-0.0002210260572610423,
0.04066888615489006,
-0.005695297382771969,
0.054978564381599426,
-0.04836241900920868,
-0.029100948944687843,
-0.08028005808591843,
0.01207642536610365,
0.05363538861274719,
-0.14141559600830078,
-0.06619355827569962,
0.013628322631120682,
0.01452553179115057,
-0.0011192923411726952,
-0.017952220514416695,
0.02862987294793129,
-0.062420058995485306,
-0.0013131335144862533,
-0.024375688284635544,
-0.06534010916948318,
-0.034488022327423096,
-0.06331996619701385,
-0.02909165993332863,
0.04338618740439415,
-0.0544305182993412,
-0.004755237139761448,
-0.006628468632698059,
0.026479024440050125,
0.008448973298072815,
-0.0978325605392456,
0.036022912710905075,
-0.059003956615924835,
0.02623572200536728,
-0.0004926808178424835,
0.018534816801548004,
-0.029736274853348732,
-0.031133705750107765,
0.0053587243892252445,
-0.08620678633451462,
0.03175845742225647,
-0.015557610429823399,
0.0007396903238259256,
0.02513372339308262,
0.059664543718099594,
0.0015707204584032297,
0.024823052808642387,
0.06781159341335297,
0.032338280230760574,
0.036384496837854385,
-0.06105310097336769,
-0.01665732078254223,
0.02122090756893158,
0.01362377405166626,
0.019211165606975555,
0.008174058981239796,
0.08147913217544556,
-0.08461657166481018,
-0.0277638491243124,
0.09279725700616837,
0.06307930499315262,
0.03482360765337944,
0.024995634332299232,
-0.0812411904335022,
0.00040301590343005955,
-0.021922642365098,
0.0865355059504509,
0.05737012252211571,
0.013314347714185715,
-0.08686306327581406,
0.04214160144329071,
0.01988130435347557,
-0.028752772137522697,
-0.038993045687675476,
0.06838235259056091,
-0.0352206751704216,
0.018117723986506462,
-0.023775046691298485,
0.05073259025812149,
0.028928788378834724,
-0.02006925642490387,
0.024903537705540657,
-0.01082131452858448,
-0.014214197173714638,
0.05192216485738754,
0.040316931903362274,
-0.0548463836312294,
1.1188840122546247e-33,
0.043547675013542175,
0.03169439360499382,
0.07257749140262604,
0.06203514337539673,
0.041108839213848114,
-0.04618033021688461,
-0.03853940963745117,
0.03399357572197914,
-0.0859638899564743,
-0.025077976286411285,
-0.016510317102074623,
0.05226771533489227,
-0.03765765205025673,
0.11322074383497238,
0.045907218009233475,
-0.01133783906698227,
-0.029127368703484535,
0.05049582198262215,
0.006589948199689388,
0.07169610261917114,
0.021683217957615852,
-0.02433374896645546,
-0.026778243482112885,
0.05190785974264145,
0.027039019390940666,
0.026919152587652206,
0.0673505961894989,
-0.045864176005125046,
-0.056738294661045074,
0.005009524058550596,
-0.12187585234642029,
-0.0677349865436554,
0.012954739853739738,
-0.02495545521378517,
0.057412825524806976,
-0.02600335329771042,
-0.02532755769789219,
-0.04889123514294624,
-0.0015547231305390596,
-0.042291611433029175,
-0.06808104366064072,
-0.007934589870274067,
-0.09291492402553558,
0.00030363493715412915,
-0.10190178453922272,
-0.004710067063570023,
-0.013383006677031517,
0.029620744287967682,
-0.0032435201574116945,
-0.05776544660329819,
0.012959311716258526,
0.07514077425003052,
0.0486551895737648,
-0.041225776076316833,
0.0017927909502759576,
-0.05477514863014221,
0.062150150537490845,
0.03338739275932312,
0.0850193127989769,
0.0555119551718235,
-0.033351872116327286,
0.08065076917409897,
0.011526267975568771,
0.040044140070676804,
-0.01378465723246336,
0.0035923782270401716,
-0.005840555764734745,
0.05503318831324577,
0.0730770081281662,
0.039858147501945496,
-0.07549789547920227,
0.019924018532037735,
0.06037504971027374,
-0.0822279155254364,
0.052239738404750824,
0.009660398587584496,
0.0661955252289772,
-0.07824978232383728,
0.04463234171271324,
0.00530571024864912,
0.04657892882823944,
0.038170862942934036,
0.050807349383831024,
-0.0863705426454544,
0.0003751774784177542,
-0.05671805515885353,
0.02116735279560089,
-0.09858697652816772,
-0.03646240383386612,
-0.028002088889479637,
-0.038955360651016235,
-0.001578501076437533,
-0.05074148625135422,
-0.06905738264322281,
-0.04249158874154091,
-9.854438220946979e-34,
0.024443112313747406,
-0.006920402403920889,
-0.042990297079086304,
0.08516539633274078,
-0.007294060662388802,
-0.06957881897687912,
-0.023588189855217934,
0.012845148332417011,
-0.007704836782068014,
-0.053628794848918915,
-0.012976471334695816,
-0.039285074919462204,
0.025244712829589844,
-0.038842491805553436,
0.03347005695104599,
-0.007351034786552191,
-0.0745573490858078,
-0.03188984841108322,
-0.0038722914177924395,
0.03154797479510307,
0.005058341659605503,
0.09736346453428268,
-0.00989900529384613,
0.06056835502386093,
-0.09396808594465256,
0.06214665621519089,
-0.08222249895334244,
-0.034329358488321304,
-0.0606219656765461,
-0.05956078693270683,
-0.013409128412604332,
-0.07710415869951248,
0.05234525352716446,
-0.04357326030731201,
0.030808720737695694,
-0.006183042656630278,
0.09987543523311615,
-0.001923139556311071,
-0.012801647186279297,
0.1376720666885376,
0.07958832383155823,
-0.01118128839880228,
-0.014864207245409489,
-0.023851465433835983,
-0.0016923007788136601,
0.014707901515066624,
-0.13221287727355957,
-0.029175924137234688,
-0.004334187135100365,
0.031176812946796417,
-0.025578897446393967,
0.03898032382130623,
-0.06113571301102638,
-0.06636877357959747,
0.006303433328866959,
-0.11646390706300735,
-0.04184190556406975,
-0.03631911426782608,
0.03815150260925293,
0.05106887221336365,
-0.07558471709489822,
-0.06738447397947311,
-0.016394861042499542,
-0.03318295255303383,
0.053969092667102814,
-0.07063137739896774,
0.008547435514628887,
-0.09186866879463196,
-0.05608532205224037,
0.030579598620533943,
-0.0027622689958661795,
-0.0760694220662117,
-0.011714695952832699,
0.036643192172050476,
-0.010316472500562668,
0.023623881861567497,
0.052967268973588943,
-0.02210024558007717,
0.025946656242012978,
0.00015824260481167585,
0.048632584512233734,
0.01206767838448286,
0.08497119694948196,
0.04738503322005272,
0.0569680891931057,
0.11033689230680466,
0.018958033993840218,
0.11935519427061081,
-0.005368493031710386,
0.03710459545254707,
-0.01797395385801792,
0.019534239545464516,
-0.004258624278008938,
0.041896235197782516,
0.012667674571275711,
-4.321264768236688e-8,
-0.047405559569597244,
0.05301652103662491,
0.009326010011136532,
0.012481719255447388,
-0.051789071410894394,
0.00098918448202312,
-0.005970764439553022,
0.07596448808908463,
0.09405007213354111,
0.06145962327718735,
0.13039198517799377,
-0.08195870369672775,
-0.05970125272870064,
-0.010444669052958488,
0.06764038652181625,
0.04782862216234207,
-0.032828669995069504,
0.01876145973801613,
-0.05741256847977638,
-0.04176299646496773,
0.06255808472633362,
0.10384619235992432,
0.006549561861902475,
-0.022382846102118492,
0.041203323751688004,
-0.05658702924847603,
-0.036216530948877335,
0.16523313522338867,
0.018655691295862198,
-0.05270708352327347,
0.024375546723604202,
0.007103488780558109,
0.004176788497716188,
-0.043373748660087585,
0.013981390744447708,
-0.011719658970832825,
0.015140904113650322,
0.06367245316505432,
0.04186520725488663,
-0.002473712433129549,
0.0020493515767157078,
0.05557866021990776,
-0.09539853036403656,
0.0421147458255291,
0.06137431040406227,
-0.020806636661291122,
-0.08202629536390305,
-0.034341905266046524,
-0.020550444722175598,
-0.032877255231142044,
-0.0663849264383316,
-0.04721370339393616,
0.004871995188295841,
0.10386160761117935,
0.078443743288517,
0.04406287148594856,
-0.024911511689424515,
0.037382397800683975,
0.0578661784529686,
0.0664096251130104,
0.0818222165107727,
-0.007551003247499466,
-0.0752599760890007,
0.025850776582956314
] |
Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow | https://openreview.net/forum?id=nEDToD1R8M | [
"Fu-Yun Wang",
"Ling Yang",
"Zhaoyang Huang",
"Mengdi Wang",
"Hongsheng Li"
] | Poster | Diffusion models have greatly improved visual generation but are hindered by slow generation speed due to the computationally intensive nature of solving generative ODEs. Rectified flow, a widely recognized solution, improves generation speed by straightening the ODE path. Its
key components include: 1) using the linear interpolating diffusion form of flow-matching, 2) employing $\boldsymbol v$-prediction, and 3) performing rectification (a.k.a. reflow). In this paper, we argue that the success of rectification primarily lies in using a pretrained diffusion model to obtain matched pairs of noise and samples, followed by retraining with these matched noise-sample pairs. Based on this, components 1) and 2) are unnecessary. Furthermore, we highlight that straightness is not an essential training target for rectification; rather, it is a specific case of flow-matching models. The more critical training target is to achieve a first-order approximate ODE path, which is inherently curved for models like DDPM and Sub-VP. Building on this insight, we propose Rectified Diffusion, which generalizes the design space and application scope of rectification to encompass the broader category of diffusion models, rather than being restricted to flow-matching models. We validate our method on Stable Diffusion v1-5 and Stable Diffusion XL. Our method not only greatly simplifies the training procedure of rectified flow-based previous works (e.g., InstaFlow) but also achieves superior performance with even lower training cost.
Our code is available at https://github.com/G-U-N/Rectified-Diffusion. | Diffusion models, Rectified-Flow | null | 847 | 2410.07303 | [
-0.09718313813209534,
-0.11477188766002655,
0.07680314779281616,
0.023039313033223152,
0.004724143538624048,
-0.05475659668445587,
-0.07461631298065186,
-0.08419286459684372,
0.04325437918305397,
-0.09982015937566757,
-0.055638957768678665,
-0.010210626758635044,
-0.012070572935044765,
-0.019356250762939453,
-0.060417864471673965,
-0.035941094160079956,
0.015584333799779415,
0.133576437830925,
-0.014336715452373028,
-0.0449298620223999,
0.002481957199051976,
-0.03322862833738327,
-0.063343346118927,
-0.03924567624926567,
0.043010301887989044,
0.09825535118579865,
0.07136964797973633,
-0.02118498459458351,
0.01619143970310688,
-0.08327987790107727,
0.07645739614963531,
0.038850173354148865,
-0.06881134957075119,
-0.0067975628189742565,
-0.11358968168497086,
0.052724439650774,
-0.041630830615758896,
0.02896079421043396,
0.022060127928853035,
0.010073228739202023,
-0.024031152948737144,
0.04294777661561966,
0.04475376009941101,
0.04034417122602463,
0.16801932454109192,
-0.060509514063596725,
0.011671599000692368,
-0.064510278403759,
-0.04334418475627899,
0.013570700772106647,
0.01602662354707718,
-0.03432108461856842,
-0.11132650822401047,
0.05949074402451515,
0.006099103484302759,
0.05897287279367447,
0.10534558445215225,
0.014657964929938316,
0.025987708941102028,
-0.005728329531848431,
-0.03489844128489494,
-0.08069618046283722,
-0.01677335985004902,
-0.006713735871016979,
-0.07000209391117096,
-0.05559871345758438,
0.04235163703560829,
0.04942389577627182,
0.038093723356723785,
0.06568282097578049,
-0.04419010132551193,
0.06583905220031738,
-0.02268669195473194,
-0.0555163137614727,
-0.039946865290403366,
0.016032051295042038,
0.07670903205871582,
0.034140244126319885,
-0.0053422399796545506,
-0.05891424044966698,
0.06408853828907013,
0.018334828317165375,
0.04982779547572136,
-0.058392275124788284,
0.01258679386228323,
0.003360231639817357,
0.0016859255265444517,
-0.013413267210125923,
0.016451120376586914,
-0.0151227917522192,
-0.023664001375436783,
0.09876459091901779,
-0.07349706441164017,
-0.02429153583943844,
0.060507286339998245,
0.07415630668401718,
-0.01874389685690403,
-0.04090176522731781,
0.07663826644420624,
0.07032088935375214,
0.022478625178337097,
0.02928088791668415,
0.009486193768680096,
0.001665061921812594,
0.030520739033818245,
-0.035194817930459976,
0.06586035341024399,
0.1003500372171402,
-0.011690395884215832,
-0.01067431177943945,
-0.011003128252923489,
-0.015420692041516304,
0.08340124785900116,
0.031161904335021973,
0.05498620122671127,
-0.060404855757951736,
-0.03996226191520691,
-0.07087342441082001,
0.012045920826494694,
0.06478376686573029,
-0.03681385889649391,
0.011735711246728897,
-0.10125863552093506,
-0.024768486618995667,
0.037597186863422394,
-0.04723922908306122,
-0.020512336865067482,
3.29636497227203e-33,
-0.029950076714158058,
0.031192142516374588,
0.012950235046446323,
0.014731151983141899,
0.0781581774353981,
-0.004194782581180334,
-0.020336786285042763,
-0.054726846516132355,
0.013709528371691704,
-0.0044995504431426525,
-0.07071308046579361,
-0.04623647779226303,
-0.119853176176548,
0.12449526786804199,
-0.007362211123108864,
-0.047678276896476746,
-0.008561020717024803,
0.06789102405309677,
-0.07593134045600891,
0.0292362067848444,
-0.03286508470773697,
-0.02374170906841755,
-0.014625382609665394,
-0.07570252567529678,
0.010369145311415195,
0.022267920896410942,
0.05486823990941048,
0.00322708860039711,
0.01878870464861393,
0.009656386449933052,
-0.09177185595035553,
0.006927576381713152,
-0.02619236148893833,
-0.007813647389411926,
0.0003765529254451394,
0.011404508724808693,
0.04469004645943642,
-0.024729080498218536,
-0.020420173183083534,
-0.03182452917098999,
0.009920856915414333,
0.09895984083414078,
-0.05655890330672264,
-0.0035619258414953947,
-0.05782656371593475,
0.0022766259498894215,
-0.016839392483234406,
0.05924895778298378,
-0.007022390142083168,
-0.0196186825633049,
-0.009708421304821968,
0.04161882400512695,
-0.052919115871191025,
-0.07210355252027512,
-0.05837768688797951,
0.01302473433315754,
0.07505721598863602,
0.01850500889122486,
-0.0062660397961735725,
0.046721506863832474,
0.007685788907110691,
0.06617710739374161,
-0.034320008009672165,
0.007060305215418339,
0.04109928011894226,
0.02479591779410839,
-0.02396315522491932,
0.06780675798654556,
0.01740354299545288,
-0.008070782758295536,
-0.033081140369176865,
0.02627016231417656,
-0.07232626527547836,
-0.08959857374429703,
0.10609658807516098,
-0.08877786248922348,
0.07041273266077042,
0.07343495637178421,
0.021208250895142555,
-0.05934225767850876,
-0.1209649071097374,
0.031224334612488747,
-0.05519618093967438,
-0.09835992008447647,
-0.015308528207242489,
-0.028978753834962845,
0.01222464069724083,
-0.0627715215086937,
-0.019075796008110046,
-0.096216581761837,
-0.028427304700016975,
-0.00726692471653223,
-0.002690335502848029,
0.03592193126678467,
-0.0014384364476427436,
-2.9899187378004443e-33,
-0.014864501543343067,
0.11300982534885406,
0.020202919840812683,
0.10228638350963593,
-0.01055334322154522,
-0.014672836288809776,
0.011505776084959507,
0.046431686729192734,
0.1098339855670929,
-0.024432824924588203,
-0.015081729739904404,
-0.03228653222322464,
-0.04266756400465965,
-0.013095014728605747,
0.019477862864732742,
-0.01671954058110714,
0.02840448170900345,
-0.04974411055445671,
-0.05077211931347847,
-0.02255478873848915,
0.03913893550634384,
0.09070105105638504,
-0.014202209189534187,
-0.014842003583908081,
-0.06911780685186386,
0.08150564134120941,
-0.062011562287807465,
0.07209601998329163,
0.012265503406524658,
-0.04344964399933815,
-0.0007663203869014978,
-0.03240383416414261,
-0.048300933092832565,
-0.044209402054548264,
-0.04227926954627037,
0.08787433058023453,
0.04143616184592247,
0.02918715961277485,
0.0002050661132670939,
0.07846575230360031,
0.06891916692256927,
0.020669149234890938,
-0.08264320343732834,
0.058589108288288116,
-0.009478185325860977,
-0.023498009890317917,
-0.04619887098670006,
0.009953667409718037,
0.039474040269851685,
-0.025610122829675674,
-0.00553039088845253,
0.10578334331512451,
-0.052767183631658554,
0.0015389375621452928,
-0.02258719690144062,
-0.07640746235847473,
-0.09933953732252121,
-0.018600784242153168,
0.035514701157808304,
-0.010660311207175255,
-0.06357154250144958,
-0.0529206246137619,
-0.008028069511055946,
-0.05836787447333336,
-0.011030394583940506,
-0.013802101835608482,
-0.040668193250894547,
-0.002094724215567112,
0.06683246791362762,
0.04811672493815422,
-0.011850782670080662,
0.038358788937330246,
-0.024956541135907173,
0.025020673871040344,
0.04070223122835159,
-0.16378721594810486,
0.04275762662291527,
-0.03386412933468819,
-0.005346008110791445,
0.027735473588109016,
-0.03284759074449539,
0.059727150946855545,
0.03056829422712326,
0.03596330061554909,
0.06130661442875862,
-0.014952083118259907,
-0.013732818886637688,
-0.0405353419482708,
0.030255284160375595,
-0.012574083171784878,
0.0012490388471633196,
-0.021144989877939224,
0.042412836104631424,
0.0642489641904831,
-0.08766569942235947,
-5.0036344845239e-8,
-0.07445577532052994,
0.02158181741833687,
0.02451770193874836,
0.0390215702354908,
0.051877520978450775,
0.005844019819051027,
-0.022294308990240097,
0.09280630201101303,
-0.02184343710541725,
-0.09214762598276138,
0.061004363000392914,
0.0035686898045241833,
-0.022797677665948868,
0.01754419133067131,
-0.000609569891821593,
0.14350134134292603,
0.001463669235818088,
-0.0065753343515098095,
-0.0011333443690091372,
0.008712227456271648,
0.011732959188520908,
0.0460672564804554,
-0.060351233929395676,
0.07905545830726624,
0.08163738250732422,
-0.0040162657387554646,
0.0026490797754377127,
-0.023236175999045372,
0.028939398005604744,
0.01094467006623745,
-0.029018785804510117,
0.057349979877471924,
0.0713946595788002,
0.01908298023045063,
0.04397198185324669,
0.044988058507442474,
0.07258976995944977,
0.03170839697122574,
-0.03370268642902374,
-0.036742232739925385,
-0.03182896226644516,
0.07985533773899078,
0.018089694902300835,
-0.003172804368659854,
0.00794087816029787,
-0.021735956892371178,
0.041470594704151154,
-0.09292028844356537,
0.014725620858371258,
-0.026866797357797623,
-0.004131661262363195,
0.0005798956844955683,
-0.022865256294608116,
-0.016133718192577362,
0.06121186539530754,
-0.012562284246087074,
-0.03271816670894623,
-0.038141343742609024,
0.05497646704316139,
0.02579394355416298,
-0.04565000161528587,
0.0720687136054039,
0.06891614198684692,
-0.01200820878148079
] |
Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models | https://openreview.net/forum?id=iJi7nz5Cxc | [
"Fu-Yun Wang",
"Yunhao Shui",
"Jingtan Piao",
"Keqiang Sun",
"Hongsheng Li"
] | Poster | Diffusion models have made substantial advances in image generation, yet models trained on large, unfiltered datasets often yield outputs misaligned with human preferences. Numerous methods have already been proposed to fine-tune pre-trained diffusion models, achieving notable improvements in aligning generated outputs with human preferences. However, we point out that existing preference alignment methods neglect the critical role of handling unconditional/negative-conditional outputs, leading to a diminished capacity to avoid generating undesirable outcomes. This oversight limits the efficacy of classifier-free guidance (CFG), which relies on the contrast between conditional generation and unconditional/negative-conditional generation to optimize output quality. In response, we propose a straightforward but consistently effective approach that involves training a model specifically attuned to negative preferences. This method does not require new training strategies or datasets but rather involves minor modifications to existing techniques. Our approach integrates seamlessly with models such as SD15, SDXL, video diffusion models and models that have undergone preference optimization, consistently enhancing their ability to produce more human preferences aligned outputs. | Diffusion, Preference Optimization | null | 846 | null | [
-0.02433915063738823,
-0.1101001501083374,
-0.004261130001395941,
0.04668210819363594,
0.05984191969037056,
-0.006197409704327583,
-0.07254913449287415,
-0.022430764511227608,
0.03622923418879509,
-0.05025773122906685,
0.016199884936213493,
-0.09800934791564941,
0.01897868700325489,
0.02474549598991871,
-0.036730244755744934,
-0.009871229529380798,
0.09732364118099213,
0.007976164110004902,
-0.12283103913068771,
-0.0038312613032758236,
-0.035862311720848083,
-0.030628366395831108,
0.002046014182269573,
-0.06497233361005783,
-0.03436405584216118,
0.0015606899978592992,
0.07053422182798386,
-0.03897552192211151,
0.039831310510635376,
-0.08106427639722824,
0.05338167026638985,
-0.010828799568116665,
-0.015368549153208733,
-0.00021398351236712188,
-0.03386903554201126,
0.03135235980153084,
-0.10466067492961884,
-0.004934392869472504,
-0.01931626908481121,
-0.03629531338810921,
-0.04611843451857567,
0.00888154562562704,
-0.0011983129661530256,
0.017113734036684036,
0.012064286507666111,
-0.0041317567229270935,
-0.014491303823888302,
-0.04189351946115494,
0.020283197984099388,
-0.0034108965191990137,
-0.014387170784175396,
0.016025198623538017,
-0.06500610709190369,
0.017564233392477036,
-0.020491061732172966,
0.08492740988731384,
0.03900860622525215,
0.06783723831176758,
0.01144266314804554,
-0.06020403280854225,
-0.08254047483205795,
-0.15095196664333344,
-0.021782906726002693,
-0.02756892889738083,
0.005752872675657272,
0.023048045113682747,
-0.018220029771327972,
0.07101375609636307,
-0.039475005120038986,
0.07401858270168304,
-0.05668945983052254,
0.06897763162851334,
-0.02420094981789589,
0.01862325519323349,
-0.042290326207876205,
0.07100503146648407,
0.09138227254152298,
-0.016935881227254868,
0.027017731219530106,
-0.039929162710905075,
0.028010167181491852,
0.0025101660285145044,
0.04326081648468971,
0.051867734640836716,
0.08778481185436249,
0.006338298786431551,
-0.015962617471814156,
0.07309513539075851,
0.0078575499355793,
0.016441484913229942,
-0.12740279734134674,
0.04198240116238594,
-0.018311869353055954,
-0.023306993767619133,
-0.025274986401200294,
0.07674388587474823,
-0.03683234378695488,
-0.10089964419603348,
-0.04609788581728935,
0.05830204859375954,
0.006030640099197626,
0.04015360400080681,
-0.02488395757973194,
-0.050460755825042725,
0.08430473506450653,
-0.10084173083305359,
0.013213294558227062,
0.06740113347768784,
0.05479447543621063,
-0.04321528598666191,
0.006150145083665848,
0.08548112213611603,
-0.009426625445485115,
-0.04061645641922951,
0.018189378082752228,
-0.04431803151965141,
0.01248414907604456,
0.02764005772769451,
0.003713209182024002,
0.03934700787067413,
-0.014832713641226292,
-0.030503712594509125,
-0.08304999768733978,
0.03148908540606499,
0.010545982979238033,
-0.04105045646429062,
-0.04209868982434273,
4.101085629641921e-33,
-0.07709486037492752,
-0.03379375860095024,
0.0723731592297554,
0.003444413188844919,
0.11561638116836548,
0.018375173211097717,
-0.01832677237689495,
-0.02861829102039337,
-0.054474037140607834,
-0.04629074037075043,
0.015905333682894707,
-0.03431319072842598,
-0.11914709210395813,
0.1663856953382492,
0.051155589520931244,
-0.02387404255568981,
-0.05139283835887909,
0.039657752960920334,
-0.051256969571113586,
0.043497826904058456,
0.022549420595169067,
0.030066868290305138,
-0.01279638521373272,
-0.056587208062410355,
-0.020546289160847664,
-0.027507882565259933,
0.06021884083747864,
-0.06534173339605331,
-0.018927287310361862,
-0.025243183597922325,
-0.10815668851137161,
-0.013988076709210873,
0.07478001713752747,
-0.021734656766057014,
0.040016382932662964,
0.016236215829849243,
-0.09569123387336731,
0.004026402253657579,
-0.040049321949481964,
0.0019559101201593876,
-0.012715911492705345,
0.11188898235559464,
-0.014423862099647522,
0.04247400537133217,
-0.09323522448539734,
-0.012903168797492981,
0.054668109863996506,
0.008573517203330994,
-0.028868241235613823,
-0.05457526445388794,
0.06631875783205032,
0.007638365961611271,
0.011252895928919315,
-0.07524438202381134,
-0.06808208674192429,
-0.038129426538944244,
0.07348956912755966,
0.013636902906000614,
0.003585424041375518,
-0.07795001566410065,
0.06851904094219208,
0.05483260378241539,
-0.03737961873412132,
-0.0027151445392519236,
-0.003809846704825759,
-0.02207092009484768,
0.017595719546079636,
0.04597173258662224,
0.025469819083809853,
-0.016678377985954285,
-0.01827600784599781,
0.055416930466890335,
-0.04436860606074333,
-0.09032220393419266,
0.04386868327856064,
0.044151242822408676,
0.04128788784146309,
-0.036804068833589554,
0.0902610719203949,
-0.056751102209091187,
0.03838860243558884,
-0.004410856869071722,
0.024436097592115402,
-0.12486451119184494,
0.014305206947028637,
-0.05897270515561104,
0.003988528158515692,
-0.04172166809439659,
-0.044861551374197006,
-0.01642772927880287,
-0.008603000082075596,
0.0008199838921427727,
-0.03040456771850586,
0.06762173771858215,
0.01046728529036045,
-3.836283119258138e-33,
-0.005064389202743769,
0.07293664664030075,
0.03611006960272789,
0.039567817002534866,
-0.02548634074628353,
0.06816983222961426,
0.01886242814362049,
0.001955926651135087,
0.07849827408790588,
-0.07662806659936905,
0.0010629871394485235,
-0.017095843330025673,
-0.01328163594007492,
-0.01798764243721962,
-0.038199782371520996,
-0.024808740243315697,
-0.007992619648575783,
-0.008202550932765007,
-0.027768153697252274,
-0.031860705465078354,
-0.016793889924883842,
0.10069211572408676,
-0.04062101989984512,
0.073104627430439,
-0.057978127151727676,
0.05463700741529465,
0.024416230618953705,
0.08904238790273666,
-0.048424433916807175,
-0.03603188320994377,
-0.050135642290115356,
0.046812284737825394,
0.014710502699017525,
-0.03477412462234497,
-0.040798209607601166,
0.03671950101852417,
0.03302707150578499,
0.03368247672915459,
-0.022743701934814453,
0.16930285096168518,
-0.005352749489247799,
-0.01460151094943285,
-0.06746039539575577,
0.0956430658698082,
-0.06261657178401947,
-0.03522254526615143,
-0.027017081156373024,
0.00019039688049815595,
0.054412491619586945,
-0.00362299638800323,
-0.0032110062893480062,
0.09573356807231903,
-0.07348925620317459,
0.05764847993850708,
-0.01224837452173233,
-0.08989749848842621,
-0.08032679557800293,
-0.06814666092395782,
0.1235228031873703,
0.036799367517232895,
-0.0234003234654665,
-0.017109785228967667,
0.004205487202852964,
-0.11874030530452728,
-0.07614529132843018,
-0.028683701530098915,
-0.0062606302089989185,
0.00631909491494298,
0.011725598014891148,
0.062445830553770065,
0.00727100670337677,
0.006297450512647629,
-0.02038555033504963,
-0.03676047921180725,
0.0012007649056613445,
-0.06747668236494064,
0.055089280009269714,
-0.021565735340118408,
0.02590300515294075,
-0.03440924733877182,
-0.030387133359909058,
0.043865080922842026,
0.01850520260632038,
0.002028567949309945,
0.055795103311538696,
0.05626412108540535,
0.0556161105632782,
0.04682168737053871,
0.026547148823738098,
0.0576988086104393,
-0.02926076389849186,
0.017425838857889175,
0.03809504956007004,
0.07512368261814117,
-0.06037850305438042,
-4.491389660188361e-8,
-0.07171531021595001,
-0.01621313951909542,
0.05145338550209999,
0.10066121071577072,
-0.010473139584064484,
0.013394853100180626,
0.06585419923067093,
0.07783295959234238,
0.013302575796842575,
-0.01703932136297226,
0.004626624286174774,
0.08762640506029129,
-0.05508212000131607,
-0.03541349247097969,
-0.023106848821043968,
0.09505216032266617,
0.025861606001853943,
0.0235928762704134,
0.028782395645976067,
0.04477745294570923,
0.017126141116023064,
-0.0009465194307267666,
0.04497118666768074,
-0.017379315569996834,
0.10601339489221573,
-0.07493845373392105,
-0.029916882514953613,
0.03100869059562683,
0.07904551178216934,
0.016664642840623856,
-0.005247048567980528,
0.0038813273422420025,
-0.022730598226189613,
-0.07390724867582321,
0.036457359790802,
0.10639426857233047,
-0.02686961740255356,
-0.007037246599793434,
-0.0074621145613491535,
-0.013050392270088196,
0.010134922340512276,
0.08270645141601562,
-0.009415666572749615,
0.011231407523155212,
-0.049342408776283264,
0.0017731154803186655,
0.09613252431154251,
-0.06476692110300064,
0.049022458493709564,
-0.021084140986204147,
0.014788191765546799,
-0.06786233931779861,
-0.008497055619955063,
0.065847247838974,
0.037281583994627,
-0.04741169884800911,
-0.00806828960776329,
-0.02313460409641266,
0.035712290555238724,
0.1095765009522438,
0.03579523786902428,
-0.009424532763659954,
-0.032050374895334244,
0.013146298937499523
] |
Asymmetric Factorized Bilinear Operation for Vision Transformer | https://openreview.net/forum?id=MJyqwBVgMs | [
"Junjie Wu",
"Qilong Wang",
"Jiangtao Xie",
"Pengfei Zhu",
"Qinghua Hu"
] | Poster | As a core component of Transformer-like deep architectures, a feed-forward network (FFN) for channel mixing is responsible for learning features of each token. Recent works show channel mixing can be enhanced by increasing computational burden or can be slimmed at the sacrifice of performance. Although some efforts have been made, existing works are still struggling to solve the paradox of performance and complexity trade-offs. In this paper, we propose an Asymmetric Factorized Bilinear Operation (AFBO) to replace FFN of vision transformer (ViT), which attempts to efficiently explore rich statistics of token features for achieving better performance and complexity trade-off. Specifically, our AFBO computes second-order statistics via a spatial-channel factorized bilinear operation for feature learning, which replaces a simple linear projection in FFN and enhances the feature learning ability of ViT by modeling second-order correlation among token features. Furthermore, our AFBO presents two structured-sparsity channel mapping strategies, namely Grouped Cross Channel Mapping (GCCM) and Overlapped Cycle Channel Mapping (OCCM). They decompose bilinear operation into grouped channel features by considering information interaction between groups, significantly reducing computational complexity while guaranteeing model performance. Finally, our AFBO is built with GCCM and OCCM in an asymmetric way, aiming to achieve a better trade-off. Note that our AFBO is model-agnostic, which can be flexibly integrated with existing ViTs. Experiments are conducted with twenty ViTs on various tasks, and the results show our AFBO is superior to its counterparts while improving existing ViTs in terms of generalization and robustness. | Vision Transformer, Channel Mixer, Factorized Bilinear Operation | null | 843 | null | [
0.02287612110376358,
-0.11355960369110107,
0.050881512463092804,
0.012340117245912552,
-0.009926517494022846,
-0.013458010740578175,
-0.06686519086360931,
-0.029626356437802315,
0.01911953091621399,
-0.08156953752040863,
-0.06671734154224396,
-0.010665555484592915,
-0.027213379740715027,
0.06313935667276382,
-0.0159451924264431,
0.06615447998046875,
0.03532114997506142,
0.15025800466537476,
-0.11999041587114334,
-0.054647207260131836,
0.02131268009543419,
-0.04245442524552345,
-0.0486951619386673,
-0.023854032158851624,
0.08602474629878998,
0.020983753725886345,
0.02303461544215679,
-0.07165059447288513,
-0.004607868380844593,
-0.039567213505506516,
-0.012317623943090439,
0.06478998064994812,
-0.009776302613317966,
-0.05337965115904808,
-0.056370172649621964,
-0.0025269368197768927,
0.06276310980319977,
-0.020724650472402573,
0.043753307312726974,
-0.008230594918131828,
0.029936784878373146,
0.015803027898073196,
0.04861621558666229,
0.020963426679372787,
0.04270639643073082,
0.09905873239040375,
0.07770320028066635,
-0.03862062096595764,
0.019152995198965073,
0.000447731843451038,
-0.012692137621343136,
0.0289994515478611,
-0.08381012082099915,
0.08799947053194046,
-0.05896429717540741,
0.024030013009905815,
-0.0008947667665779591,
-0.006863293703645468,
-0.037005092948675156,
0.004572002682834864,
-0.0038747231010347605,
-0.035143159329891205,
0.060918401926755905,
0.01986548863351345,
-0.06252972781658173,
-0.07003064453601837,
0.06921199709177017,
-0.01184395793825388,
-0.0018141682958230376,
-0.019880332052707672,
-0.0440078005194664,
0.03981010988354683,
-0.0497230626642704,
-0.017565904185175896,
0.07485353201627731,
0.052601732313632965,
0.10827981680631638,
-0.03009720705449581,
0.062492139637470245,
-0.07845064997673035,
0.07447397708892822,
0.015691692009568214,
0.04728252440690994,
-0.03890620917081833,
0.15529018640518188,
-0.03153940662741661,
-0.05543899163603783,
0.016486797481775284,
-0.04697015881538391,
0.017816560342907906,
-0.11822184175252914,
-0.017744576558470726,
0.075539730489254,
-0.10344003140926361,
0.028556758537888527,
-0.021882589906454086,
-0.0037098126485943794,
-0.05891772359609604,
-0.05071790888905525,
0.0898144468665123,
-0.059334658086299896,
-0.05039481818675995,
0.08068768680095673,
0.006585196126252413,
0.017734985798597336,
-0.024029839783906937,
0.027901843190193176,
0.07242020219564438,
0.039095815271139145,
-0.038943372666835785,
0.039170462638139725,
0.029197154566645622,
0.002174782333895564,
-0.033615708351135254,
0.02540874294936657,
0.06495628505945206,
-0.00758324284106493,
0.01752135530114174,
0.09273196011781693,
0.032544560730457306,
-0.018146660178899765,
-0.023528117686510086,
-0.04582076147198677,
0.06618107855319977,
-0.0032595000229775906,
0.043604571372270584,
0.0003475262492429465,
4.4346505187869764e-33,
-0.02974018268287182,
0.04154317453503609,
0.0573592483997345,
-0.041254185140132904,
-0.0006607240648008883,
0.012624911963939667,
0.018421929329633713,
0.027181215584278107,
0.01231472659856081,
0.03621306270360947,
-0.017456134781241417,
0.006669721566140652,
-0.017867887392640114,
0.09158538281917572,
0.012905986979603767,
-0.07171064615249634,
-0.063936747610569,
-0.040081776678562164,
0.001909396843984723,
-0.0535108707845211,
0.013230239041149616,
-0.03581823408603668,
-0.01627924107015133,
0.016507219523191452,
0.01984337344765663,
-0.07212453335523605,
0.04042022302746773,
-0.03540077432990074,
-0.03947671130299568,
-0.002148493891581893,
-0.0859212875366211,
0.014812999404966831,
-0.02524409629404545,
-0.050523947924375534,
0.04606890305876732,
-0.11416278034448624,
0.0451849065721035,
-0.059713467955589294,
-0.00015893355885054916,
-0.0071968925185501575,
-0.023199107497930527,
0.030062828212976456,
-0.08379995077848434,
-0.060385700315237045,
-0.05811162292957306,
-0.0522051565349102,
0.030306266620755196,
0.07506170868873596,
-0.003433243604376912,
0.010555088520050049,
0.03040909394621849,
-0.014452384784817696,
-0.06747417896986008,
0.020583823323249817,
0.11843429505825043,
-0.008432758040726185,
0.011090912856161594,
0.03693830594420433,
0.07608620822429657,
0.07022880762815475,
-0.09703324735164642,
0.031711410731077194,
-0.014751177281141281,
0.07033875584602356,
-0.00804110150784254,
-0.018514808267354965,
0.08673083782196045,
0.07655290514230728,
-0.0261760875582695,
-0.09291408210992813,
-0.025319330394268036,
0.0785483568906784,
-0.05907444283366203,
-0.07659907639026642,
0.0318894237279892,
0.013686850666999817,
0.031857531517744064,
-0.055480264127254486,
-0.03497575223445892,
0.02399553917348385,
-0.02835419774055481,
0.07323002070188522,
0.04416176304221153,
0.029862355440855026,
-0.010821369476616383,
0.05002158135175705,
0.03327130153775215,
-0.07939819246530533,
-0.06495960801839828,
-0.0034445219207555056,
-0.029934361577033997,
0.0016575201880186796,
-0.008715275675058365,
0.029787922278046608,
0.02622104063630104,
-3.31056788276584e-33,
-0.057083550840616226,
1.2606267318915343e-7,
0.0309155136346817,
0.02743983268737793,
-0.024826334789395332,
-0.09863316267728806,
0.0004782217729371041,
-0.085137277841568,
0.061447903513908386,
-0.04561125487089157,
-0.02324254997074604,
-0.009901834651827812,
-0.037697892636060715,
-0.053012389689683914,
0.07780848443508148,
0.015393031761050224,
-0.06045108661055565,
0.06744375824928284,
-0.013929396867752075,
-0.017845647409558296,
0.036000557243824005,
0.06279663741588593,
-0.0018586396472528577,
0.05694740638136864,
-0.12943708896636963,
0.009840825572609901,
-0.015827316790819168,
0.03435371071100235,
0.057907942682504654,
-0.004327775444835424,
0.04644111171364784,
-0.024093929678201675,
0.016759753227233887,
-0.04778707027435303,
-0.04223638400435448,
0.057894591242074966,
0.013569775968790054,
-0.00010582159302430227,
-0.03442321717739105,
0.009243203327059746,
0.0375530831515789,
-0.03441603481769562,
-0.06733816862106323,
-0.04271114245057106,
-0.032726533710956573,
0.012591025792062283,
0.053950171917676926,
-0.028169799596071243,
-0.009722069837152958,
0.03490377962589264,
0.01229453831911087,
0.023591874167323112,
-0.11856617778539658,
-0.013301901519298553,
-0.06844054907560349,
-0.004330968484282494,
0.013499750755727291,
0.10026877373456955,
0.1457415074110031,
-0.037354473024606705,
0.03447046875953674,
-0.04713710397481918,
0.02938740700483322,
-0.10196776688098907,
0.05170711502432823,
-0.020342351868748665,
0.018059473484754562,
0.0035042501986026764,
0.05674374848604202,
0.0893140509724617,
0.005128071177750826,
-0.008143576793372631,
-0.02624940313398838,
0.03966116905212402,
0.01885354146361351,
-0.002079663099721074,
0.08451443910598755,
-0.08574669808149338,
-0.0018186499364674091,
0.06420298665761948,
-0.035547174513339996,
-0.045980267226696014,
-0.00672565633431077,
0.07383384555578232,
0.045422039926052094,
0.12065586447715759,
0.04311806708574295,
0.013585257343947887,
0.07145213335752487,
-0.04580365866422653,
-0.028819670900702477,
0.05411556735634804,
-0.07969021797180176,
0.06124516576528549,
0.03699526935815811,
-4.371663209212784e-8,
-0.044932279735803604,
-0.003320010146126151,
-0.034329500049352646,
-0.05528843030333519,
-0.004553413949906826,
-0.10205778479576111,
0.0883016586303711,
0.061395976692438126,
-0.04815886169672012,
-0.06685717403888702,
0.05958029255270958,
0.02109036222100258,
-0.0494995079934597,
-0.04058662801980972,
0.03462793678045273,
0.03591987490653992,
0.04441588744521141,
-0.027497604489326477,
0.019888075068593025,
-0.023898251354694366,
-0.0074312807992100716,
0.011264486238360405,
0.05274316295981407,
-0.04068254306912422,
-0.0568995326757431,
-0.06309134513139725,
-0.013814765959978104,
0.07507845759391785,
0.14333787560462952,
-0.030900297686457634,
0.06466461718082428,
0.04316040873527527,
0.03895602002739906,
-0.0636720061302185,
-0.036285024136304855,
-0.016351010650396347,
0.0658850148320198,
0.009491394273936749,
-0.05944092199206352,
0.006393337622284889,
-0.0012940737651661038,
0.024241378530859947,
0.020117497071623802,
0.04983564838767052,
-0.06544513255357742,
0.02694660797715187,
0.03052709996700287,
-0.08768518269062042,
0.004944298882037401,
-0.04660317674279213,
0.03699873015284538,
0.04105324298143387,
0.005446971859782934,
0.08042211830615997,
-0.022072123363614082,
0.00328996148891747,
0.024486428126692772,
-0.0933784618973732,
0.04385614022612572,
0.035593871027231216,
-0.03742355480790138,
-0.010388500988483429,
-0.07929789274930954,
-0.021737542003393173
] |
ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY | https://openreview.net/forum?id=OheAR2xrtb | [
"Chenrui Tie",
"Yue Chen",
"Ruihai Wu",
"Boxuan Dong",
"Zeyi Li",
"Chongkai Gao",
"Hao Dong"
] | Poster | Imitation learning, e.g., diffusion policy, has been proven effective in various robotic manipulation tasks.
However, extensive demonstrations are required for policy robustness and generalization.
To reduce the demonstration reliance, we leverage spatial symmetry and propose ET-SEED, an efficient trajectory-level SE(3) equivariant diffusion model for generating action sequences in complex robot manipulation tasks.
Further, previous equivariant diffusion models require the per-step equivariance in the Markov process, making it difficult to learn policy under such strong constraints.
We theoretically extend equivariant Markov kernels and simplify the condition of equivariant diffusion process, thereby significantly improving training efficiency for trajectory-level SE(3) equivariant diffusion policy in an end-to-end manner.
We evaluate ET-SEED on representative robotic manipulation tasks, involving rigid body, articulated and deformable object.
Experiments demonstrate superior data efficiency and manipulation proficiency of our proposed method,
as well as its ability to generalize to unseen configurations with only a few demonstrations. Website: https://et-seed.github.io/ | Robotics; Manipulation; Equivariance | We introduces ET-SEED, an SE(3) equivariant diffusion model that leverages spatial symmetries to improve data efficiency and spatial generalization in robotic manipulation tasks while reducing training complexity. | 840 | null | [
-0.028090080246329308,
-0.11287518590688705,
0.07208744436502457,
-0.03060705028474331,
0.043452031910419464,
-0.0008028147276490927,
-0.006004280876368284,
-0.05280004441738129,
0.02500738948583603,
0.05566171184182167,
0.03890678286552429,
-0.010772640816867352,
-0.0022814820986241102,
0.04196155071258545,
0.04657987132668495,
-0.010045042261481285,
0.06868687272071838,
0.056844037026166916,
-0.0662800520658493,
-0.005848910193890333,
0.018791211768984795,
0.005154763348400593,
0.1367608606815338,
0.009775491431355476,
0.019481247290968895,
0.03636839613318443,
0.07154668122529984,
-0.07827679067850113,
-0.030035024508833885,
-0.05993632599711418,
0.025768466293811798,
-0.0786939486861229,
-0.023468460887670517,
-0.10158095508813858,
-0.051386989653110504,
0.0803995281457901,
-0.07603469491004944,
0.004009542521089315,
-0.0250301044434309,
-0.05489593371748924,
0.005341026932001114,
-0.0019081714563071728,
0.049727801233530045,
0.031244385987520218,
0.05788420885801315,
0.07740551233291626,
0.01272209919989109,
-0.11889934539794922,
0.07626355439424515,
0.010514029301702976,
0.022314272820949554,
-0.003240980440750718,
-0.005911452230066061,
-0.016641708090901375,
-0.012949117459356785,
-0.011601578444242477,
0.038554973900318146,
-0.028262149542570114,
0.10481087863445282,
-0.07022341340780258,
0.03553197532892227,
-0.1026010811328888,
0.01063645537942648,
-0.023723941296339035,
-0.09568466991186142,
-0.075501449406147,
0.07351414859294891,
-0.01725073903799057,
-0.00011308239481877536,
0.031827665865421295,
-0.02471933886408806,
0.02272551693022251,
-0.03300754725933075,
0.0448264554142952,
-0.0019579806830734015,
0.03576747328042984,
-0.00597265362739563,
-0.044146064668893814,
-0.03765290975570679,
-0.015693211928009987,
0.029946882277727127,
-0.011117769405245781,
0.06297456473112106,
0.04012821614742279,
-0.04299217090010643,
-0.042044587433338165,
-0.007866814732551575,
0.05710779130458832,
0.08541078865528107,
-0.022548574954271317,
0.05812276527285576,
0.08881599456071854,
0.010112281888723373,
-0.08302102982997894,
0.02120283991098404,
-0.05645686015486717,
-0.003256213152781129,
-0.07389968633651733,
-0.016632206737995148,
0.048724059015512466,
0.005883772857487202,
0.05327477306127548,
-0.025284407660365105,
0.030952222645282745,
0.016104329377412796,
-0.07580486685037613,
-0.025858985260128975,
-0.07371240109205246,
0.02588481269776821,
0.012681091204285622,
0.03716116398572922,
-0.06196816638112068,
0.038782186806201935,
0.05077935382723808,
0.021311936900019646,
-0.002424852456897497,
-0.02308570221066475,
0.008721530437469482,
-0.004872831050306559,
0.08877240866422653,
0.01649225689470768,
-0.06180781498551369,
-0.035563305020332336,
0.06814804673194885,
0.06381009519100189,
-0.025485934689641,
-0.01945558376610279,
-1.5037299113309248e-34,
0.08656661212444305,
0.002934081247076392,
0.06514724344015121,
0.03731841221451759,
0.02010386995971203,
0.032679393887519836,
0.0556790828704834,
-0.027201315388083458,
-0.05130655691027641,
-0.03378314897418022,
-0.059651196002960205,
-0.021343762055039406,
-0.09765667468309402,
0.11579182744026184,
0.01878228411078453,
0.00792278815060854,
-0.00356982066296041,
0.0368315726518631,
-0.03146630898118019,
0.023491961881518364,
0.054403580725193024,
-0.04665546119213104,
-0.08007203787565231,
-0.04938463494181633,
0.0031177522614598274,
-0.03942183405160904,
0.004480740986764431,
-0.019098130986094475,
0.0257932897657156,
-0.01921759732067585,
-0.05976860970258713,
-0.013210191391408443,
-0.0989183858036995,
0.08440422266721725,
0.034502480179071426,
0.02726888097822666,
-0.023094775155186653,
-0.06687361001968384,
0.02607986330986023,
-0.04514314979314804,
-0.018137270584702492,
-0.016170676797628403,
0.025634881108999252,
-0.017991147935390472,
-0.11780890822410583,
-0.023459767922759056,
0.06233077868819237,
0.03464222699403763,
-0.03169883042573929,
-0.03626292571425438,
0.027301451191306114,
0.05108189955353737,
0.05471375584602356,
-0.0977608934044838,
-0.01087732333689928,
-0.019636059179902077,
0.05418021231889725,
0.03341158479452133,
-0.014381347224116325,
-0.0045879920944571495,
0.05634254589676857,
0.02238469384610653,
-0.04730233550071716,
0.04731632024049759,
0.0150150116533041,
0.05363544449210167,
-0.06462056934833527,
0.1012757420539856,
0.10175562649965286,
0.011282370425760746,
-0.03546605259180069,
0.015315966680645943,
-0.03734953701496124,
-0.13567928969860077,
0.0706208273768425,
0.01841798983514309,
0.0019744953606277704,
-0.08579142391681671,
0.010103849694132805,
0.058776646852493286,
-0.09542544186115265,
-0.016660291701555252,
-0.05391364544630051,
-0.03474288061261177,
0.02709270268678665,
-0.0237234216183424,
0.009157261811196804,
-0.06687990576028824,
0.005997680593281984,
-0.05521759018301964,
-0.021849580109119415,
-0.11547388136386871,
-0.0951283648610115,
0.08287165313959122,
0.0032961913384497166,
-1.402108998674189e-33,
0.06141647323966026,
-0.012744356878101826,
0.0436711423099041,
0.10018610954284668,
-0.04977304860949516,
-0.024812903255224228,
-0.006439217831939459,
-0.028326617553830147,
0.016518164426088333,
-0.07378142327070236,
-0.06133219599723816,
0.002851941157132387,
0.010746938176453114,
-0.037839993834495544,
0.03630159795284271,
-0.013790912926197052,
-0.0159828532487154,
0.058669719845056534,
-0.00817982479929924,
0.029836678877472878,
-0.025880618020892143,
0.05961756408214569,
-0.012951644137501717,
0.01208875235170126,
-0.03828011825680733,
0.009961620904505253,
0.025950640439987183,
0.043785542249679565,
0.0035414390731602907,
0.03810059651732445,
-0.04074908792972565,
-0.029714185744524002,
0.022742729634046555,
0.07301755994558334,
-0.0832676887512207,
0.06026949733495712,
0.03267378732562065,
0.0658411905169487,
0.009566521272063255,
0.05209861695766449,
-0.016980845481157303,
-0.054051101207733154,
0.004199068993330002,
0.08455145359039307,
0.002225975738838315,
-0.09768715500831604,
0.00029447051929309964,
0.0523136630654335,
-0.03366342931985855,
0.017588404938578606,
-0.0636463612318039,
0.12170297652482986,
-0.14009296894073486,
-0.034023918211460114,
-0.0021235193125903606,
-0.029101183637976646,
-0.03428879380226135,
-0.04254745692014694,
0.04196946695446968,
-0.04258882626891136,
-0.0748380497097969,
0.05030488595366478,
0.005933624226599932,
0.07683495432138443,
-0.03377116471529007,
0.06085524335503578,
-0.046923547983169556,
0.0006529797101393342,
0.06064007431268692,
0.02400859445333481,
0.016993477940559387,
-0.01653837226331234,
0.04343656823039055,
-0.061046574264764786,
0.11150249093770981,
-0.04096211493015289,
0.04422488436102867,
-0.028161773458123207,
-0.009266726672649384,
-0.018497129902243614,
0.018049679696559906,
-0.003349924925714731,
0.05781074985861778,
0.06399960070848465,
-0.036159299314022064,
0.08085285872220993,
-0.023001769557595253,
0.041843757033348083,
-0.029803264886140823,
0.04440879449248314,
-0.026942482218146324,
-0.01074075885117054,
0.049428556114435196,
0.09683065861463547,
-0.018889015540480614,
-4.117118024282718e-8,
-0.055838845670223236,
0.06839966028928757,
0.135953888297081,
0.019611451774835587,
0.07750189304351807,
0.04815266653895378,
0.017598522827029228,
0.008850142359733582,
-0.05358083173632622,
0.015345346182584763,
0.03902611881494522,
-0.01202575396746397,
0.004976149648427963,
-0.058175403624773026,
-0.01023543905466795,
0.06045018136501312,
0.020730916410684586,
0.01571488566696644,
-0.0003374502994120121,
0.04685492441058159,
-0.01561732031404972,
-0.0589735247194767,
0.040093500167131424,
-0.01165422797203064,
0.03544441983103752,
-0.012575379572808743,
-0.1122719869017601,
0.0041646952740848064,
0.09313315153121948,
-0.00015282401000149548,
0.011833899654448032,
-0.0942581370472908,
-0.004409960936754942,
0.023552387952804565,
-0.01865663006901741,
0.012497013434767723,
-0.003983859438449144,
-0.03551063314080238,
-0.02065429277718067,
-0.036175038665533066,
0.006099693011492491,
0.053863804787397385,
0.00987091101706028,
-0.0005508549511432648,
-0.053068630397319794,
0.07544247806072235,
-0.07032421231269836,
-0.1627749353647232,
0.023777760565280914,
-0.05232705548405647,
0.0070472643710672855,
-0.0517616793513298,
-0.029187945649027824,
0.10707972198724747,
0.06830685585737228,
0.08071347326040268,
0.005238346289843321,
-0.05080988258123398,
0.027158210054039955,
0.10383192449808121,
-0.011500298045575619,
-0.00044205356971360743,
-0.05021294206380844,
-0.06773287057876587
] |
An Optimal Discriminator Weighted Imitation Perspective for Reinforcement Learning | https://openreview.net/forum?id=9JtG4nN7ql | [
"Haoran Xu",
"Shuozhe Li",
"Harshit Sikchi",
"Scott Niekum",
"Amy Zhang"
] | Poster | We introduce Iterative Dual Reinforcement Learning (IDRL), a new method that takes an optimal discriminator-weighted imitation view of solving RL. Our method is motivated by a simple experiment in which we find training a discriminator using the offline dataset plus an additional expert dataset and then performing discriminator-weighted behavior cloning gives strong results on various types of datasets. That optimal discriminator weight is quite similar to the learned visitation distribution ratio in Dual-RL, however, we find that current Dual-RL methods do not correctly estimate that ratio. In IDRL, we propose a correction method to iteratively approach the optimal visitation distribution ratio in the offline dataset given no addtional expert dataset. During each iteration, IDRL removes zero-weight suboptimal transitions using the learned ratio from the previous iteration and runs Dual-RL on the remaining subdataset. This can be seen as replacing the behavior visitation distribution with the optimized visitation distribution from the previous iteration, which theoretically gives a curriculum of improved visitation distribution ratios that are closer to the optimal discriminator weight. We verify the effectiveness of IDRL on various kinds of offline datasets, including D4RL datasets and more realistic corrupted demonstrations. IDRL beats strong Primal-RL and Dual-RL baselines in terms of both performance and stability, on all datasets. | imitation learning, offline RL, deep RL, dual RL | An optimal discriminator-weighted imitation view of solving Reinforcement Learning. | 838 | null | [
-0.11687751859426498,
-0.09903363138437271,
0.06459785252809525,
-0.014115354046225548,
-0.03711090236902237,
-0.004215447697788477,
0.025578605011105537,
0.04426460340619087,
0.010284238494932652,
0.028572000563144684,
-0.006230469793081284,
-0.008873127400875092,
0.03685091435909271,
0.046373602002859116,
-0.012506337836384773,
0.051318924874067307,
0.12188965082168579,
0.052509404718875885,
-0.0941070169210434,
-0.020458810031414032,
0.003828159999102354,
-0.022418517619371414,
0.03706140071153641,
-0.015953045338392258,
0.00046558791655115783,
-0.004238713067024946,
-0.01514306291937828,
-0.022522110491991043,
0.030590200796723366,
-0.03326285257935524,
0.0321451798081398,
0.01102123036980629,
0.06808806210756302,
-0.08774978667497635,
-0.057670339941978455,
0.037441059947013855,
-0.08955967426300049,
-0.007880930788815022,
0.035979628562927246,
-0.028863297775387764,
-0.0018348194425925612,
0.06734291464090347,
-0.013464747928082943,
0.02684742957353592,
0.022223342210054398,
0.017364224418997765,
-0.041481468826532364,
-0.06317728757858276,
0.08715201169252396,
0.04444665461778641,
0.050213128328323364,
0.02855401672422886,
0.02801838330924511,
0.0409788116812706,
0.02894132025539875,
0.03385613113641739,
0.07235299050807953,
0.01416152622550726,
0.05622170493006706,
-0.016478052362799644,
0.010344631969928741,
-0.07897795736789703,
-0.022573987022042274,
-0.018665466457605362,
-0.09153519570827484,
-0.06144038960337639,
-0.021653803065419197,
0.02835031785070896,
0.050254013389348984,
0.019662894308567047,
-0.05481419339776039,
0.05391333997249603,
-0.028432486578822136,
0.03324563056230545,
0.05052419379353523,
-0.05998065322637558,
0.025029851123690605,
-0.06848964840173721,
0.006411026697605848,
-0.0940977931022644,
-0.01890527456998825,
0.009874481707811356,
0.006931256037205458,
0.009657124988734722,
0.0987463891506195,
-0.007191249169409275,
-0.018066536635160446,
0.06319893151521683,
0.14648181200027466,
0.050123415887355804,
0.047643501311540604,
0.04488472640514374,
0.013696229085326195,
-0.013347592204809189,
-0.0804649218916893,
-0.0019182029645889997,
-0.03699195012450218,
-0.0017751974519342184,
0.014029258862137794,
0.0714145302772522,
-0.0021323247347027063,
0.04129889979958534,
0.01086867693811655,
-0.012029862962663174,
-0.0272202268242836,
-0.08919616043567657,
0.06199437752366066,
0.0010309509234502912,
0.058958083391189575,
-0.08364805579185486,
0.02676413580775261,
-0.008045507594943047,
0.009947395883500576,
0.1033347100019455,
-0.044748030602931976,
0.027874264866113663,
0.024100858718156815,
0.00023372455325443298,
-0.1054617166519165,
0.07309392094612122,
-0.02766268327832222,
-0.10148806869983673,
0.053830962628126144,
0.06331212818622589,
-0.0024143350310623646,
-0.04297475889325142,
-0.03188133239746094,
4.140307468025045e-33,
0.07572417706251144,
0.012296469882130623,
0.05150814354419708,
-0.06096817925572395,
0.03847183659672737,
-0.03271235525608063,
0.04618850350379944,
-0.006797652691602707,
-0.07845450937747955,
-0.016386551782488823,
0.021484658122062683,
-0.008904136717319489,
-0.04733040928840637,
0.07057010382413864,
0.007622972130775452,
-0.007973234169185162,
-0.04146961495280266,
0.0009874976240098476,
-0.0059811826795339584,
-0.06450121104717255,
0.06393826007843018,
-0.028123045340180397,
0.01349506713449955,
-0.059503957629203796,
0.02961334027349949,
-0.03751884773373604,
0.015322446823120117,
-0.04638705030083656,
0.02391722798347473,
0.030451316386461258,
-0.07483618706464767,
-0.020250216126441956,
-0.019058508798480034,
0.09052359312772751,
0.044085606932640076,
0.037936922162771225,
0.038440681993961334,
-0.02892891690135002,
-0.026566877961158752,
0.002412032335996628,
-0.0168930497020483,
-0.026445580646395683,
0.030339566990733147,
-0.005828142166137695,
-0.0681152269244194,
-0.12456435710191727,
0.052025746554136276,
0.02764851786196232,
-0.1038934588432312,
0.010874127969145775,
0.01257577445358038,
0.010440350510179996,
0.013064265251159668,
-0.09604154527187347,
-0.0489414818584919,
0.013148768804967403,
-0.02862115390598774,
0.04531305283308029,
0.009714537300169468,
0.03735104575753212,
0.030584225431084633,
0.013268962502479553,
0.04176405072212219,
0.1044875830411911,
0.010680199600756168,
0.08976667374372482,
-0.03662848100066185,
-0.03840511292219162,
0.0740681141614914,
0.033509425818920135,
0.01595386676490307,
-0.0030082869343459606,
-0.02163463458418846,
-0.028999438509345055,
0.10123848170042038,
-0.011226333677768707,
0.013513868674635887,
-0.04844629764556885,
0.007833032868802547,
0.03551517054438591,
-0.04767793416976929,
0.04265899211168289,
-0.06858640909194946,
-0.12030349671840668,
-0.020637819543480873,
-0.06984195113182068,
0.046649619936943054,
-0.0256754532456398,
-0.006885938812047243,
-0.010590698570013046,
-0.06631577759981155,
-0.07460804283618927,
-0.11109654605388641,
0.01486227661371231,
0.07383594661951065,
-4.1677802403717e-33,
0.05906010791659355,
-0.04807325452566147,
0.06368745118379593,
0.010644258931279182,
-0.054506246000528336,
-0.02954006753861904,
0.00201132008805871,
-0.015893084928393364,
-0.0013230415061116219,
-0.005765658803284168,
-0.02237704209983349,
0.039402611553668976,
-0.00452318275347352,
0.0271750558167696,
0.002823233138769865,
-0.0298254806548357,
-0.03453366085886955,
0.1037060022354126,
0.019517401233315468,
0.007144669070839882,
0.056738074868917465,
0.001372316968627274,
-0.11406517773866653,
0.10987082123756409,
-0.05312564969062805,
0.01949903927743435,
-0.058412231504917145,
0.12228284776210785,
0.039773665368556976,
0.0341230072081089,
-0.030392484739422798,
-0.0709325522184372,
-0.0032190997153520584,
-0.017230037599802017,
-0.08851452171802521,
0.10536496341228485,
0.04803493246436119,
0.10003454983234406,
-0.11503792554140091,
0.11432988941669464,
-0.008634362369775772,
-0.05432980880141258,
-0.01150368619710207,
-0.013227714225649834,
0.015828482806682587,
-0.055869124829769135,
-0.04965744912624359,
0.04607304185628891,
0.014958330430090427,
-0.05813727527856827,
-0.04401962831616402,
0.05489367991685867,
-0.01887131668627262,
0.0727163776755333,
-0.00879533588886261,
-0.03145236521959305,
0.03416340425610542,
0.03763951361179352,
0.022213701158761978,
-0.06787744164466858,
-0.017405245453119278,
-0.014557285234332085,
-0.021923089399933815,
0.046806201338768005,
-0.025109414011240005,
-0.013354485854506493,
0.043176453560590744,
0.05383220314979553,
-0.026568138971924782,
-0.037236254662275314,
-0.03765640780329704,
-0.02021799609065056,
0.0530080646276474,
-0.11119306832551956,
0.02131253108382225,
-0.06313779950141907,
0.01790231466293335,
0.0034894368145614862,
0.002237999811768532,
-0.01014024019241333,
-0.03495986759662628,
-0.013161619193851948,
0.07847940921783447,
0.03191668167710304,
-0.0014532931381836534,
0.09722394496202469,
0.023231951519846916,
0.09399141371250153,
0.010250971652567387,
-0.012443211860954762,
-0.03816945105791092,
0.006500510964542627,
0.0062907421961426735,
0.002363866660743952,
-0.03284111246466637,
-4.938407016652491e-8,
-0.0962432399392128,
0.02447541244328022,
0.12010228633880615,
0.04743415489792824,
0.11743688583374023,
-0.006010138429701328,
-0.04054292291402817,
0.007766273338347673,
-0.09243584424257278,
-0.03914282098412514,
0.014995412901043892,
0.06661497056484222,
0.03659551218152046,
-0.10596781224012375,
-0.027070390060544014,
0.07959052175283432,
0.04129427671432495,
-0.08030039817094803,
0.003900727955624461,
0.03623146563768387,
0.01158329751342535,
-0.04156437888741493,
0.023284362629055977,
-0.056719861924648285,
-0.02958068624138832,
-0.05294569209218025,
-0.08256591111421585,
-0.006709577981382608,
0.07934029400348663,
-0.025480931624770164,
0.0013409147504717112,
-0.024335218593478203,
0.06698264181613922,
-0.019099297001957893,
0.05946117267012596,
0.0022824539337307215,
0.052535444498062134,
-0.07203862071037292,
-0.03736615180969238,
-0.02236947789788246,
0.007929517887532711,
0.011316864751279354,
0.012239049188792706,
0.013698981143534184,
0.04435976222157478,
0.009758706204593182,
-0.06488009542226791,
-0.09830685704946518,
0.053745172917842865,
-0.08284410834312439,
0.053928375244140625,
-0.04657107964158058,
-0.049040086567401886,
0.08252845704555511,
0.08812777698040009,
0.030134906992316246,
-0.016019662842154503,
-0.04620538651943207,
0.030391884967684746,
0.08536205440759659,
0.03215031325817108,
-0.029573429375886917,
-0.10091031342744827,
-0.02407761476933956
] |
DICE: End-to-end Deformation Capture of Hand-Face Interactions from a Single Image | https://openreview.net/forum?id=rfrtFwnF62 | [
"Qingxuan Wu",
"Zhiyang Dou",
"Sirui Xu",
"Soshi Shimada",
"Chen Wang",
"Zhengming Yu",
"Yuan Liu",
"Cheng Lin",
"Zeyu Cao",
"Taku Komura",
"Vladislav Golyanik",
"Christian Theobalt",
"Wenping Wang",
"Lingjie Liu"
] | Poster | Reconstructing 3D hand-face interactions with deformations from a single image is a challenging yet crucial task with broad applications in AR, VR, and gaming. The challenges stem from self-occlusions during single-view hand-face interactions, diverse spatial relationships between hands and face, complex deformations, and the ambiguity of the single-view setting. The previous state-of-the-art, Decaf, employs a global fitting optimization guided by contact and deformation estimation networks trained on studio-collected data with 3D annotations. However, Decaf suffers from a time-consuming optimization process and limited generalization capability due to its reliance on 3D annotations of hand-face interaction data. To address these issues, we present DICE, the first end-to-end method for Deformation-aware hand-face Interaction reCovEry from a single image. DICE estimates the poses of hands and faces, contacts, and deformations simultaneously using a Transformer-based architecture. It features disentangling the regression of local deformation fields and global mesh vertex locations into two network branches, enhancing deformation and contact estimation for precise and robust hand-face mesh recovery. To improve generalizability, we propose a weakly-supervised training approach that augments the training set using in-the-wild images without 3D ground-truth annotations, employing the depths of 2D keypoints estimated by off-the-shelf models and adversarial priors of poses for supervision. Our experiments demonstrate that DICE achieves state-of-the-art performance on a standard benchmark and in-the- wild data in terms of accuracy and physical plausibility. Additionally, our method operates at an interactive rate (20 fps) on an Nvidia 4090 GPU, whereas Decaf requires more than 15 seconds for a single image. The code will be available at: https://github.com/Qingxuan-Wu/DICE. | interaction, deformation, end-to-end, mesh recovery | null | 837 | 2406.17988 | [
-0.10279794037342072,
-0.016507212072610855,
0.06176236644387245,
0.008765979669988155,
0.04114704951643944,
-0.021796615794301033,
-0.0035559721291065216,
-0.04148126393556595,
0.059923358261585236,
-0.0287590604275465,
-0.08995641767978668,
-0.08318718522787094,
-0.08210195600986481,
0.11040645837783813,
-0.013471214100718498,
-0.05903001129627228,
0.0005654482520185411,
0.08676782995462418,
-0.0389309860765934,
0.11125405877828598,
0.035991717129945755,
0.01840502955019474,
-0.017232267186045647,
-0.10007604211568832,
0.09936313331127167,
-0.014222727157175541,
0.021824147552251816,
-0.031478170305490494,
-0.004097908269613981,
-0.06399394571781158,
-0.023002764210104942,
0.016563132405281067,
-0.06827326864004135,
0.039391592144966125,
-0.07417185604572296,
0.016793852671980858,
0.004643010441213846,
0.08021845668554306,
-0.08750654757022858,
-0.023226313292980194,
-0.08341573923826218,
0.086049884557724,
0.018325969576835632,
0.015847966074943542,
0.12436864525079727,
0.07846268266439438,
-0.01807798258960247,
0.013304845429956913,
0.08067982643842697,
0.007493470795452595,
0.013128389604389668,
-0.08060409873723984,
0.025861181318759918,
0.0085779232904315,
0.03123452514410019,
0.060922104865312576,
0.0668351948261261,
-0.07856366038322449,
0.01669781282544136,
0.047807395458221436,
0.035025883466005325,
-0.020179472863674164,
-0.01411977969110012,
0.0337589867413044,
-0.05998311564326286,
-0.030740296468138695,
0.1286877989768982,
-0.018741434440016747,
-0.055490873754024506,
-0.008305823430418968,
-0.10005661100149155,
-0.004774418193846941,
-0.027523459866642952,
-0.04560432955622673,
0.04957486689090729,
0.014356354251503944,
-0.029638057574629784,
-0.048422351479530334,
0.03838774561882019,
0.002433513756841421,
0.08039845526218414,
0.03526458516716957,
0.027690188959240913,
0.019527604803442955,
-0.03284694626927376,
-0.012733781710267067,
-0.03678152337670326,
0.06610400974750519,
0.034087080508470535,
0.0393335334956646,
-0.12435507774353027,
0.018411777913570404,
-0.06937870383262634,
-0.005813200026750565,
0.03364703059196472,
0.004143874626606703,
0.007292327471077442,
-0.029288211837410927,
-0.04037465527653694,
0.10966942459344864,
-0.056626688688993454,
-0.07516942918300629,
0.02081144042313099,
0.019216913729906082,
0.04746365174651146,
-0.004233435727655888,
0.008230594918131828,
-0.023608285933732986,
-0.01993652619421482,
-0.006284047849476337,
0.020119626075029373,
-0.04769932106137276,
-0.06291109323501587,
-0.014801202341914177,
0.047079749405384064,
0.0067631835117936134,
0.006626762915402651,
-0.00752468453720212,
0.034378208220005035,
-0.0022954640444368124,
-0.06467776000499725,
-0.01776129938662052,
0.03703157976269722,
-0.015274944715201855,
0.024414710700511932,
0.019243698567152023,
-0.03922172263264656,
-1.6116832028719906e-33,
0.049561623483896255,
0.05426739528775215,
0.04464706778526306,
0.02580968476831913,
-0.013442113064229488,
-0.047242097556591034,
-0.059910714626312256,
0.020094098523259163,
-0.019301576539874077,
0.027849536389112473,
0.022946937009692192,
-0.013666755519807339,
-0.012188997119665146,
0.04909900948405266,
0.022478291764855385,
0.03509300574660301,
-0.02373558282852173,
0.06401365995407104,
-0.0942760482430458,
0.05478755012154579,
-0.02520543895661831,
-0.053418148308992386,
-0.024095335975289345,
0.019749395549297333,
-0.03875616192817688,
0.03957575187087059,
0.04962464049458504,
0.029601596295833588,
0.0673246681690216,
-0.03940504789352417,
-0.06601940840482712,
0.024528518319129944,
0.06408347934484482,
-0.06780616194009781,
0.061187852174043655,
-0.008155496791005135,
0.013684187084436417,
-0.024033434689044952,
-0.08300284296274185,
0.028346188366413116,
0.05805276706814766,
0.07709042727947235,
-0.04812353104352951,
-0.07311419397592545,
-0.08442389965057373,
-0.01933128945529461,
0.0975763350725174,
0.07741300016641617,
-0.07231919467449188,
0.006713071838021278,
0.04185592383146286,
0.05415531247854233,
-0.02337755262851715,
0.014477412216365337,
-0.08725744485855103,
-0.060129933059215546,
0.03144361078739166,
-0.04744973033666611,
0.04234983026981354,
0.01354128960520029,
-0.018941668793559074,
-0.02294781804084778,
-0.04424756392836571,
0.026330869644880295,
-0.031243102625012398,
-0.005182115826755762,
0.007010766305029392,
0.10599861294031143,
0.05509747564792633,
0.011451355181634426,
-0.07357277721166611,
0.009105582721531391,
0.028530366718769073,
-0.07945161312818527,
0.002489152131602168,
0.02231683023273945,
0.03908047452569008,
0.03599729761481285,
-0.006874543149024248,
0.0564495325088501,
-0.004489745944738388,
0.05430254712700844,
-0.010916641913354397,
-0.05931983143091202,
-0.061365775763988495,
-0.049102723598480225,
0.042131926864385605,
-0.06402790546417236,
-0.013978108763694763,
-0.01927795447409153,
-0.06555361300706863,
-0.05539896339178085,
-0.031008677557110786,
-0.05992401763796806,
0.026436438784003258,
2.1910626952152028e-33,
0.0291044432669878,
0.0006080510793253779,
-0.06360820680856705,
-0.0037499782629311085,
0.0183563269674778,
-0.052504558116197586,
0.02962900698184967,
0.044300395995378494,
0.08837135881185532,
-0.07777010649442673,
-0.029291197657585144,
0.003958348650485277,
0.049727488309144974,
0.011974958702921867,
0.05091778188943863,
-0.012091174721717834,
-0.08888900279998779,
-0.017299238592386246,
-0.046677522361278534,
0.037830132991075516,
0.044243745505809784,
0.04501042142510414,
0.019404755905270576,
-0.03315532207489014,
-0.06958580017089844,
0.07643260806798935,
-0.014365093782544136,
-0.006834530737251043,
0.015140668489038944,
0.029755426570773125,
0.048238612711429596,
-0.0020023733377456665,
0.030738884583115578,
0.024259740486741066,
-0.07659994810819626,
0.033551786094903946,
-0.015472532249987125,
-0.0014777613105252385,
0.007085686083883047,
-0.033313993364572525,
0.06340686231851578,
0.0007924322853796184,
-0.11682291328907013,
0.0957469493150711,
0.011572086252272129,
0.007237783167511225,
-0.0733318030834198,
-0.010651160031557083,
0.02477656863629818,
0.06999560445547104,
-0.08271464705467224,
0.06846848875284195,
-0.101486936211586,
-0.05377523601055145,
-0.016428617760539055,
-0.1104317381978035,
-0.00527184596285224,
-0.015129356645047665,
0.0705314576625824,
-0.0017830895958468318,
-0.02892882749438286,
-0.046460896730422974,
-0.076241634786129,
0.05165347456932068,
0.06420194357633591,
0.05679495632648468,
-0.001394021906889975,
-0.028468342497944832,
0.003921062685549259,
0.07540750503540039,
-0.007497300859540701,
0.025277847424149513,
0.02192971110343933,
0.002463668817654252,
0.02929520234465599,
-0.051534827798604965,
0.013895830139517784,
0.021248240023851395,
0.036582283675670624,
-0.022305093705654144,
0.015830297023057938,
-0.13307030498981476,
0.053824376314878464,
0.11364439129829407,
0.055398765951395035,
0.08053120225667953,
-0.015411467291414738,
0.07797618210315704,
-0.012630581855773926,
0.09992628544569016,
-0.10530386865139008,
0.05727655440568924,
0.03274843469262123,
0.015185710974037647,
-0.020626921206712723,
-3.6293755556471297e-8,
-0.09322477877140045,
0.03258313238620758,
0.02530493400990963,
0.0040005045011639595,
-0.06178074702620506,
0.06097731739282608,
0.08062966912984848,
0.049336619675159454,
0.014189575798809528,
0.013022834435105324,
0.01817222870886326,
-0.05065605044364929,
0.026273423805832863,
-0.02913561649620533,
-0.022503597661852837,
0.08417230099439621,
-0.05053972080349922,
0.04592767357826233,
-0.04371452331542969,
0.04627230018377304,
-0.017763221636414528,
0.0018644403899088502,
0.049410130828619,
-0.01345726940780878,
0.07891363650560379,
-0.06565910577774048,
-0.03485565260052681,
0.12988229095935822,
-0.026354797184467316,
-0.03850257024168968,
-0.04000215232372284,
0.00221358728595078,
0.05726460739970207,
-0.04877655208110809,
-0.01689862459897995,
0.04776761308312416,
0.03595651686191559,
-0.01850048266351223,
0.08223431557416916,
-0.038270413875579834,
-0.03640766814351082,
0.0365506075322628,
-0.041482750326395035,
-0.0011103403521701694,
0.015144867822527885,
0.055210743099451065,
0.09669789671897888,
-0.08617175370454788,
0.02990509569644928,
-0.014252020046114922,
-0.03571008890867233,
-0.012415464036166668,
-0.017909187823534012,
0.10022620856761932,
0.048982489854097366,
0.0678296759724617,
-0.007699803914874792,
0.03300470486283302,
0.14313282072544098,
0.07329636067152023,
-0.045217618346214294,
-0.031710319221019745,
-0.06714803725481033,
0.017087671905755997
] |
CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models | https://openreview.net/forum?id=jt1h2dnmng | [
"Zheng Chong",
"Xiao Dong",
"Haoxiang Li",
"shiyue Zhang",
"Wenqing Zhang",
"Hanqing Zhao",
"xujie zhang",
"Dongmei Jiang",
"Xiaodan Liang"
] | Poster | Virtual try-on methods based on diffusion models achieve realistic effects but often require additional encoding modules, a large number of training parameters, and complex preprocessing, which increases the burden on training and inference. In this work, we re-evaluate the necessity of additional modules and analyze how to improve training efficiency and reduce redundant steps in the inference process. Based on these insights, we propose CatVTON, a simple and efficient virtual try-on diffusion model that transfers in-shop or worn garments of arbitrary categories to target individuals by concatenating them along spatial dimensions as inputs of the diffusion model. The efficiency of CatVTON is reflected in three aspects: (1) Lightweight network. CatVTON consists only of a VAE and a simplified denoising UNet, removing redundant image and text encoders as well as cross-attentions, and includes just 899.06M parameters. (2) Parameter-efficient training. Through experimental analysis, we identify self-attention modules as crucial for adapting pre-trained diffusion models to the virtual try-on task, enabling high-quality results with only 49.57M training parameters. (3) Simplified inference. CatVTON eliminates unnecessary preprocessing, such as pose estimation, human parsing, and captioning, requiring only a person image and garment reference to guide the virtual try-on process, reducing over 49% memory usage compared to other diffusion-based methods. Extensive experiments demonstrate that CatVTON achieves superior qualitative and quantitative results compared to baseline methods and demonstrates strong generalization performance in in-the-wild scenarios, despite being trained solely on public datasets with 73K samples. | diffusion models, virtual try-on, parameter-efficient training | A high-quality virtual try-on diffusion model with parameter efficiency and simplified inference. | 833 | 2407.15886 | [
0.02303825505077839,
-0.07980175316333771,
0.02164328470826149,
0.054507117718458176,
0.09529940038919449,
0.0060427202843129635,
0.03621638938784599,
-0.03214511275291443,
-0.007108861580491066,
-0.08967381715774536,
0.02823694236576557,
-0.03884095326066017,
-0.02249763160943985,
0.07287035137414932,
0.0023630911018699408,
-0.03281088545918465,
0.12434317171573639,
0.11619815975427628,
-0.13215628266334534,
-0.0015002336585894227,
-0.011287021450698376,
-0.040508635342121124,
0.04869643226265907,
-0.02904757298529148,
-0.03603338822722435,
-0.00269750552251935,
-0.06769714504480362,
-0.11320720613002777,
-0.03405611589550972,
-0.05659819394350052,
0.02727566473186016,
-0.014657062478363514,
-0.028289733454585075,
0.14170421659946442,
-0.06912990659475327,
-0.008003548718988895,
-0.08733911067247391,
-0.028153790161013603,
0.009351497516036034,
-0.002256625797599554,
-0.05526037514209747,
-0.007329259067773819,
-0.06714370846748352,
0.01621747948229313,
0.09515121579170227,
0.016252148896455765,
0.057075418531894684,
0.0018921366427093744,
0.030611151829361916,
0.019476866349577904,
-0.0062277051620185375,
-0.0678250715136528,
-0.013185424730181694,
0.08025123924016953,
-0.024954212829470634,
0.024376090615987778,
-0.0029800208285450935,
0.018677540123462677,
0.019621843472123146,
-0.01941605657339096,
0.0056716823019087315,
-0.04012023285031319,
-0.03361492604017258,
-0.00048415263881906867,
0.021545076742768288,
-0.05861494690179825,
-0.0009946355130523443,
0.08121530711650848,
-0.017875509336590767,
-0.019462788477540016,
0.05196218565106392,
0.06265980005264282,
-0.1358555406332016,
0.0605049692094326,
0.061498526483774185,
0.02945033647119999,
0.032594408839941025,
-0.04483026638627052,
0.02859819121658802,
-0.034236546605825424,
0.005329117178916931,
-0.016800878569483757,
0.06980340927839279,
-0.015999391674995422,
0.06334582716226578,
0.0271956417709589,
-0.01849118247628212,
0.06929340213537216,
-0.06000758334994316,
-0.059352315962314606,
-0.01960267871618271,
0.022608645260334015,
-0.02283267304301262,
-0.043729718774557114,
0.02490074932575226,
0.005929818842560053,
-0.02362176589667797,
0.007196610793471336,
-0.00412836205214262,
0.0923817828297615,
0.0019277125829830766,
-0.01260343287140131,
-0.006962290499359369,
0.010705133900046349,
0.01376646850258112,
-0.06472717970609665,
0.11238480359315872,
0.07160589843988419,
0.026153508573770523,
-0.030009739100933075,
-0.02672436460852623,
-0.01428636722266674,
0.005628062877804041,
-0.0681057721376419,
0.034251149743795395,
0.00449845427647233,
0.007942251861095428,
0.055816732347011566,
0.06742904335260391,
0.036018695682287216,
0.006107095628976822,
0.012136034667491913,
-0.05136511102318764,
-0.007433888968080282,
-0.001579913543537259,
-0.08038046211004257,
-0.03311474248766899,
3.472032317426851e-33,
-0.049107518047094345,
-0.005768171977251768,
-0.00023586083261761814,
0.016995921730995178,
0.06704118847846985,
-0.036339111626148224,
-0.009859414771199226,
-0.016307735815644264,
-0.019836755469441414,
-0.062297459691762924,
-0.07074456661939621,
-0.034714389592409134,
-0.06521837413311005,
0.13249348104000092,
0.09337037056684494,
-0.025349603965878487,
-0.01678439974784851,
0.017119143158197403,
-0.034029122442007065,
0.029122501611709595,
0.0012354732025414705,
-0.015659447759389877,
0.03840256854891777,
-0.0012624997179955244,
-0.015400481410324574,
-0.03147757425904274,
0.020273428410291672,
-0.03868621215224266,
0.041138775646686554,
0.019348109140992165,
-0.08167240023612976,
0.008906354196369648,
-0.015910115092992783,
0.01387231145054102,
-0.028250573202967644,
0.001167476293630898,
-0.02919550985097885,
0.010863278992474079,
0.025537211447954178,
-0.013250293210148811,
-0.005076774396002293,
0.11663087457418442,
0.00633347500115633,
-0.012615411542356014,
-0.10437948256731033,
-0.01681414619088173,
0.05748942121863365,
0.045807961374521255,
-0.04996321722865105,
0.03230529651045799,
0.05620737746357918,
-0.033768441528081894,
-0.04435129463672638,
-0.09140637516975403,
-0.06442298740148544,
-0.026574207469820976,
0.10945849865674973,
0.008027294650673866,
0.015996070578694344,
0.010066507384181023,
0.05879496410489082,
-0.005454338155686855,
0.052892424166202545,
0.06776215881109238,
0.08828514069318771,
-0.04048535227775574,
0.013023887760937214,
0.05760301649570465,
0.07732471823692322,
-0.004996892996132374,
-0.04692385718226433,
0.029531076550483704,
0.010145682841539383,
-0.12219550460577011,
0.09522203356027603,
-0.0322188064455986,
0.0085151894018054,
-0.10263944417238235,
0.0001761800958774984,
-0.01833982765674591,
-0.01691303960978985,
-0.025388197973370552,
-0.03526841476559639,
-0.07453842461109161,
-0.06337478756904602,
0.00489787058904767,
0.018498487770557404,
-0.13118132948875427,
-0.016833599656820297,
-0.0114451227709651,
-0.011555726639926434,
0.054187558591365814,
-0.038722965866327286,
-0.0029825945384800434,
0.04354292154312134,
-1.9169201475301677e-33,
0.042318474501371384,
0.1042005643248558,
-0.01670764945447445,
0.05755830928683281,
-0.07390951365232468,
0.08582300692796707,
0.014743407256901264,
-0.01542353443801403,
0.002290859119966626,
0.00867992453277111,
-0.0010118962964043021,
-0.06599881500005722,
0.016059167683124542,
-0.003960016183555126,
-0.0021296688355505466,
-0.0012835742672905326,
-0.006377640645951033,
0.031784821301698685,
-0.039559464901685715,
-0.028353679925203323,
-0.021533316001296043,
0.10007298737764359,
-0.062275998294353485,
0.0047570266760885715,
-0.09016413241624832,
0.020781949162483215,
-0.02922239899635315,
0.05444399267435074,
-0.011803919449448586,
-0.07227937132120132,
-0.056010909378528595,
-0.02884540893137455,
-0.012040119618177414,
-0.02068673074245453,
0.016988474875688553,
0.12115181237459183,
0.0679633691906929,
0.036348238587379456,
-0.007930747233331203,
0.020192760974168777,
0.08567294478416443,
-0.026540299877524376,
-0.08609795570373535,
0.021051757037639618,
-0.0950961485505104,
-0.06292557716369629,
-0.09072956442832947,
-0.0015753157204017043,
-0.011159265413880348,
0.054241906851530075,
0.022767310962080956,
0.04232931509613991,
-0.10664569586515427,
0.0025668921880424023,
-0.0422699935734272,
-0.014897040091454983,
-0.02829637937247753,
-0.014258833602070808,
0.029869383201003075,
-0.0361456423997879,
-0.08507391065359116,
-0.016522016376256943,
0.044043220579624176,
-0.06881134957075119,
-0.03797968477010727,
-0.07744258642196655,
-0.07390345633029938,
0.01510412897914648,
0.06584181636571884,
-0.024140235036611557,
0.008640216663479805,
0.0395190492272377,
0.06576510518789291,
-0.0011056730290874839,
-0.0670609101653099,
-0.11963257938623428,
0.06851015985012054,
-0.03253337740898132,
-0.03677923604846001,
-0.019251598045229912,
-0.044500771909952164,
-0.05444192513823509,
-0.011202755384147167,
0.016429077833890915,
0.08790995925664902,
0.10286957025527954,
0.025063002482056618,
0.07248371839523315,
-0.009054298512637615,
0.05793970078229904,
-0.02134030871093273,
0.05162959173321724,
0.12519755959510803,
0.07868024706840515,
-0.08564142882823944,
-4.977386680593554e-8,
-0.08796734362840652,
-0.0011151928920298815,
0.037966035306453705,
0.10454856604337692,
0.040968846529722214,
0.012128169648349285,
0.022221624851226807,
0.12414281815290451,
-0.0455644428730011,
0.034402910619974136,
0.0487193800508976,
0.02857612632215023,
-0.06205851212143898,
-0.011028161272406578,
0.03661365807056427,
0.08548744022846222,
0.049802590161561966,
0.00016294053057208657,
-0.000568322604522109,
0.00492524541914463,
0.0414591059088707,
0.05022208020091057,
-0.010212209075689316,
0.004976609721779823,
0.042946238070726395,
-0.004792380146682262,
-0.04796936362981796,
0.12870638072490692,
0.0019416826544329524,
-0.020265450701117516,
0.05950237065553665,
0.022776074707508087,
-0.051566239446401596,
-0.057544831186532974,
0.0020196447148919106,
0.04666915535926819,
0.06796971708536148,
-0.036237169057130814,
-0.027801403775811195,
-0.024512335658073425,
0.02706248313188553,
-0.0271578598767519,
-0.01806839555501938,
0.025872088968753815,
-0.0020434968173503876,
-0.004998542834073305,
0.032981012016534805,
-0.09669915586709976,
0.038617417216300964,
0.028095275163650513,
0.03684372827410698,
-0.0449521467089653,
-0.03376321122050285,
0.07950211316347122,
-0.0077371420338749886,
-0.019660191610455513,
0.05618328973650932,
0.016704754903912544,
0.024373451247811317,
0.10158427804708481,
0.07222042977809906,
0.0176963210105896,
-0.027475865557789803,
-0.02266312763094902
] |
Hessian-Free Online Certified Unlearning | https://openreview.net/forum?id=C3TrHWanh5 | [
"Xinbao Qiao",
"Meng Zhang",
"Ming Tang",
"Ermin Wei"
] | Poster | Machine unlearning strives to uphold the data owners' right to be forgotten by enabling models to selectively forget specific data.
Recent advances suggest pre-computing and storing statistics extracted from second-order information and implementing unlearning through Newton-style updates.
However, the Hessian matrix operations are extremely costly and previous works conduct unlearning for empirical risk minimizer with the convexity assumption, precluding their applicability to high-dimensional over-parameterized models and the nonconvergence condition.
In this paper, we propose an efficient Hessian-free unlearning approach.
The key idea is to maintain a statistical vector for each training data, computed through affine stochastic recursion of the difference between the retrained and learned models.
We prove that our proposed method outperforms the state-of-the-art methods in terms of the unlearning and generalization guarantees, the deletion capacity, and the time/storage complexity, under the same regularity conditions.
Through the strategy of recollecting statistics for removing data, we develop an online unlearning algorithm that achieves near-instantaneous data removal, as it requires only vector addition.
Experiments demonstrate that our proposed scheme surpasses existing results by orders of magnitude in terms of time/storage costs with millisecond-level unlearning execution, while also enhancing test accuracy. | machine unlearning; certified data removal; privacy | null | 830 | 2404.01712 | [
-0.0896296426653862,
-0.11332528293132782,
-0.04310821369290352,
0.09296024590730667,
0.10066905617713928,
-0.007555332500487566,
-0.020536309108138084,
-0.08131527155637741,
0.017129089683294296,
0.047649774700403214,
0.055323150008916855,
0.057649336755275726,
0.0267120860517025,
-0.01914566569030285,
-0.05646401271224022,
-0.009387630969285965,
-0.014430973678827286,
0.06290122121572495,
-0.0853152871131897,
-0.0370018295943737,
-0.040336765348911285,
0.07978442311286926,
-0.04599611461162567,
0.07954690605401993,
0.07775343209505081,
-0.05958954244852066,
0.08002243936061859,
0.01805366948246956,
-0.06842724978923798,
0.001983178546652198,
0.035961396992206573,
-0.014456436969339848,
0.041335005313158035,
-0.05047951266169548,
-0.040589217096567154,
0.051715876907110214,
-0.0025031499098986387,
0.03907580301165581,
-0.02558119222521782,
0.014789726585149765,
-0.017517127096652985,
0.0005985936732031405,
-0.04894129931926727,
0.04996205121278763,
0.05739531293511391,
0.01923184283077717,
-0.0231730155646801,
-0.09008916467428207,
0.04574798792600632,
0.03286822885274887,
0.010012886486947536,
-0.00019612211326602846,
-0.09618779271841049,
0.06301012635231018,
-0.015395062044262886,
-0.03342755138874054,
0.022838955745100975,
-0.042377665638923645,
-0.08365947753190994,
0.030530359596014023,
-0.004547005519270897,
-0.11489012837409973,
-0.03973725065588951,
-0.021376347169280052,
-0.047401294112205505,
-0.006957935635000467,
0.017023954540491104,
0.022596389055252075,
0.03250845521688461,
0.07995882630348206,
-0.06919194012880325,
0.08714176714420319,
-0.042295489460229874,
0.021466359496116638,
0.008001197129487991,
0.03299522399902344,
0.08459259569644928,
-0.010836188681423664,
0.050462204962968826,
-0.030267801135778427,
-0.008203434757888317,
0.048997532576322556,
0.07661177963018417,
0.011634456925094128,
0.06147124990820885,
0.004005004186183214,
0.04883822053670883,
0.03302764147520065,
0.12688903510570526,
-0.047972600907087326,
0.020751584321260452,
0.0421685054898262,
-0.033442363142967224,
4.167480085470743e-7,
-0.01883777044713497,
-0.035206232219934464,
0.0477604903280735,
-0.01711982674896717,
-0.024774964898824692,
0.11141079664230347,
-0.014742436818778515,
0.03959023579955101,
-0.014887562952935696,
0.016819411888718605,
-0.03780896216630936,
-0.041555773466825485,
0.011722346767783165,
0.07625481486320496,
-0.015521283261477947,
-0.06992717832326889,
0.04790911450982094,
0.05367078259587288,
-0.036149073392152786,
0.019804568961262703,
-0.006226500496268272,
0.0769035592675209,
-0.007362482603639364,
-0.042543455958366394,
-0.020065803080797195,
0.1317809522151947,
-0.07010328024625778,
-0.016628460958600044,
0.0527690052986145,
0.06919582188129425,
-0.03586669638752937,
-0.0036675736773759127,
-0.08968646079301834,
3.03414726376311e-33,
-0.0402015745639801,
0.013341236859560013,
0.043932415544986725,
-0.05218752846121788,
0.0452948734164238,
0.038885995745658875,
0.001939314417541027,
0.03676065057516098,
0.02379295602440834,
0.006559559144079685,
-0.006266423035413027,
0.011961225420236588,
-0.06456732749938965,
0.009878342971205711,
0.053398266434669495,
0.003003875957801938,
-0.04655253514647484,
0.08193869143724442,
0.013012954033911228,
-0.0953645259141922,
0.10754940658807755,
-0.05852261558175087,
-0.03007327951490879,
-0.04896199330687523,
-0.03206263482570648,
-0.04262382909655571,
0.0628814771771431,
-0.004946763627231121,
-0.0006072797696106136,
0.023663751780986786,
-0.049599539488554,
0.0047994046472013,
0.011928624473512173,
0.03162725642323494,
0.035664256662130356,
0.006567835342139006,
-0.0008338482002727687,
-0.04925425350666046,
0.008960386738181114,
-0.00880932155996561,
0.003143968991935253,
0.037156686186790466,
-0.013547435402870178,
-0.07003103196620941,
-0.11701949685811996,
-0.1298941671848297,
-0.007094138767570257,
0.02432139404118061,
-0.001742839114740491,
-0.08183148503303528,
0.043428368866443634,
0.024942487478256226,
0.02168560028076172,
-0.04422302916646004,
-0.15442544221878052,
0.028818385675549507,
0.057626646012067795,
-0.02783079631626606,
-0.00925138033926487,
0.021766917780041695,
-0.02015163004398346,
0.07623264193534851,
0.005972404032945633,
0.029645215719938278,
-0.016887756064534187,
0.03538073971867561,
0.014695710502564907,
0.085858054459095,
0.012234126217663288,
-0.049583952873945236,
-0.013972030021250248,
0.0006785001023672521,
-0.025820881128311157,
-0.09534996002912521,
0.03300771862268448,
0.00817430205643177,
0.07092884927988052,
-0.02595556527376175,
0.051209088414907455,
-0.0062082745134830475,
0.03449394926428795,
0.04449476674199104,
-0.010406726971268654,
-0.049819353967905045,
-0.028488734737038612,
-0.10298880189657211,
0.02683723159134388,
-0.027378136292099953,
-0.029702557250857353,
0.009562725201249123,
-0.06811684370040894,
0.003606193233281374,
-0.05105859413743019,
-0.058508384972810745,
0.05040007829666138,
-2.0447722395843768e-33,
-0.06428711116313934,
-0.006200395990163088,
-0.02908952906727791,
0.04796183481812477,
0.008431610651314259,
-0.036379121243953705,
-0.054101742804050446,
0.026036662980914116,
-0.009180476889014244,
-0.13791094720363617,
0.00644436851143837,
-0.05463099107146263,
-0.015546517446637154,
-0.036361660808324814,
0.047747790813446045,
0.043609071522951126,
-0.10406195372343063,
0.03475222736597061,
-0.08602846413850784,
0.005762993358075619,
0.01382638793438673,
-0.012567272409796715,
-0.05506713688373566,
-0.026722470298409462,
-0.07534609735012054,
-0.0008475309587083757,
-0.011885551735758781,
0.14807134866714478,
-0.0006698538782075047,
-0.0018956154817715287,
-0.11025010049343109,
-0.013968819752335548,
-0.025366824120283127,
0.004039959516376257,
-0.02433072216808796,
0.03163812682032585,
0.0019404897466301918,
0.05483455955982208,
-0.007205347530543804,
0.10159130394458771,
0.05195584148168564,
0.06593945622444153,
-0.10828626155853271,
-0.03963115066289902,
0.04361090064048767,
-0.009003685787320137,
-0.07312742620706558,
0.019248003140091896,
0.11555377393960953,
-0.003792101051658392,
-0.042202990502119064,
0.024471454322338104,
-0.09962614625692368,
0.04251018539071083,
-0.03272486850619316,
0.005145909730345011,
-0.0027658522594720125,
-0.03561077639460564,
0.003862338839098811,
0.0037638258654624224,
-0.0381503589451313,
0.021283524110913277,
-0.09162801504135132,
0.052376653999090195,
0.02986259013414383,
-0.09575891494750977,
0.01606328785419464,
0.037422072142362595,
0.00621391274034977,
0.03954910486936569,
-0.04385615885257721,
0.062586709856987,
0.005681131966412067,
-0.0487976036965847,
-0.018492288887500763,
-0.047402944415807724,
0.03516671806573868,
-0.0031224344857037067,
-0.05326678231358528,
-0.002228270284831524,
0.08189269155263901,
0.004355915356427431,
0.058141469955444336,
0.017767038196325302,
0.05477225407958031,
0.08796098828315735,
0.08441286534070969,
-0.005287550389766693,
0.029518187046051025,
-0.029326962307095528,
-0.026541905477643013,
0.022993924096226692,
0.04633341729640961,
0.048567380756139755,
-0.04320386052131653,
-3.814815485725376e-8,
-0.0061197299510240555,
-0.01429255772382021,
0.07544564455747604,
-0.00048485121806152165,
0.07643463462591171,
-0.03356529027223587,
0.0019204698037356138,
0.1294499933719635,
-0.027734430506825447,
-0.034644581377506256,
0.05435783043503761,
-0.03327739238739014,
-0.06931570917367935,
-0.05627264082431793,
-0.08407901972532272,
0.033012088388204575,
0.0016558659262955189,
0.04828554019331932,
0.01022990420460701,
0.003463412169367075,
-0.02890491858124733,
0.011534912511706352,
0.03187163546681404,
0.0344562828540802,
0.09785018861293793,
-0.059884317219257355,
0.048575494438409805,
0.07854673266410828,
0.12775497138500214,
0.03417811915278435,
-0.0380416065454483,
-0.052244413644075394,
0.0741252452135086,
-0.00797736644744873,
0.02238970622420311,
0.07531420886516571,
0.026924628764390945,
0.019879864528775215,
-0.08409096300601959,
-0.05128839239478111,
-0.015409647487103939,
0.11398142576217651,
-0.03345859795808792,
0.028256133198738098,
0.06326683610677719,
-0.024557577446103096,
-0.03476767987012863,
-0.011085831560194492,
0.09156883507966995,
0.03624414652585983,
0.019582198932766914,
-0.0066989632323384285,
0.03658169135451317,
0.09549041092395782,
0.015546931885182858,
0.05345133692026138,
-0.02506561018526554,
0.017805716022849083,
-0.019616784527897835,
0.02938205376267433,
0.01239681988954544,
-0.032357972115278244,
-0.055705998092889786,
-0.0135116558521986
] |
Video-STaR: Self-Training Enables Video Instruction Tuning with Any Supervision | https://openreview.net/forum?id=JYV2hrtFSv | [
"Orr Zohar",
"Xiaohan Wang",
"Yonatan Bitton",
"Idan Szpektor",
"Serena Yeung-Levy"
] | Poster | The performance and reasoning capabilities of Large Multi-modal Models (LMMs) is dependent on the size and quality of their training datasets. However, collecting datasets that support chain-of-thought instruction tuning is highly challenging. Existing video instruction tuning datasets are often derived by prompting large language models with video captions to generate question-answer pairs, which makes them predominantly descriptive rather than reasoning-focused.
Meanwhile, many labeled video datasets with diverse labels and supervision exist -- however, we find that their integration into LMMs is non-trivial.
Herein, we present $\underline{\text{Video}}$ $\underline{\text{S}}\text{elf}$-$\underline{\text{T}}\text{raining}$ $\text{with}$ $\underline{\text{a}}\text{ugmented}$ $\underline{\text{R}}\text{easoning}$ (Video-STaR), the first self-training approach for video instruction tuning.
Video-STaR allows the utilization of *any* labeled video dataset for video instruction tuning.
In Video-STaR, an LMM cycles between instruction generation and finetuning, which we show (I) improves general video understanding and (II) adapts LMMs to novel downstream tasks with existing supervision.
During instruction generation, an LMM is prompted to propose an answer. The answers are then filtered only to those that contain the original video labels, and the LMM is then re-trained on the generated dataset.
By training exclusively on generated answers containing the correct video labels, Video-STaR leverages these existing labels as weak supervision for video instruction tuning.
Our results demonstrate that Video-STaR-augmented LMMs achieve notable improvements in (I) general Video QA, where TempCompass performance improved by 6.1%, *and* (II) downstream tasks, with a 9.9% increase in Kinetics700-QA accuracy and a 4.0% improvement in action quality assessment on FineDiving, while also exhibiting better interpretability. | Video Understanding, Visual Instruction Tuning, Self-Training, Chain-of-thought reasoning | Introducing the first self-training method for video-LMMs | 826 | null | [
-0.0009324653656221926,
-0.07180391252040863,
0.010551643557846546,
-0.03642794117331505,
0.0856889858841896,
0.020857078954577446,
0.06339895725250244,
-0.0095550287514925,
0.03658250719308853,
-0.0690297856926918,
-0.007886629551649094,
-0.05731537193059921,
-0.0020347388926893473,
0.04526243358850479,
0.010729797184467316,
-0.013115099631249905,
0.04526253417134285,
-0.009603708051145077,
-0.129499614238739,
-0.09990079700946808,
0.0612633042037487,
-0.008720126934349537,
0.05180182307958603,
0.02681465819478035,
0.010959207080304623,
0.0040248678997159,
-0.028981192037463188,
-0.011196937412023544,
0.07145719230175018,
-0.10126309096813202,
0.03220731392502785,
0.04879133775830269,
0.0009009605273604393,
0.043174270540475845,
-0.05291656777262688,
0.029351336881518364,
-0.009060405194759369,
-0.012007505632936954,
-0.053208090364933014,
-0.04328886792063713,
-0.04070811718702316,
0.041408319026231766,
0.06445633620023727,
-0.005120045971125364,
0.02331598661839962,
-0.0026219915598630905,
-0.023515382781624794,
-0.047019217163324356,
-0.024935906752943993,
0.01659788005053997,
-0.14077258110046387,
-0.02502315677702427,
0.013591726310551167,
-0.02607576921582222,
-0.0008351357537321746,
0.0117906853556633,
0.08232347667217255,
0.0011260801693424582,
0.030473431572318077,
-0.046506334096193314,
-0.050107602030038834,
-0.07948505133390427,
-0.08554792404174805,
0.02702554687857628,
-0.0007454918813891709,
-0.0425424799323082,
0.03208652883768082,
-0.01664649322628975,
-0.0322774201631546,
-0.03310806676745415,
-0.10184727609157562,
0.040310028940439224,
-0.01952606439590454,
0.07433651387691498,
-0.02679978311061859,
0.04457911476492882,
0.04435548558831215,
-0.05438268184661865,
0.06267131119966507,
-0.047946490347385406,
0.04531000182032585,
-0.0458064042031765,
0.06737085431814194,
0.009873107075691223,
0.07895661890506744,
-0.013228638097643852,
0.04323892295360565,
0.020830661058425903,
-0.021016191691160202,
0.021894415840506554,
-0.08362407982349396,
-0.05390066280961037,
0.011774040758609772,
0.00458143837749958,
0.02488590031862259,
0.0853864997625351,
0.03673287481069565,
-0.13714845478534698,
-0.01338772289454937,
0.10607220977544785,
0.06008854880928993,
0.06157654523849487,
0.02232762984931469,
0.02619125135242939,
0.048632193356752396,
-0.015120195224881172,
0.06512049585580826,
0.06560453772544861,
0.05524329096078873,
-0.09581183642148972,
0.026261137798428535,
0.0294350478798151,
-0.009330777451395988,
-0.015800466760993004,
0.028287392109632492,
0.004829333629459143,
-0.014970226213335991,
-0.01814119704067707,
-0.0028158463537693024,
0.00010027434473158792,
-0.01858779415488243,
0.01798159070312977,
0.017885081470012665,
0.049942631274461746,
0.08351099491119385,
-0.08351372927427292,
-0.04981338977813721,
2.4414119974603967e-34,
0.1334175020456314,
0.03518514335155487,
0.049806930124759674,
-0.006930925417691469,
0.04011540487408638,
-0.02740580029785633,
-0.02174094133079052,
0.04831349477171898,
-0.03497779369354248,
0.03088129498064518,
0.009700220078229904,
-0.03519391268491745,
-0.0782189890742302,
0.08427093923091888,
0.08921416103839874,
-0.008836145512759686,
-0.05172714963555336,
-0.010152786038815975,
-0.07017415761947632,
0.03530057519674301,
0.02729303203523159,
0.06233639270067215,
0.058018479496240616,
0.013493014499545097,
0.031065819784998894,
-0.012682097032666206,
0.07516809552907944,
-0.06580575555562973,
-0.00041853904258459806,
0.005974244326353073,
-0.06447938084602356,
-0.04813845455646515,
-0.014305450953543186,
0.07497812062501907,
0.052660129964351654,
-0.031142523512244225,
-0.05331948399543762,
-0.055518776178359985,
0.03675241395831108,
-0.03606314957141876,
0.06314165890216827,
0.04114489629864693,
-0.012340269051492214,
0.009424116462469101,
-0.058675188571214676,
-0.021591970697045326,
0.05361773818731308,
0.0548224113881588,
-0.10602113604545593,
-0.04418322443962097,
0.03317301347851753,
0.09895946085453033,
-0.0286561232060194,
-0.1688179075717926,
0.01830635592341423,
0.04372837021946907,
0.07045301049947739,
0.06089915707707405,
0.02920638397336006,
-0.02263147011399269,
-0.019410088658332825,
0.021261965855956078,
0.022495867684483528,
0.08230678737163544,
0.03588755056262016,
0.04392841085791588,
0.01444255094975233,
0.06617238372564316,
0.1250716894865036,
-0.026291321963071823,
-0.09387582540512085,
-0.018094642087817192,
0.03557677939534187,
-0.07352165877819061,
0.02423086389899254,
0.007256306242197752,
0.01912444829940796,
-0.14649474620819092,
0.058397043496370316,
0.05132310464978218,
0.03799526393413544,
0.01631825976073742,
0.031173057854175568,
-0.056194864213466644,
-0.036347612738609314,
-0.12235429883003235,
0.024533484131097794,
-0.10052543133497238,
0.020322931930422783,
-0.02757398411631584,
-0.06310800462961197,
-0.011167183518409729,
-0.0006164496298879385,
-0.05569282919168472,
-0.026670226827263832,
-2.118237376350656e-34,
0.01627185381948948,
0.004803686868399382,
-0.07040176540613174,
0.06623557955026627,
0.018205732107162476,
-0.03957460820674896,
0.06669662892818451,
0.06302386522293091,
0.045195914804935455,
-0.05612293258309364,
-0.031087195500731468,
-0.036763448268175125,
-0.06711746752262115,
-0.06576631218194962,
0.006424240302294493,
-0.039426665753126144,
-0.025830069556832314,
0.03557673841714859,
-0.005659216083586216,
-0.0005399312940426171,
-0.007668638601899147,
0.08289466053247452,
0.002921859035268426,
0.01743091642856598,
-0.02173798158764839,
0.02663242071866989,
-0.04013519361615181,
0.09062238782644272,
-0.010430554859340191,
0.013720560818910599,
0.03174548223614693,
-0.05788305774331093,
0.02172306552529335,
-0.0009596057934686542,
-0.10323149710893631,
0.03865162655711174,
0.0745832696557045,
-0.02517583966255188,
-0.06558868288993835,
0.08848483115434647,
0.031792134046554565,
-0.017519304528832436,
-0.06518014520406723,
-0.026573553681373596,
-0.014925345778465271,
-0.0675172358751297,
-0.12594552338123322,
-0.004623596556484699,
-0.05466277524828911,
0.003997340332716703,
-0.05681269243359566,
0.0001653835061006248,
-0.045409779995679855,
-0.002237715758383274,
-0.0019055630546063185,
-0.12441552430391312,
-0.027434980496764183,
0.023054391145706177,
0.05260148644447327,
0.04639505222439766,
-0.061196889728307724,
-0.03105921298265457,
0.028822867199778557,
-0.04230446368455887,
0.0002818694047164172,
-0.06375115364789963,
-0.03294266760349274,
-0.05577336996793747,
0.02031109295785427,
0.014333156868815422,
0.1200001910328865,
0.002959375036880374,
0.020722582936286926,
0.027127116918563843,
0.009510899893939495,
0.03762257099151611,
0.00279622757807374,
-0.03425624594092369,
-0.016614407300949097,
-0.10492780059576035,
-0.05345236882567406,
-0.0373748280107975,
0.05440395325422287,
0.1024998128414154,
0.05225295200943947,
0.053031619638204575,
0.08021484315395355,
0.10290612280368805,
-0.006449517328292131,
0.0596311092376709,
0.0010169798042625189,
0.0400744266808033,
0.04624079540371895,
0.0873628705739975,
-0.02195323444902897,
-4.586165402997722e-8,
-0.08037396520376205,
0.06009819358587265,
-0.0010763561585918069,
0.07399968057870865,
0.051328085362911224,
-0.012899908237159252,
0.004179035313427448,
0.025857754051685333,
0.0041435640305280685,
-0.057349976152181625,
0.06066327914595604,
0.01705585978925228,
-0.0465070866048336,
-0.0017024945700541139,
0.026174446567893028,
0.04295327141880989,
0.023953063413500786,
0.01390504278242588,
-0.028111154213547707,
0.07296424359083176,
0.04152141511440277,
0.0009357779053971171,
0.01142131444066763,
-0.0048576355911791325,
-0.025534190237522125,
-0.002152357017621398,
-0.04072994366288185,
0.09729200601577759,
0.042020656168460846,
0.060046713799238205,
0.004173362627625465,
0.01861453428864479,
0.0009559648460708559,
-0.11015405505895615,
0.031740330159664154,
0.005845454055815935,
0.013016492128372192,
-0.05012989416718483,
0.007913405075669289,
-0.027011344209313393,
0.030695384368300438,
0.04156394675374031,
-0.017178894951939583,
0.029183054342865944,
0.05789434164762497,
0.05968953296542168,
-0.05090916156768799,
-0.10244790464639664,
-0.007824458181858063,
-0.0007243475411087275,
-0.028231415897607803,
0.005840438883751631,
-0.08254197984933853,
0.08204327523708344,
0.09910831600427628,
-0.0029560045804828405,
0.00945196207612753,
-0.0009071780950762331,
-0.02694154530763626,
0.04878601059317589,
0.0067987218499183655,
0.015130676329135895,
0.0015344716375693679,
-0.03524433821439743
] |
Field-DiT: Diffusion Transformer on Unified Video, 3D, and Game Field Generation | https://openreview.net/forum?id=w6YS9A78fq | [
"Kangfu Mei",
"Mo Zhou",
"Vishal M. Patel"
] | Poster | The probabilistic field models the distribution of continuous functions defined over metric spaces. While these models hold great potential for unifying data generation across various modalities, including images, videos, and 3D geometry, they still struggle with long-context generation beyond simple examples. This limitation can be attributed to their MLP architecture, which lacks sufficient inductive bias to capture global structures through uniform sampling.
To address this, we propose a new and simple model that incorporates a view-wise sampling algorithm to focus on local structure learning, along with autoregressive generation to preserve global geometry. It adapts cross-modality conditions, such as text prompts for text-to-video generation, camera poses for 3D view generation, and control actions for game generation.
Experimental results across various modalities demonstrate the effectiveness of our model, with its 675M parameter size, and highlight its potential as a foundational framework for scalable, modality-unified visual content generation. Our project page can be found at https://kfmei.com/Field-DiT/. | Generative Model, Diffusion Model, Diffusion Probabilistic Fields, World Model | null | 824 | null | [
-0.0638175830245018,
-0.051762428134679794,
0.04061617702245712,
-0.07051675766706467,
0.10922657698392868,
-0.019706077873706818,
-0.03138488903641701,
-0.028820352628827095,
0.0868849903345108,
-0.018293000757694244,
-0.018380189314484596,
-0.05846460908651352,
-0.019929274916648865,
0.06364043802022934,
0.01861054264008999,
0.0013830940006300807,
0.04902397096157074,
0.09123852849006653,
-0.11989529430866241,
0.013769692741334438,
-0.041611894965171814,
-0.02454494684934616,
0.03428869694471359,
-0.025594288483262062,
0.03892169147729874,
0.05576628819108009,
0.09302143007516861,
-0.026194266974925995,
0.03997707739472389,
-0.05887533351778984,
0.07826244086027145,
0.01377204991877079,
-0.046831853687763214,
0.010168928653001785,
-0.05819668993353844,
0.09309178590774536,
-0.06194661557674408,
0.015347715467214584,
-0.025915557518601418,
-0.03456650674343109,
-0.01921025663614273,
0.05973288044333458,
0.023215578868985176,
0.05712498351931572,
0.1220187246799469,
0.02083173580467701,
-0.05698421970009804,
0.0004923304077237844,
0.0618714839220047,
0.04583599045872688,
-0.05078753083944321,
-0.03403811156749725,
-0.019551701843738556,
0.013764803297817707,
-0.01013067364692688,
0.09776937961578369,
0.06618861854076385,
-0.024375954642891884,
0.03844806179404259,
-0.012531602755188942,
-0.028627555817365646,
-0.025564203038811684,
-0.017518699169158936,
-0.011336429044604301,
-0.04217730090022087,
-0.03021475113928318,
0.04634057357907295,
0.02535010501742363,
-0.03966964781284332,
-0.0465632863342762,
-0.04489562287926674,
0.027319910004734993,
-0.03647991642355919,
0.0100615955889225,
-0.005669179372489452,
0.05345470458269119,
-0.050703827291727066,
-0.08945376425981522,
-0.013209020718932152,
-0.06158557906746864,
0.12305506318807602,
0.0103991087526083,
0.08065181225538254,
0.016995863988995552,
-0.003536789445206523,
0.00015836706734262407,
-0.04788573458790779,
0.07599752396345139,
-0.029161347076296806,
0.014715062454342842,
-0.08692079782485962,
-0.009943473152816296,
-0.051186155527830124,
-0.04164322465658188,
-0.06627418845891953,
-0.015825165435671806,
-0.015970593318343163,
-0.12323073297739029,
0.000009680891707830597,
0.08429108560085297,
0.015207735821604729,
-0.006438825745135546,
0.0853082537651062,
0.03988669440150261,
0.02908097393810749,
-0.07489214837551117,
0.042841482907533646,
0.052184704691171646,
-0.06790733337402344,
-0.010057244449853897,
0.05014674738049507,
0.029473716393113136,
-0.0481363944709301,
0.0314454659819603,
0.056342098861932755,
-0.029169589281082153,
0.004281498026102781,
-0.034736186265945435,
0.01968979462981224,
0.01146439928561449,
0.05737646669149399,
0.01122184470295906,
-0.08034753054380417,
0.04626567289233208,
-0.021730618551373482,
-0.03535091504454613,
0.05373947322368622,
6.637569819404649e-34,
0.04983644187450409,
-0.006286663468927145,
0.07882565259933472,
0.12760964035987854,
0.0980537086725235,
0.05524745211005211,
0.052913665771484375,
0.0190906822681427,
0.002441385295242071,
-0.045231737196445465,
-0.010710585862398148,
-0.03133928403258324,
-0.0497429259121418,
0.051139820367097855,
0.11340834945440292,
-0.06531526893377304,
-0.054841019213199615,
0.015235110186040401,
0.0050786216743290424,
0.12558798491954803,
0.027889344841241837,
-0.012195856310427189,
0.0158923901617527,
-0.021198607981204987,
-0.05936732515692711,
0.07130151242017746,
-0.004676833748817444,
0.02168401889503002,
-0.03845716267824173,
-0.01716862991452217,
-0.112308070063591,
-0.028242358937859535,
-0.001933715189807117,
-0.014355694875121117,
0.04924730211496353,
0.019168773666024208,
-0.061580073088407516,
-0.0683906078338623,
-0.07816849648952484,
-0.030536144971847534,
0.022505488246679306,
0.062215566635131836,
0.022281548008322716,
-0.05073310807347298,
-0.10891637206077576,
0.008800058625638485,
0.0742868036031723,
0.0313466414809227,
-0.09731332957744598,
-0.04033495858311653,
0.11459247767925262,
-0.008441061712801456,
-0.031802549958229065,
-0.09235652536153793,
0.01818137615919113,
-0.017371002584695816,
-0.00021560669119935483,
-0.08035232871770859,
0.036194946616888046,
0.03243211656808853,
0.028960712254047394,
0.022772768512368202,
-0.05090462416410446,
0.04112536460161209,
0.042721938341856,
0.011713054031133652,
0.0006022278685122728,
0.013122022151947021,
0.07930462807416916,
0.0185810886323452,
0.03146837279200554,
0.014170360751450062,
-0.08112543821334839,
-0.042334847152233124,
-0.00000849991283757845,
0.003283635713160038,
-0.004815137013792992,
-0.011907178908586502,
0.0058877673000097275,
0.09853481501340866,
-0.08970528095960617,
-0.033693984150886536,
0.026931431144475937,
-0.05244633927941322,
-0.008153792470693588,
-0.033802520483732224,
0.04984106123447418,
-0.09710531681776047,
-0.054952509701251984,
-0.12157758325338364,
-0.0541134811937809,
-0.006238711066544056,
-0.024551503360271454,
0.014263903722167015,
0.02465018257498741,
-1.5385838928046566e-33,
0.050374835729599,
-0.02928907424211502,
-0.009625058621168137,
0.06304977089166641,
-0.0035361573100090027,
0.01054616179317236,
0.01459812093526125,
0.060026176273822784,
0.05277363955974579,
-0.04289089888334274,
-0.10567445307970047,
0.02192586101591587,
0.006605020258575678,
-0.054428160190582275,
0.02253592014312744,
0.009077916853129864,
0.033669039607048035,
-0.04717480391263962,
-0.07756742089986801,
0.028387058526277542,
0.04010726138949394,
0.039672575891017914,
-0.03494216874241829,
-0.029476817697286606,
-0.06809936463832855,
0.04580792412161827,
-0.04714084044098854,
0.0019990073051303625,
0.04357300326228142,
0.02296287938952446,
0.007763199042528868,
0.002153529552742839,
0.06843646615743637,
0.015805684030056,
-0.12201796472072601,
0.04312646761536598,
0.04869775474071503,
0.05136636644601822,
-0.005034321453422308,
0.07493020594120026,
-0.000264334084931761,
-0.033877477049827576,
-0.0721696987748146,
0.06877434998750687,
-0.05112650990486145,
0.0350436270236969,
-0.025248045101761818,
0.007823796011507511,
-0.0012414399534463882,
0.05012065917253494,
-0.08695478737354279,
0.06108669564127922,
-0.08458274602890015,
-0.02052806317806244,
-0.04518617317080498,
-0.09291750937700272,
-0.11028321832418442,
-0.03581531345844269,
0.1282826066017151,
-0.008099670521914959,
-0.02190876193344593,
-0.05567885562777519,
-0.06061450019478798,
-0.025410570204257965,
-0.0016918479232117534,
-0.0005817469791509211,
-0.0783519446849823,
-0.029638120904564857,
-0.01861736923456192,
0.07258040457963943,
0.00006787291204091161,
-0.007817941717803478,
0.008534294553101063,
-0.0043208422139286995,
-0.048263829201459885,
-0.03459448367357254,
0.031405914574861526,
0.022449029609560966,
0.12158876657485962,
-0.058011218905448914,
0.07673175632953644,
-0.034849029034376144,
0.016195593401789665,
0.08164692670106888,
0.05476270988583565,
-0.0026871354784816504,
0.006922000087797642,
0.009208610281348228,
0.008602320216596127,
0.04688829928636551,
0.004638992715626955,
0.04857555031776428,
0.009909729473292828,
0.06514260172843933,
0.015475734136998653,
-4.454419766375395e-8,
-0.1070241928100586,
0.05050219967961311,
-0.03911019116640091,
0.022515619173645973,
-0.04908066615462303,
-0.04886860400438309,
0.04276975616812706,
-0.0015677111223340034,
0.05193040519952774,
-0.05666884779930115,
0.009257378056645393,
-0.060923486948013306,
-0.012958621606230736,
-0.01950613223016262,
-0.02735763229429722,
0.07580754905939102,
0.027488116174936295,
0.009456630796194077,
0.013734659180045128,
0.10205334424972534,
-0.012494639493525028,
0.055530063807964325,
0.007915299385786057,
-0.05281631648540497,
0.0208100788295269,
-0.020072871819138527,
-0.09485894441604614,
-0.02099970541894436,
0.11152731627225876,
-0.03400168567895889,
0.020328881219029427,
0.0003666682750917971,
0.013276876881718636,
-0.004338984843343496,
-0.017358237877488136,
0.019537199288606644,
-0.002423142548650503,
-0.0641893744468689,
-0.00907909870147705,
-0.0705791637301445,
-0.01516783144325018,
0.061866603791713715,
-0.045383792370557785,
0.005236357916146517,
0.018435291945934296,
0.04131375625729561,
0.09101435542106628,
-0.06589307636022568,
-0.009807062335312366,
-0.014761959202587605,
-0.0368519052863121,
-0.052685271948575974,
0.02049579843878746,
0.004006491973996162,
0.11327008903026581,
0.040275946259498596,
0.04320801421999931,
0.019536109641194344,
0.0954442247748375,
0.03855455294251442,
-0.0412103533744812,
0.025121411308646202,
-0.07216615974903107,
0.008136213757097721
] |
Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations | https://openreview.net/forum?id=94kQgWXojH | [
"Nicholas Jiang",
"Anish Kachinthaya",
"Suzanne Petryk",
"Yossi Gandelsman"
] | Poster | We investigate the internal representations of vision-language models (VLMs) to address hallucinations, a persistent challenge despite advances in model size and training. We project VLMs’ internal image representations to their language vocabulary and observe more confident output probabilities on real objects than hallucinated objects. We additionally use these output probabilities to spatially localize real objects. Building on this approach, we introduce a knowledge erasure algorithm that removes hallucinations by linearly orthogonalizing image features with respect to hallucinated object features. We show that targeted edits to a model’s latent representations can reduce hallucinations by up to 25.7% on the COCO2014 dataset while preserving performance. Our findings demonstrate how a deeper understanding of VLMs’ latent representations can enhance reliability and enable novel capabilities, such as zero-shot segmentation. | Vision language models, hallucinations, logit lens, interpretability | null | 823 | 2410.02762 | [
0.04112919047474861,
-0.0629008412361145,
0.011895977891981602,
0.05632861331105232,
0.053563084453344345,
0.037716206163167953,
-0.04255613684654236,
-0.08130712807178497,
0.04993081092834473,
-0.05683382600545883,
-0.07370448857545853,
-0.059394046664237976,
0.008088621310889721,
0.07267813384532928,
-0.021824322640895844,
0.0143789267167449,
0.04452414810657501,
0.12951251864433289,
-0.07835068553686142,
-0.050644442439079285,
0.032724205404520035,
0.007521239574998617,
0.03706175461411476,
-0.004356371704488993,
0.10828793048858643,
0.024472905322909355,
-0.0186968632042408,
-0.09762009978294373,
0.015819421038031578,
-0.047934390604496,
-0.010236806236207485,
0.0826544314622879,
-0.0027664287481456995,
-0.0022567445412278175,
0.008332883007824421,
0.05884687975049019,
-0.04496386647224426,
0.04574035853147507,
-0.02228095382452011,
-0.0763300433754921,
-0.0011932493653148413,
0.05388622358441353,
0.024796126410365105,
0.048699334263801575,
-0.03077492117881775,
-0.01399553194642067,
0.013066831044852734,
0.012540035881102085,
0.0004015040467493236,
-0.053538478910923004,
-0.05215577036142349,
-0.007755593862384558,
-0.013218909502029419,
0.05086696892976761,
-0.05626561492681503,
0.007858323864638805,
0.06843316555023193,
0.005195532459765673,
0.006294116843491793,
0.01198346633464098,
-0.020499562844634056,
-0.05953904986381531,
0.03331824019551277,
-0.022165894508361816,
-0.029888970777392387,
-0.02058115415275097,
-0.002125775208696723,
0.016128357499837875,
0.002427108120173216,
-0.020427117124199867,
0.0033662898931652308,
0.036566365510225296,
-0.07020659744739532,
0.04344891011714935,
0.05683988705277443,
0.017596324905753136,
0.006336852442473173,
-0.12379835546016693,
0.052369989454746246,
-0.07053705304861069,
0.03740964084863663,
0.03178417682647705,
0.04736175015568733,
-0.004863656125962734,
0.1203794926404953,
0.004242556169629097,
0.03576213866472244,
0.046118054538965225,
-0.03512465953826904,
0.01977006532251835,
-0.05403364822268486,
-0.09840273857116699,
-0.014076299965381622,
-0.022181112319231033,
0.03300720080733299,
-0.030194267630577087,
0.0003678114735521376,
-0.0735318660736084,
-0.03840300440788269,
0.02854592725634575,
-0.013082250021398067,
-0.02991815097630024,
0.006403480656445026,
-0.01789676398038864,
0.028617726638913155,
-0.0230033490806818,
0.11766119301319122,
0.04247452691197395,
0.03296411409974098,
-0.002233268693089485,
0.008530143648386002,
0.022618139162659645,
-0.022305792197585106,
0.035760313272476196,
0.03262447938323021,
-0.05009916424751282,
0.004592285957187414,
0.025137318298220634,
0.013028576038777828,
0.04205351322889328,
-0.02953460067510605,
0.023856140673160553,
0.028747716918587685,
-0.011241179890930653,
-0.0071490793488919735,
-0.07854402810335159,
-0.058378320187330246,
-6.784723230233998e-36,
0.03705471009016037,
0.02313072793185711,
0.03898127004504204,
0.010421833023428917,
0.09820456802845001,
-0.03773033991456032,
-0.040753189474344254,
-0.0004371796967461705,
0.0424182265996933,
-0.052680131047964096,
-0.011238192208111286,
-0.026095334440469742,
-0.03797734156250954,
0.0895090401172638,
-0.01156814768910408,
0.033439476042985916,
-0.05537433177232742,
0.004699298646301031,
-0.057651929557323456,
0.015568796545267105,
-0.03635967895388603,
0.033666517585515976,
-0.07373598963022232,
-0.0005290540866553783,
-0.01604795828461647,
0.02409042976796627,
0.037744224071502686,
-0.07551229000091553,
-0.038497909903526306,
0.017760101705789566,
-0.15715278685092926,
0.11417542397975922,
0.03252774477005005,
0.037226464599370956,
0.0069765797816216946,
-0.017599642276763916,
0.0627032071352005,
-0.06557171046733856,
0.053980231285095215,
-0.011462253518402576,
0.033854953944683075,
0.07422665506601334,
-0.04412464052438736,
-0.07558050006628036,
-0.08801207691431046,
-0.0033894716762006283,
0.046225495636463165,
0.08602598309516907,
-0.10015430301427841,
0.042112380266189575,
0.059153951704502106,
0.07157441973686218,
-0.03541618585586548,
0.004783520009368658,
-0.0686093345284462,
-0.027599412947893143,
0.04472300410270691,
-0.06458023935556412,
0.058981072157621384,
-0.022191502153873444,
0.0673673078417778,
0.10044083744287491,
0.010458179749548435,
0.06419103592634201,
-0.03173688054084778,
-0.0191692803055048,
-0.014028314501047134,
-0.011040039360523224,
0.021975077688694,
-0.08371023833751678,
-0.10311785340309143,
0.029572390019893646,
-0.03421318903565407,
-0.012243540026247501,
-0.029858969151973724,
-0.03130267560482025,
0.02680346742272377,
-0.029367664828896523,
-0.007899410091340542,
0.06214252486824989,
-0.005284768994897604,
-0.052560314536094666,
-0.04895233362913132,
-0.07116109132766724,
-0.022329676896333694,
-0.033753372728824615,
0.016894938424229622,
-0.09440916776657104,
-0.06390997022390366,
-0.04246767237782478,
-0.013435686938464642,
-0.09448928385972977,
0.04978213831782341,
0.016523171216249466,
-0.032521288841962814,
-8.744355538973144e-34,
-0.0035366632509976625,
-0.015361600555479527,
0.06541246920824051,
0.0020405324175953865,
-0.03465031087398529,
-0.09274999797344208,
0.022271426394581795,
0.029758518561720848,
0.04765727370977402,
-0.06566895544528961,
-0.026049744337797165,
0.012290929444134235,
0.010137591511011124,
-0.015472888946533203,
0.04105961322784424,
-0.1045602485537529,
0.034518029540777206,
-0.008109578862786293,
-0.0568728931248188,
0.02910318225622177,
0.013142290525138378,
0.09816961735486984,
-0.0989624559879303,
0.008347551338374615,
-0.08516497910022736,
0.06388751417398453,
0.013526647351682186,
0.010176676325500011,
0.01291279960423708,
-0.03470038250088692,
0.016214095056056976,
0.03621833026409149,
0.00236110738478601,
-0.023194678127765656,
-0.04786785691976547,
0.01731528528034687,
0.008911210112273693,
-0.06263585388660431,
-0.08197823166847229,
-0.009329638443887234,
0.0601394847035408,
-0.014690984971821308,
-0.07023786008358002,
0.02184942178428173,
-0.028619801625609398,
-0.038997966796159744,
0.011174093931913376,
0.015625696629285812,
0.010561595670878887,
0.06304944306612015,
-0.042669638991355896,
0.024409115314483643,
-0.1270066201686859,
-0.006066512782126665,
-0.08860808610916138,
-0.08265798538923264,
-0.044746845960617065,
0.01828063651919365,
0.14956574141979218,
0.01275797188282013,
-0.0869336873292923,
0.038529638200998306,
-0.016353482380509377,
-0.05097000673413277,
-0.007058754097670317,
-0.015341716818511486,
-0.08592840284109116,
0.04792876914143562,
0.08865683525800705,
0.024882007390260696,
0.04466502368450165,
0.003918461501598358,
0.00209653377532959,
0.10764319449663162,
0.0014564420562237501,
-0.04757066071033478,
-0.004718413110822439,
-0.022031323984265327,
0.019167553633451462,
-0.08426032215356827,
0.006752086337655783,
-0.05358804389834404,
0.011826656758785248,
0.07446452975273132,
0.1121276393532753,
0.08338568359613419,
0.004202929325401783,
0.004242594353854656,
0.003971919883042574,
0.0489819310605526,
-0.06553837656974792,
-0.00278608500957489,
-0.02293277159333229,
0.08499467372894287,
-0.08720318228006363,
-3.2747141887057296e-8,
-0.09772492945194244,
0.0200502872467041,
-0.023036697879433632,
0.01644153520464897,
-0.01151334960013628,
-0.08671964704990387,
-0.022656822577118874,
0.14605645835399628,
0.0002150703512597829,
-0.02096426673233509,
0.12845498323440552,
-0.020260682329535484,
-0.011182289570569992,
-0.11138395220041275,
-0.017640898004174232,
0.064247265458107,
0.04513314366340637,
0.13338156044483185,
0.012211352586746216,
-0.03310411423444748,
0.04046616330742836,
0.005374856758862734,
0.033488351851701736,
-0.03706461563706398,
0.030934378504753113,
-0.004080579616129398,
-0.0022606311831623316,
0.08502164483070374,
0.003458341583609581,
-0.01224250067025423,
0.050879452377557755,
0.07075578719377518,
0.08634926378726959,
0.06501295417547226,
0.0029713897965848446,
0.06483562290668488,
-0.005446483381092548,
-0.00003190550341969356,
0.011327342130243778,
-0.05021176487207413,
0.06553026288747787,
0.006772132590413094,
0.0059303524903953075,
0.010207387618720531,
-0.028598111122846603,
-0.016346869990229607,
0.06260427087545395,
-0.10050445050001144,
0.034903548657894135,
-0.013226869516074657,
-0.024024007841944695,
0.07748329639434814,
-0.02030608057975769,
0.12220878154039383,
0.0733332708477974,
0.04987208545207977,
0.05037666857242584,
-0.011892279610037804,
0.018787972629070282,
0.03737896680831909,
0.006294759921729565,
0.007712458726018667,
-0.06873417645692825,
0.017546987161040306
] |
Robustness of Quantum Algorithms for Nonconvex Optimization | https://openreview.net/forum?id=JyQYYjtO88 | [
"Weiyuan Gong",
"Chenyi Zhang",
"Tongyang Li"
] | Poster | In this paper, we systematically study quantum algorithms for finding an $\epsilon$-approximate second-order stationary point ($\epsilon$-SOSP) of a $d$-dimensional nonconvex function, a fundamental problem in nonconvex optimization, with noisy zeroth- or first-order oracles as inputs. We first prove that, up to noise of $O(\epsilon^{10}/d^5)$, perturbed accelerated gradient descent equipped with quantum gradient estimation takes $O(\log d/\epsilon^{1.75})$ quantum queries to find an $\epsilon$-SOSP. We then prove that standard perturbed gradient descent is robust to the noise of $O(\epsilon^6/d^4)$ and $O(\epsilon/d^{0.5+\zeta})$ for any $\zeta>0$ on the zeroth- and first-order oracles, respectively, which provides a quantum algorithm with poly-logarithmic query complexity. Furthermore, we propose a stochastic gradient descent algorithm using quantum mean estimation on the Gaussian smoothing of noisy oracles, which is robust to $O(\epsilon^{1.5}/d)$ and $O(\epsilon/\sqrt{d})$ noise on the zeroth- and first-order oracles, respectively. The quantum algorithm takes $O(d^{2.5}/\epsilon^{3.5})$ and $O(d^2/\epsilon^3)$ queries to the two oracles, giving a polynomial speedup over the classical counterparts. As a complement, we characterize the domains where quantum algorithms can find an $\epsilon$-SOSP with poly-logarithmic, polynomial, or exponential number of queries in $d$, or the problem is information-theoretically unsolvable even with an infinite number of queries. In addition, we prove an $\Omega(\epsilon^{-12/7})$ lower bound on $\epsilon$ for any randomized classical and quantum algorithm to find an $\epsilon$-SOSP using either noisy zeroth- or first-order oracles. | Nonconvex optimization, Robustness, Quantum algorithms | null | 822 | 2212.02548 | [
-0.07749295234680176,
-0.02727465145289898,
-0.01838536374270916,
0.03987043350934982,
-0.05151400342583656,
-0.01240502204746008,
0.021915026009082794,
-0.08759640902280807,
0.012239782139658928,
-0.04538063332438469,
0.0012783226557075977,
0.02005280740559101,
0.04504520073533058,
0.005139673128724098,
-0.0731329694390297,
0.028286946937441826,
0.10030851513147354,
0.06371286511421204,
-0.002158146118745208,
0.034322820603847504,
-0.0072128428146243095,
0.06489812582731247,
0.03365297615528107,
-0.013761669397354126,
0.12185557186603546,
-0.06026162952184677,
-0.02178208902478218,
0.010222097858786583,
0.015743065625429153,
0.028310582041740417,
-0.020476773381233215,
-0.014959696680307388,
-0.029443414881825447,
-0.04534895345568657,
0.012381480075418949,
-0.006123701110482216,
0.030050193890929222,
-0.0024303344544023275,
0.01799933612346649,
0.00003892109452863224,
-0.02109408751130104,
0.08149141073226929,
-0.08645385503768921,
-0.00011530062329256907,
-0.0014320514164865017,
0.03585260733962059,
0.04941129311919212,
-0.032702941447496414,
0.01388885173946619,
-0.14816199243068695,
-0.013077132403850555,
0.008606956340372562,
-0.006558930967003107,
-0.0297817662358284,
-0.041135236620903015,
-0.031560130417346954,
-0.011790631338953972,
0.003930253442376852,
0.031080177053809166,
-0.005632414948195219,
-0.0010082991793751717,
-0.07217980176210403,
0.06469909846782684,
0.014745153486728668,
0.03589801862835884,
-0.020390521734952927,
-0.004447752609848976,
-0.0388721227645874,
-0.004459980875253677,
0.09001544117927551,
-0.07700244337320328,
-0.0005498916143551469,
-0.08507855981588364,
0.03217830881476402,
0.004672233015298843,
-0.011053758673369884,
0.0686456710100174,
0.02849586494266987,
-0.01042979210615158,
-0.022508136928081512,
0.04637942463159561,
-0.01815040409564972,
-0.014158312231302261,
-0.03038698434829712,
0.06739579141139984,
-0.013614891096949577,
-0.011172724887728691,
0.02308027632534504,
0.026015667244791985,
-0.10157769173383713,
-0.06664139032363892,
-0.030883535742759705,
-0.11522243916988373,
-0.05858386680483818,
0.07442589849233627,
-0.01877479813992977,
0.04698948934674263,
0.016064738854765892,
0.05566222965717316,
0.045780956745147705,
0.06343892216682434,
0.05374119430780411,
-0.00833064690232277,
0.0026595783419907093,
0.024704083800315857,
0.008564162068068981,
0.015585102140903473,
0.03379698842763901,
0.02790149487555027,
-0.017062293365597725,
0.049040284007787704,
0.02677181176841259,
0.010389681905508041,
-0.009585794992744923,
-0.023564934730529785,
0.01677600108087063,
0.03574148938059807,
0.025306744500994682,
-0.011886699125170708,
0.046155836433172226,
-0.06811415404081345,
-0.026363668963313103,
0.01888297125697136,
0.06790316104888916,
0.02505631372332573,
0.06075461208820343,
-0.09415227174758911,
1.834241201352105e-33,
0.03670363873243332,
0.10533881187438965,
0.0577087439596653,
-0.10096535086631775,
-0.007074978668242693,
0.060618143528699875,
0.13120532035827637,
-0.040107131004333496,
-0.061547230929136276,
0.04213487729430199,
0.0030392450280487537,
-0.05545945093035698,
-0.04731786623597145,
-0.04202002286911011,
0.09847116470336914,
0.016257701441645622,
-0.07608421891927719,
-0.06068138778209686,
0.03406595066189766,
-0.06709490716457367,
0.11315035074949265,
-0.08433420956134796,
-0.08265633136034012,
0.011163747869431973,
0.00278309709392488,
-0.03853249549865723,
0.1576988250017166,
-0.005607214290648699,
-0.04611876979470253,
0.001236651442013681,
-0.09034675359725952,
0.020488062873482704,
-0.028271645307540894,
-0.05205613374710083,
0.06772950291633606,
0.020886415615677834,
-0.018313806504011154,
0.02748166397213936,
-0.008878549560904503,
-0.046283092349767685,
-0.031716395169496536,
0.01928151771426201,
0.053311336785554886,
-0.030253564938902855,
-0.037806179374456406,
-0.013391485437750816,
-0.03431491553783417,
0.052800558507442474,
0.05110843852162361,
-0.09695816040039062,
-0.0377785786986351,
0.017247920855879784,
-0.09059741348028183,
0.04799913987517357,
0.04893150553107262,
-0.06656355410814285,
0.03740415722131729,
-0.0057533904910087585,
-0.057459864765405655,
0.07501854002475739,
-0.00718893576413393,
-0.03397117555141449,
-0.049661099910736084,
-0.029584739357233047,
-0.0726708397269249,
0.02228095382452011,
0.009063255973160267,
0.03212639316916466,
-0.034635454416275024,
0.03841052204370499,
-0.004323400557041168,
0.0018407455645501614,
0.022655131295323372,
-0.08879213780164719,
0.10372401028871536,
-0.008292089216411114,
0.05676228180527687,
-0.10316210985183716,
0.03733743727207184,
-0.0385454036295414,
0.018796633929014206,
-0.05663623660802841,
0.005362604279071093,
0.036575112491846085,
-0.030220674350857735,
-0.15813001990318298,
-0.017177345231175423,
-0.017466401681303978,
-0.06756793707609177,
-0.022594796493649483,
-0.07766067236661911,
-0.049104269593954086,
0.04167213290929794,
0.04926418513059616,
-0.0991203710436821,
-2.4045313672985763e-33,
-0.032291166484355927,
-0.021281328052282333,
0.0034991854336112738,
0.034809570759534836,
0.01600506156682968,
-0.0755377784371376,
-0.0031789380591362715,
0.010450916364789009,
-0.015074474737048149,
-0.06912091374397278,
-0.05929199606180191,
0.04216805472970009,
0.0764097198843956,
0.09598991274833679,
0.026244955137372017,
0.09338048845529556,
-0.034579601138830185,
0.06884881854057312,
0.0431385338306427,
0.09643533825874329,
0.03222016617655754,
0.02282005175948143,
0.00830900389701128,
0.04862328618764877,
0.006412227172404528,
0.0746920257806778,
0.016042066738009453,
0.05708998441696167,
-0.09448320418596268,
-0.01444937102496624,
-0.016893640160560608,
-0.06036924198269844,
-0.08148858696222305,
0.008120516315102577,
-0.0842113047838211,
0.010611973702907562,
-0.041762661188840866,
-0.006889375392347574,
-0.03364163264632225,
0.03832469880580902,
-0.025314919650554657,
-0.014730136841535568,
0.002364270156249404,
-0.01970483548939228,
0.07013238966464996,
0.03318791091442108,
-0.0795622318983078,
-0.012458343990147114,
-0.039818160235881805,
0.10619942098855972,
-0.027380183339118958,
0.029607810080051422,
-0.012951265089213848,
0.15683390200138092,
-0.0530938096344471,
-0.012619109824299812,
-0.05831877514719963,
0.05045539140701294,
0.047836463898420334,
0.0064598494209349155,
-0.05099714547395706,
0.04651358351111412,
-0.057759515941143036,
0.1126270666718483,
0.03444546461105347,
-0.012378941290080547,
-0.027366574853658676,
-0.007399266120046377,
0.029813701286911964,
-0.015087948180735111,
-0.08766350895166397,
-0.06690669804811478,
0.0036094968672841787,
0.044912707060575485,
-0.020885717123746872,
0.05910186469554901,
0.03492676466703415,
-0.023824481293559074,
0.0010153987677767873,
-0.005769082810729742,
0.05064627528190613,
0.10638628900051117,
0.030379146337509155,
0.04073694348335266,
0.04171021282672882,
0.05225079879164696,
0.06749825179576874,
0.0008916464284993708,
0.024368925020098686,
-0.015203225426375866,
0.006269039586186409,
-0.011784696020185947,
0.06380148977041245,
-0.11249849200248718,
0.0012360885739326477,
-3.980703056072343e-8,
-0.06284049898386002,
-0.05786514654755592,
-0.00004732061279355548,
-0.05773189291357994,
0.14355400204658508,
-0.03340645879507065,
0.019969379529356956,
0.009172990918159485,
-0.01591334491968155,
-0.06377194076776505,
0.02938014268875122,
-0.07477612048387527,
-0.06834638863801956,
-0.025824811309576035,
-0.004789137281477451,
0.04498171806335449,
-0.045839160680770874,
-0.021898774430155754,
-0.008249937556684017,
-0.10338626801967621,
-0.07406026870012283,
0.107772096991539,
0.0023101314436644316,
-0.00872742384672165,
-0.0005947754834778607,
-0.03869210556149483,
0.07623106241226196,
-0.011645465157926083,
0.08649515360593796,
0.03263130411505699,
-0.03740578889846802,
-0.016757596284151077,
0.04596436023712158,
0.05206803232431412,
-0.049609921872615814,
0.07486896216869354,
0.04086269065737724,
0.011108470149338245,
-0.021065086126327515,
-0.012170160189270973,
-0.060155440121889114,
0.04449276626110077,
0.016483377665281296,
-0.0023479696828871965,
0.06217409297823906,
-0.010163974016904831,
-0.01160485204309225,
-0.05387326702475548,
0.06775220483541489,
0.06435530632734299,
0.020546847954392433,
0.004225776996463537,
0.01470230519771576,
-0.030953682959079742,
-0.02934543415904045,
0.016381515190005302,
-0.07692544907331467,
-0.08478918671607971,
-0.01347512286156416,
0.060132697224617004,
0.033573638647794724,
-0.0014831025619059801,
0.027282310649752617,
0.0215484406799078
] |
Fourier Head: Helping Large Language Models Learn Complex Probability Distributions | https://openreview.net/forum?id=4hPwLg7zD3 | [
"Nate Gillman",
"Daksh Aggarwal",
"Michael Freeman",
"Chen Sun"
] | Poster | As the quality of large language models has improved, there has been increased interest in using them to model non-linguistic tokens. For example, the Decision Transformer recasts agentic decision making as a sequence modeling problem, using a decoder-only LLM to model the distribution over the discrete action space for an Atari agent. However, when adapting LLMs to non-linguistic domains, it remains unclear if softmax over discrete bins captures the continuous structure of the tokens and the potentially complex distributions needed for high quality token generation. We introduce a neural network layer, constructed using Fourier series, which we can easily substitute for any linear layer if we want the outputs to have a more continuous structure. We perform extensive analysis on synthetic datasets, as well as on large-scale decision making and time series forecasting tasks. We also provide theoretical evidence that this layer can better learn signal from data while ignoring high-frequency noise. All of our results support the effectiveness of our proposed Fourier head in scenarios where the underlying data distribution has a natural continuous structure. For example, the Fourier head improves a Decision Transformer agent's returns across four benchmark Atari games by as much as 377\%, and increases a state-of-the-art times series foundation model's forecasting performance by 3.5\% across 20 benchmarks unseen during training.
We release our implementation at https://nategillman.com/fourier-head | LLM, Fourier, smooth function, multi-class classification | Using Fourier series, we build a neural network layer which learns categorical distributions that have a continuous structure. | 818 | 2410.22269 | [
-0.11429668217897415,
-0.11913058906793594,
0.0031088434625416994,
-0.026287024840712547,
-0.007575050462037325,
0.08396584540605545,
-0.015054108574986458,
-0.030904583632946014,
0.08155164867639542,
-0.03219961002469063,
-0.059436094015836716,
-0.03919263929128647,
-0.0008932239143177867,
0.053686365485191345,
-0.016579709947109222,
-0.03263574466109276,
0.08461684733629227,
0.1111249178647995,
-0.0509130023419857,
-0.08523482084274292,
0.06032261252403259,
0.02606513537466526,
0.027134066447615623,
-0.036741938441991806,
0.04714485630393028,
0.059725843369960785,
0.006845591589808464,
-0.036642592400312424,
0.02051912061870098,
-0.018720446154475212,
0.0466216541826725,
0.0008348322589881718,
0.024903196841478348,
-0.0071916538290679455,
-0.08518578857183456,
0.055909980088472366,
-0.05600999295711517,
-0.013126303441822529,
0.04819421470165253,
-0.025407737120985985,
0.0016768458299338818,
0.02657482586801052,
-0.012991571798920631,
-0.010845358483493328,
0.0567278191447258,
-0.037065763026475906,
0.036378223448991776,
-0.06860049068927765,
-0.0027186363004148006,
0.04801885038614273,
-0.00017494458006694913,
0.027162615209817886,
0.024537859484553337,
0.08739154785871506,
-0.05218211188912392,
0.003159524407237768,
0.01436100248247385,
0.05816170200705528,
-0.02196390926837921,
-0.04192553460597992,
-0.058930762112140656,
-0.1117846816778183,
0.04769851639866829,
-0.045757245272397995,
-0.04912994056940079,
-0.0019375255797058344,
-0.021838029846549034,
0.0830083042383194,
0.004759054630994797,
0.02568509615957737,
-0.042580973356962204,
0.06518059968948364,
-0.07987915724515915,
0.0019929655827581882,
0.010188571177423,
0.002728438237681985,
0.030744923278689384,
-0.02244800701737404,
0.05914241448044777,
-0.027491968125104904,
-0.014535791240632534,
-0.03367731720209122,
0.06161556765437126,
-0.017404992133378983,
0.11086993664503098,
0.07925957441329956,
-0.04557503014802933,
0.06560954451560974,
-0.05577786639332771,
0.006211468949913979,
-0.08643456548452377,
-0.014788031578063965,
0.13290849328041077,
0.004390446934849024,
0.022128816694021225,
0.11149342358112335,
-0.07500346004962921,
-0.03335234895348549,
-0.0006657723570242524,
0.08052526414394379,
-0.007933822460472584,
-0.011284174397587776,
0.0012315323110669851,
-0.06723292171955109,
0.027186786755919456,
-0.04855183884501457,
0.0156526081264019,
-0.0012047376949340105,
0.017125163227319717,
-0.09383127838373184,
-0.008810050785541534,
0.023086365312337875,
0.034984249621629715,
0.054014649242162704,
0.044703129678964615,
-0.01598098874092102,
-0.006576664745807648,
-0.034617457538843155,
-0.03504015877842903,
0.07485577464103699,
0.014544173143804073,
0.025187809020280838,
-0.04506026208400726,
0.08727014064788818,
0.021851828321814537,
0.019159942865371704,
-0.022968871518969536,
1.688388433695968e-33,
-0.025897325947880745,
0.07984351366758347,
0.01434555184096098,
-0.035478439182043076,
0.05032510310411453,
-0.0048979055136442184,
0.07066959887742996,
0.04645930603146553,
0.042992547154426575,
0.018661191686987877,
-0.056463323533535004,
0.09279381483793259,
-0.10766282677650452,
0.046583354473114014,
0.05120024085044861,
-0.024096548557281494,
-0.03393950313329697,
0.009416969493031502,
0.012069059535861015,
-0.047443076968193054,
0.02638350985944271,
-0.019014079123735428,
0.029032045975327492,
-0.07970447093248367,
-0.03121970221400261,
0.05213206633925438,
0.03791312873363495,
-0.06485241651535034,
-0.04074819013476372,
-0.00460117869079113,
-0.13491129875183105,
-0.009823059663176537,
0.011005692183971405,
-0.024720292538404465,
0.09344237297773361,
-0.03356028348207474,
-0.037892233580350876,
0.028484100475907326,
-0.07369055598974228,
0.017283760011196136,
-0.02115166373550892,
0.01748250052332878,
-0.05307571962475777,
0.04831334203481674,
-0.05147264897823334,
-0.04505905508995056,
0.006426542066037655,
-0.034041859209537506,
0.0015296840574592352,
-0.0786079615354538,
0.04843622446060181,
-0.03314394876360893,
-0.012893440201878548,
-0.061980437487363815,
0.054215021431446075,
0.023486196994781494,
-0.02308139018714428,
0.03628481924533844,
-0.031492576003074646,
0.07464620471000671,
0.021310102194547653,
0.03228485956788063,
0.06050606817007065,
0.0755993127822876,
0.02640438824892044,
0.05701855942606926,
-0.03053426742553711,
0.11664386093616486,
0.06524543464183807,
-0.014560742303729057,
0.04312628507614136,
0.09427712112665176,
0.009065723046660423,
-0.03504433110356331,
0.08058999478816986,
0.0013085926184430718,
0.05111178010702133,
-0.08510082960128784,
-0.09889114648103714,
0.03142160922288895,
0.050298236310482025,
-0.05108404532074928,
-0.020508138462901115,
-0.033224355429410934,
-0.01698155514895916,
-0.0013525839895009995,
0.04050097241997719,
-0.03320150077342987,
0.007529503665864468,
-0.06052584573626518,
-0.11224450916051865,
-0.057965148240327835,
-0.009395070374011993,
0.022923260927200317,
-0.020213492214679718,
-2.0222003603256888e-33,
-0.06750534474849701,
0.04242192953824997,
-0.058088354766368866,
0.03537119925022125,
-0.06064971163868904,
-0.04564271122217178,
-0.0550607405602932,
0.06610770523548126,
0.03296928480267525,
0.022907322272658348,
-0.04859084635972977,
-0.02813282050192356,
0.09154589474201202,
0.010760942474007607,
0.02097325585782528,
-0.0742911547422409,
-0.04345914348959923,
-0.026150263845920563,
0.05434558913111687,
-0.03253215178847313,
-0.05435190349817276,
0.028871705755591393,
-0.1814013123512268,
0.03793554753065109,
-0.033878717571496964,
0.049838703125715256,
-0.10729186236858368,
0.0591142438352108,
0.02432878687977791,
0.016210105270147324,
-0.03562035784125328,
-0.08072103559970856,
-0.0027424339205026627,
-0.0035006606485694647,
-0.06173968315124512,
0.030245672911405563,
0.0849689468741417,
0.037866510450839996,
-0.08056094497442245,
0.022747071459889412,
0.01557182241231203,
-0.08978615701198578,
-0.024282129481434822,
-0.015971016138792038,
-0.0037833687383681536,
-0.01787925697863102,
-0.08215650171041489,
-0.023183029145002365,
-0.003823506413027644,
0.03042585216462612,
0.03386417031288147,
0.043399710208177567,
-0.058616988360881805,
0.010282156988978386,
-0.07844486832618713,
-0.03601713106036186,
-0.057499729096889496,
0.007502383086830378,
0.07327967882156372,
-0.030955953523516655,
-0.08221489191055298,
-0.04819580912590027,
0.044996701180934906,
-0.055998630821704865,
0.023765133693814278,
-0.029133034870028496,
-0.026645246893167496,
0.0066388314589858055,
0.04722535237669945,
-0.007532328367233276,
0.10034570097923279,
-0.02039482444524765,
0.036199405789375305,
0.09375638514757156,
0.005287505686283112,
-0.026213541626930237,
-0.002500348724424839,
-0.08085307478904724,
-0.047489140182733536,
0.029981477186083794,
-0.062125205993652344,
0.024223001673817635,
-0.023088568821549416,
0.029602721333503723,
0.11277300119400024,
0.07116968929767609,
0.047955531626939774,
0.028073444962501526,
0.022919554263353348,
-0.013386066071689129,
0.042474351823329926,
0.07066884636878967,
-0.005147221963852644,
0.049877192825078964,
-0.05177697911858559,
-4.800566344442814e-8,
-0.02639772929251194,
0.039371006190776825,
0.003396134590730071,
0.00044696571421809494,
0.05892978608608246,
-0.03531091287732124,
-0.017376387491822243,
-0.013485024683177471,
-0.04442708194255829,
-0.06276953965425491,
0.0495467483997345,
-0.01699833571910858,
-0.040421970188617706,
-0.031927891075611115,
0.005949551705271006,
0.0845903530716896,
0.014498171396553516,
0.009892423637211323,
0.04605720564723015,
-0.06900163739919662,
0.04802256077528,
0.12869296967983246,
0.018454425036907196,
-0.018405186012387276,
0.017119580879807472,
-0.0685557946562767,
-0.03201385959982872,
0.014901508577167988,
0.03213243559002876,
0.09895312786102295,
0.019573330879211426,
0.010187750682234764,
0.023021342232823372,
-0.0660635381937027,
-0.02797391265630722,
0.018463371321558952,
0.0430806428194046,
-0.05066850036382675,
-0.038586266338825226,
-0.007182758767157793,
0.03343122452497482,
0.06964307278394699,
-0.04289593547582626,
-0.0028907114174216986,
0.033719323575496674,
-0.03965025022625923,
0.05205422267317772,
-0.13091669976711273,
0.11420315504074097,
-0.019787635654211044,
0.05566508695483208,
0.010501743294298649,
-0.006297186017036438,
0.020047955214977264,
0.10285520553588867,
0.02457774616777897,
-0.04084663838148117,
-0.10147752612829208,
-0.03487658500671387,
0.02381962165236473,
-0.02236434444785118,
0.044365350157022476,
-0.024370966479182243,
0.01328661572188139
] |
Locality Sensitive Avatars From Video | https://openreview.net/forum?id=SVta2eQNt3 | [
"Chunjin Song",
"Zhijie Wu",
"Shih-Yang Su",
"Bastian Wandt",
"Leonid Sigal",
"Helge Rhodin"
] | Poster | We present locality-sensitive avatar, a neural radiance field (NeRF) based network to learn human motions from monocular videos. To this end, we estimate a canonical representation between different frames of a video with a non-linear mapping from observation to canonical space, which we decompose into a skeletal rigid motion and a non-rigid counterpart. Our key contribution is to retain fine-grained details by modeling the non-rigid part with a graph neural network (GNN) that keeps the pose information local to neighboring body parts. Compared to former canonical representation based methods which solely operate on the coordinate space of a whole shape, our locality-sensitive motion modeling can reproduce both realistic shape contours and vivid fine-grained details. We evaluate on ZJU-MoCap, SynWild, ActorsHQ, MVHumanNet and various outdoor videos. The experiments reveal that with the locality sensitive deformation to canonical feature space, we are the first to achieve state-of-the-art results across novel view synthesis, novel pose animation and 3D shape reconstruction simultaneously. Our code is available at https://github.com/ChunjinSong/lsavatar. | 3D Computer Vision, Neural Rendering, Avatar Modeling | null | 817 | null | [
-0.011634272523224354,
-0.10354563593864441,
0.07550409436225891,
-0.048368558287620544,
0.06957918405532837,
-0.011811299249529839,
-0.03246908262372017,
-0.030222788453102112,
0.02209990844130516,
-0.009141418151557446,
0.0029008041601628065,
-0.0029330970719456673,
-0.04622972011566162,
0.11760962009429932,
-0.06982696801424026,
-0.042816415429115295,
0.10385420173406601,
0.10743723064661026,
-0.06777836382389069,
0.0867290198802948,
-0.0047640991397202015,
0.028325283899903297,
0.057814501225948334,
0.018785128369927406,
0.09918880462646484,
-0.01872912049293518,
0.06168506294488907,
-0.07379286736249924,
0.02435890957713127,
-0.05953614413738251,
0.008761835284531116,
0.0188030693680048,
-0.10782109946012497,
-0.02452676370739937,
-0.03765961900353432,
0.08428853750228882,
0.0035356516018509865,
0.027632100507616997,
-0.06973245739936829,
-0.0638984963297844,
0.021744687110185623,
0.10556995123624802,
0.02068171463906765,
0.049841031432151794,
0.15288442373275757,
0.09226957708597183,
0.042961493134498596,
-0.05614352971315384,
-0.019011352211236954,
0.038129281252622604,
-0.07293479144573212,
-0.018110191449522972,
0.024016747251152992,
0.015338636934757233,
-0.008959507569670677,
0.012583140283823013,
-0.00783013179898262,
-0.11646272987127304,
0.04208879545331001,
0.03543616086244583,
0.06769494712352753,
0.0027210258413106203,
0.0010077282786369324,
-0.007838338613510132,
-0.004895367193967104,
-0.08743742853403091,
0.03826356306672096,
-0.022853972390294075,
0.003251488786190748,
-0.005776116158813238,
-0.046677228063344955,
0.010305916890501976,
-0.008195517584681511,
-0.015133346430957317,
-0.04152680188417435,
0.02398734912276268,
0.07406257838010788,
-0.007497397717088461,
0.008302914910018444,
-0.06099439412355423,
0.06837522983551025,
-0.050222836434841156,
0.06383316218852997,
-0.037643857300281525,
0.05170468986034393,
0.02496952936053276,
-0.01380965020507574,
0.10394745320081711,
-0.04441335052251816,
0.006043743807822466,
-0.1454758197069168,
0.020577404648065567,
0.027024205774068832,
-0.11374495923519135,
0.030580319464206696,
-0.0314151868224144,
0.02565748617053032,
-0.02483450248837471,
0.016149990260601044,
0.11625318229198456,
0.005911383777856827,
-0.016688272356987,
0.06831850856542587,
0.04218931868672371,
0.06790786236524582,
0.01765829138457775,
0.029918266460299492,
-0.019276021048426628,
0.062937431037426,
-0.003139032516628504,
0.023046329617500305,
-0.023027241230010986,
-0.047195807099342346,
0.05340227112174034,
0.05381409451365471,
0.021357355639338493,
0.011899387463927269,
0.012123195454478264,
-0.030903397127985954,
-0.01884514093399048,
-0.030181121081113815,
-0.01692182756960392,
-0.04039521515369415,
-0.03488665446639061,
0.023890497162938118,
-0.012410505674779415,
-0.005782334133982658,
-2.431297006324815e-34,
0.06749524176120758,
0.06040572002530098,
0.05865611508488655,
0.04843221604824066,
0.0521453432738781,
0.0019381199963390827,
-0.06050078198313713,
-0.0029105795547366142,
-0.009654881432652473,
-0.015401062555611134,
0.008565487340092659,
-0.09180402010679245,
0.04454353451728821,
0.09429588913917542,
-0.024756427854299545,
-0.027903402224183083,
-0.017868896946310997,
-0.026952479034662247,
0.01926920749247074,
0.09420805424451828,
0.0032023729290813208,
-0.0006853332160972059,
-0.038488779217004776,
0.010493477806448936,
-0.09535618126392365,
-0.029078025370836258,
0.024059582501649857,
-0.0363498255610466,
-0.0405273400247097,
-0.021209530532360077,
-0.0340392030775547,
0.009991928935050964,
-0.005941419862210751,
0.019724473357200623,
0.06836152076721191,
-0.05575774982571602,
-0.007036480121314526,
-0.10280211269855499,
0.04500611871480942,
-0.01884366013109684,
0.05938982591032982,
0.028132693842053413,
-0.06752954423427582,
-0.0584445521235466,
-0.0552656315267086,
0.0017147694015875459,
0.09591484069824219,
0.10442556440830231,
-0.06024777144193649,
0.02999303489923477,
0.0642361044883728,
0.02321845293045044,
-0.008918983861804008,
-0.10314951837062836,
-0.07375423610210419,
0.004177634604275227,
0.017305180430412292,
-0.04827743396162987,
0.03232765942811966,
0.0021536776330322027,
-0.01647343672811985,
-0.017255160957574844,
-0.041420698165893555,
-0.04493546485900879,
-0.058908477425575256,
-0.013937178999185562,
0.02413204498589039,
0.06902173906564713,
0.0882255807518959,
0.025556808337569237,
-0.029909757897257805,
0.045044172555208206,
-0.0020602094009518623,
-0.013597655110061169,
0.037264350801706314,
0.05734019726514816,
0.015659470111131668,
-0.026620347052812576,
-0.06326140463352203,
0.06853470951318741,
-0.04217700660228729,
-0.0037590591236948967,
-0.0544167123734951,
0.014432651922106743,
-0.03721926361322403,
-0.04703902453184128,
0.04040646180510521,
-0.026802878826856613,
0.0048034884966909885,
-0.033856768161058426,
-0.010950569063425064,
-0.08248793333768845,
-0.029574116691946983,
0.012575165368616581,
0.0014271611580625176,
1.9986304128947557e-33,
0.02946680597960949,
0.012221035547554493,
-0.04428635910153389,
-0.012066006660461426,
0.02654721401631832,
-0.07038212567567825,
-0.010175183415412903,
0.022977452725172043,
0.004781882278621197,
-0.06716747581958771,
-0.0426393523812294,
-0.07341068983078003,
0.004738340154290199,
-0.07710541784763336,
0.04963246360421181,
0.044190213084220886,
-0.007308941800147295,
-0.02505618892610073,
-0.01753219962120056,
-0.004431131761521101,
0.05415046960115433,
0.05986344441771507,
-0.0393296554684639,
0.0021669717971235514,
-0.07410098612308502,
0.14394772052764893,
0.012030444107949734,
0.03432176634669304,
-0.04919608309864998,
0.015276945196092129,
-0.03103645145893097,
-0.051060158759355545,
0.0023056736681610346,
-0.03922872617840767,
-0.08600881695747375,
0.06825146079063416,
0.02028580941259861,
-0.004008322488516569,
-0.03577778488397598,
-0.03316427022218704,
-0.0022139486391097307,
-0.008915354497730732,
-0.07829709351062775,
0.050594694912433624,
-0.02050568163394928,
-0.024550089612603188,
-0.039579201489686966,
0.006839540787041187,
-0.05465839058160782,
0.07492895424365997,
-0.05170794203877449,
0.07698087394237518,
-0.05154155567288399,
-0.0278448648750782,
-0.014485268853604794,
-0.06125952675938606,
-0.0038551425095647573,
-0.0057294415310025215,
0.09906768053770065,
-0.000729189720004797,
-0.012261169031262398,
-0.027403678745031357,
-0.06756461411714554,
0.058577172458171844,
-0.028043635189533234,
0.038541924208402634,
0.001945611904375255,
-0.022897979244589806,
-0.012487274594604969,
0.02934602089226246,
0.036541249603033066,
-0.016890741884708405,
0.013528259471058846,
0.020967435091733932,
0.0458681546151638,
-0.08392354100942612,
0.06340795755386353,
0.03579733893275261,
0.0594758614897728,
-0.06406907737255096,
0.04610391706228256,
-0.08312682807445526,
0.014072470366954803,
0.09825699031352997,
0.08992134034633636,
0.13842515647411346,
-0.01521693728864193,
0.04054027050733566,
0.007411758881062269,
0.07839591056108475,
-0.055263761430978775,
0.0871606096625328,
-0.08564280718564987,
0.025269782170653343,
0.0573565810918808,
-3.783266322443524e-8,
-0.11976032704114914,
0.06168682873249054,
-0.009270287118852139,
-0.006767003331333399,
-0.05442522466182709,
0.004161007236689329,
0.07126099616289139,
0.07226034998893738,
-0.027828697115182877,
0.029596922919154167,
0.08007398247718811,
-0.026532528921961784,
0.0028491285629570484,
-0.042259834706783295,
-0.05737122520804405,
0.06480573117733002,
-0.014968267641961575,
0.017898520454764366,
-0.04206326603889465,
0.08075765520334244,
-0.025836003944277763,
0.027521315962076187,
0.030741794034838676,
0.007526559289544821,
0.033913686871528625,
-0.056035902351140976,
-0.051203444600105286,
0.08003460615873337,
0.07638071477413177,
0.006178506184369326,
-0.04748912528157234,
0.04702788218855858,
0.06692960858345032,
-0.08204202353954315,
-0.04896732419729233,
-0.03384280204772949,
0.015737071633338928,
-0.04643740877509117,
-0.029256101697683334,
-0.01140563003718853,
-0.0049493745900690556,
0.08392947167158127,
0.09026113897562027,
0.03990491107106209,
0.005326363258063793,
0.06281678378582001,
0.059287700802087784,
-0.09765468537807465,
0.08160822838544846,
0.016263818368315697,
-0.042112283408641815,
-0.036482539027929306,
-0.035590771585702896,
0.12342040985822678,
0.026115979999303818,
0.029020879417657852,
-0.028399210423231125,
-0.021197320893406868,
0.05964191257953644,
0.027564624324440956,
-0.01472669467329979,
-0.009815588593482971,
-0.05048529803752899,
-0.04395103454589844
] |
Unhackable Temporal Reward for Scalable Video MLLMs | https://openreview.net/forum?id=Gf1uBeuUJW | [
"En Yu",
"Kangheng Lin",
"Liang Zhao",
"Yana Wei",
"Zining Zhu",
"Haoran Wei",
"Jianjian Sun",
"Zheng Ge",
"Xiangyu Zhang",
"Jingyu Wang",
"Wenbing Tao"
] | Poster | In the pursuit of superior video-processing MLLMs, we have encountered a perplexing paradox: the “anti-scaling law”, where more data and larger models lead to worse performance. This study unmasks the culprit: “temporal hacking”, a phenomenon where models shortcut by fixating on select frames, missing the full video narrative. In this work, we systematically establish a comprehensive theory of temporal hacking, defining it from a reinforcement learning perspective, introducing the Temporal Perplexity (TPL) score to assess this misalignment, and proposing the Unhackable Temporal Rewarding (UTR) framework to mitigate the temporal hacking. Both theoretically and empirically, TPL proves to be a reliable indicator of temporal modeling quality, correlating strongly with frame activation patterns. Extensive experiments reveal that UTR not only counters temporal hacking but significantly elevates video comprehension capabilities. This work not only advances video-AI systems but also illuminates the critical importance of aligning proxy rewards with true objectives in MLLM development. | Video MLLMs; Temporal hacking; Temporal Perplexity | Unhackable video-langauge modeling for existing video MLLM. | 814 | 2502.12081 | [
0.011713147163391113,
-0.10312993079423904,
0.020739581435918808,
-0.04291095584630966,
0.06667760759592056,
0.09984315186738968,
-0.022165311500430107,
-0.09022404253482819,
0.13151706755161285,
0.009069370105862617,
-0.0336047038435936,
-0.02672506310045719,
-0.028644846752285957,
0.07655486464500427,
-0.021444056183099747,
-0.052325718104839325,
0.08237132430076599,
-0.0010476864408701658,
-0.09188418090343475,
-0.01730477251112461,
0.02479729615151882,
-0.061496391892433167,
0.029463861137628555,
-0.0002009713207371533,
0.004303412977606058,
0.005961338523775339,
-0.02699320763349533,
-0.09011216461658478,
0.026959050446748734,
-0.031668443232774734,
0.020704304799437523,
-0.020178763195872307,
0.008473030291497707,
-0.026335997506976128,
-0.04210871830582619,
-0.016105569899082184,
-0.0760437399148941,
-0.028980184346437454,
-0.04637172445654869,
-0.07039683312177658,
0.018154948949813843,
-0.06128496676683426,
0.07195807993412018,
-0.02709979936480522,
0.013247434981167316,
-0.04438011720776558,
0.0788128674030304,
-0.06839245557785034,
-0.003446507966145873,
-0.025659339502453804,
-0.07925191521644592,
0.025975121185183525,
0.04529167339205742,
-0.019660621881484985,
-0.06959495693445206,
0.028567412868142128,
0.06556400656700134,
0.10559383034706116,
0.03223905712366104,
-0.018575608730316162,
-0.055117763578891754,
-0.07526180893182755,
-0.005114407744258642,
0.04065002501010895,
0.002099741017445922,
-0.00470865610986948,
0.035638269037008286,
-0.03564485162496567,
0.030662523582577705,
0.035296160727739334,
-0.04608297720551491,
0.037429291754961014,
-0.003384746378287673,
-0.0023951339535415173,
0.04192827269434929,
0.04542051628232002,
-0.04637681692838669,
-0.09366250038146973,
0.05536465346813202,
-0.07583809643983841,
0.03356565535068512,
0.00509156659245491,
-0.006419316399842501,
0.06007860228419304,
0.1320408433675766,
-0.03959713876247406,
0.050779905170202255,
0.03893374651670456,
0.07435091584920883,
0.011157380416989326,
-0.09637494385242462,
-0.023726293817162514,
0.05366542935371399,
0.0077317808754742146,
0.004691471811383963,
0.005206623580306768,
-0.06352833658456802,
-0.05606183037161827,
-0.0659666359424591,
0.02695517987012863,
0.013160661794245243,
0.04974983632564545,
-0.024447165429592133,
-0.004262514412403107,
0.07737068086862564,
0.01602114364504814,
0.05960093438625336,
0.04546727240085602,
0.033751457929611206,
-0.07543614506721497,
-0.0002914242213591933,
0.06304118782281876,
0.10163713246583939,
0.015096611343324184,
0.10024388879537582,
0.0275465976446867,
-0.01110357791185379,
0.002961174352094531,
-0.0032923708204180002,
0.0604841522872448,
0.08210735023021698,
-0.06785465031862259,
0.0259879007935524,
-0.0335797555744648,
0.03311862424015999,
0.007787018548697233,
0.04720233753323555,
3.776251495198857e-34,
0.018923401832580566,
0.0078200101852417,
0.019954022020101547,
0.0405876561999321,
0.07938437163829803,
0.007220221217721701,
0.016965625807642937,
0.039568156003952026,
0.010342108085751534,
0.010282292030751705,
-0.039336077868938446,
0.007635811809450388,
-0.06640778481960297,
0.10637784749269485,
0.07036551833152771,
0.03946333006024361,
-0.053977496922016144,
0.02763952501118183,
-0.0029138834215700626,
-0.021157139912247658,
0.08011829107999802,
-0.06004253402352333,
-0.010336575098335743,
-0.07383576780557632,
0.006597149185836315,
0.01566300168633461,
0.03805489093065262,
-0.05995931476354599,
0.02271420694887638,
-0.007383039221167564,
-0.055481914430856705,
-0.01023886725306511,
0.03478843718767166,
0.02831379324197769,
0.036273982375860214,
0.023538757115602493,
0.03480447456240654,
-0.03955688700079918,
0.02761409804224968,
0.06386926025152206,
-0.045684564858675,
-0.05062028020620346,
-0.06363934278488159,
-0.02322367951273918,
-0.0066436463966965675,
-0.041263215243816376,
0.0840277224779129,
0.023975305259227753,
-0.12711434066295624,
-0.009257428348064423,
0.08716025948524475,
0.09708664566278458,
0.03857582435011864,
-0.14191201329231262,
-0.06361926347017288,
0.0441475510597229,
0.04779617488384247,
-0.008254212327301502,
-0.002256651408970356,
0.0650247186422348,
0.046852853149175644,
-0.026000062003731728,
-0.03502681851387024,
0.05598197877407074,
-0.026466703042387962,
0.033962491899728775,
0.053273048251867294,
0.08305580914020538,
0.03167882561683655,
0.045560434460639954,
-0.014289578422904015,
0.05373483523726463,
0.0153704434633255,
-0.0859496220946312,
0.008455201052129269,
0.037448685616254807,
0.028852954506874084,
-0.058236874639987946,
0.06112426891922951,
0.03365210071206093,
0.008884192444384098,
-0.12291312962770462,
0.004163234960287809,
-0.0669160932302475,
-0.0002088498295051977,
-0.06156264245510101,
0.013844270259141922,
-0.06671631336212158,
0.00915849581360817,
-0.04279901459813118,
-0.042152076959609985,
0.015129451639950275,
0.03819565102458,
0.0399748757481575,
0.02731378749012947,
-9.409027799386067e-34,
-0.004343979060649872,
-0.03154877945780754,
-0.01855473592877388,
0.03270465135574341,
0.04408637434244156,
-0.06527473032474518,
-0.05315296724438667,
-0.039823468774557114,
0.04980706796050072,
-0.028849877417087555,
-0.02128630317747593,
-0.11125541478395462,
-0.04546729847788811,
-0.045943886041641235,
0.021631309762597084,
-0.06230739504098892,
0.018152743577957153,
-0.023778198286890984,
-0.08031722903251648,
-0.0016789521323516965,
0.06608936935663223,
0.053188506513834,
-0.09531402587890625,
-0.029168421402573586,
0.009540232829749584,
0.03959604725241661,
0.0008934233337640762,
0.0788450688123703,
0.014185396023094654,
0.028711026534438133,
0.016911806538701057,
-0.007007407955825329,
-0.00925820879638195,
-0.020526638254523277,
-0.0462176576256752,
0.10753266513347626,
0.08663881570100784,
-0.06771066784858704,
-0.08643685281276703,
0.06735581159591675,
0.0434391051530838,
-0.050682876259088516,
-0.10745280236005783,
0.01572425477206707,
0.01048920676112175,
0.004234942141920328,
-0.07565177977085114,
-0.006603737361729145,
-0.017169950529932976,
-0.014041746035218239,
0.03235682472586632,
0.03972352296113968,
-0.005172803066670895,
-0.024166028946638107,
-0.10187610983848572,
-0.027545001357793808,
-0.07213149219751358,
-0.0014467135770246387,
0.11120647192001343,
0.033745430409908295,
-0.04965152591466904,
0.020944571122527122,
-0.05679396539926529,
-0.009573644958436489,
-0.018457122147083282,
-0.01474829576909542,
0.03439578041434288,
-0.011441836133599281,
-0.024616170674562454,
-0.02751198410987854,
0.0095236049965024,
-0.026858912780880928,
-0.011120384559035301,
-0.018809538334608078,
0.027014432474970818,
0.027794843539595604,
-0.06709253787994385,
-0.02898784913122654,
0.014251968823373318,
-0.08815791457891464,
-0.010410805232822895,
-0.06669507175683975,
0.0837097242474556,
-0.008733655326068401,
-0.0026888037100434303,
0.10269100219011307,
0.05780811980366707,
0.08095898479223251,
0.03647274523973465,
0.002648985246196389,
-0.07153089344501495,
0.00011333009751979262,
-0.030438672751188278,
0.09533244371414185,
0.014777553267776966,
-4.063253555841584e-8,
-0.0742848813533783,
0.07506862282752991,
-0.020276859402656555,
0.01740119606256485,
0.021284321323037148,
0.019751308485865593,
-0.07502114772796631,
0.05133700743317604,
0.040206342935562134,
-0.0351533479988575,
0.09516964852809906,
-0.02244214154779911,
-0.025757383555173874,
-0.03180670738220215,
-0.018689418211579323,
0.0034265548456460238,
0.040315378457307816,
-0.03675386309623718,
0.010583153925836086,
-0.011150711216032505,
0.014482888393104076,
0.0354095995426178,
0.013362925499677658,
-0.1310989260673523,
-0.08972010016441345,
-0.011517209000885487,
-0.03935844451189041,
0.09879206866025925,
0.0861014649271965,
0.040085386484861374,
0.0032303771004080772,
-0.008132918737828732,
0.026141256093978882,
0.012646537274122238,
-0.007028514519333839,
0.016455821692943573,
0.04506315663456917,
-0.05908460170030594,
-0.03517453372478485,
-0.005773056764155626,
0.018974456936120987,
0.03654663637280464,
0.01516862865537405,
-0.0035234775859862566,
-0.10239528864622116,
0.04262617602944374,
-0.07678661495447159,
-0.09200803935527802,
0.03450937196612358,
-0.04229704663157463,
0.043880242854356766,
-0.04892043024301529,
-0.024552812799811363,
0.11326299607753754,
0.12301348894834518,
-0.00003086727156187408,
0.09204156696796417,
-0.058500438928604126,
-0.025188399478793144,
0.06132518872618675,
0.03813832253217697,
0.009886481799185276,
-0.05636461451649666,
-0.049754075706005096
] |
Understanding Long Videos with Multimodal Language Models | https://openreview.net/forum?id=OxKi02I29I | [
"Kanchana Ranasinghe",
"Xiang Li",
"Kumara Kahatapitiya",
"Michael S Ryoo"
] | Poster | Large Language Models (LLMs) have allowed recent LLM-based approaches to achieve excellent performance on long-video understanding benchmarks. We investigate how extensive world knowledge and strong reasoning skills of underlying LLMs influence this strong performance. Surprisingly, we discover that LLM-based approaches can yield surprisingly good accuracy on long-video tasks with limited video information, sometimes even with no video-specific information. Building on this, we explore injecting video-specific information into an LLM-based framework. We utilize off-the-shelf vision tools to extract three object-centric information modalities from videos, and then leverage natural language as a medium for fusing this information. Our resulting Multimodal Video Understanding (MVU) framework demonstrates state-of-the-art performance across multiple video understanding benchmarks. Strong performance also on robotics domain tasks establishes its strong generality. Code: github.com/kahnchana/mvu | long-video, visual question answering, interpretability | Investigates effects of LLM strengths on Long Video QnA tasks. Introduces Multimodal Video Understanding (MVU) framework that incorporates object-centric data from pre-trained models and sets a new state-of-the-art in long-video tasks. | 811 | 2403.16998 | [
0.05388592556118965,
-0.15709832310676575,
0.022285548970103264,
-0.06280861049890518,
0.18891185522079468,
-0.05976463854312897,
-0.04208865016698837,
-0.033304598182439804,
0.010787318460643291,
-0.03490068018436432,
-0.032120052725076675,
-0.05425514653325081,
-0.013761833310127258,
0.12508030235767365,
0.023396093398332596,
0.022482603788375854,
0.060276877135038376,
0.06663520634174347,
-0.13850519061088562,
-0.015818553045392036,
0.04389065131545067,
-0.022282695397734642,
0.09897951036691666,
-0.010789003223180771,
0.05441555008292198,
0.04787180572748184,
0.01595882698893547,
-0.07061808556318283,
0.03847390040755272,
-0.04687976464629173,
0.011274971999228,
-0.01104490552097559,
-0.019158121198415756,
0.04802466928958893,
-0.017799362540245056,
0.07782978564500809,
-0.06534723192453384,
0.0021567828953266144,
-0.05922360345721245,
-0.08352997899055481,
-0.024790238589048386,
0.04336356744170189,
0.09237075597047806,
-0.03594837710261345,
0.06534246355295181,
0.07320021092891693,
0.005689206067472696,
-0.08107902854681015,
0.08131761848926544,
0.020107612013816833,
-0.10664203763008118,
-0.03524298965930939,
0.0063110520131886005,
0.038477830588817596,
-0.061800576746463776,
0.03319362550973892,
0.0409051850438118,
-0.055337581783533096,
0.021513592451810837,
-0.030335763469338417,
0.023986704647541046,
-0.10475832223892212,
0.009306440129876137,
0.02886652946472168,
-0.0005957505200058222,
-0.05665324255824089,
0.03465597704052925,
-0.02324446104466915,
-0.018131433054804802,
0.0014877786161378026,
-0.0472618043422699,
0.03507404774427414,
0.04502522572875023,
0.058194249868392944,
-0.03093819133937359,
-0.0017994195222854614,
0.06006669998168945,
-0.03908996284008026,
0.024440065026283264,
-0.09647458791732788,
0.06981334835290909,
-0.021277256309986115,
0.05321583151817322,
0.03848888725042343,
0.14708441495895386,
-0.009066173806786537,
-0.0024957393761724234,
0.03852946311235428,
-0.03925327956676483,
0.05944677069783211,
-0.05151093006134033,
-0.07398730516433716,
0.050440266728401184,
-0.030434001237154007,
0.041551657021045685,
0.019499873742461205,
0.02207523211836815,
-0.13653093576431274,
-0.09125898778438568,
0.07950980216264725,
-0.022359881550073624,
0.031246721744537354,
-0.005549297202378511,
-0.049029767513275146,
0.022105714306235313,
0.015293334610760212,
0.0361977145075798,
0.04761483892798424,
0.021960224956274033,
-0.011440891772508621,
0.055728401988744736,
0.03703666105866432,
0.04420232027769089,
0.028046773746609688,
0.04110904410481453,
-0.00009031466470332816,
0.00693935202434659,
0.005118913482874632,
-0.001182607258670032,
0.03467177972197533,
-0.027527756989002228,
-0.02392149530351162,
-0.030553242191672325,
-0.011199265718460083,
0.06479097157716751,
0.02230791002511978,
0.025293530896306038,
5.129973978068583e-34,
0.07038126140832901,
0.03815920278429985,
0.06051772087812424,
0.07912906259298325,
0.0749850794672966,
0.01406258624047041,
0.00878024660050869,
0.07589594274759293,
-0.0680595338344574,
0.00881645455956459,
0.019161999225616455,
-0.03374547138810158,
-0.04733417183160782,
0.08051799982786179,
-0.02581303007900715,
0.011557577177882195,
-0.08362751454114914,
-0.025886375457048416,
0.00586675712838769,
0.059028949588537216,
0.049001067876815796,
-0.0071400790475308895,
-0.048531219363212585,
0.03090847097337246,
0.017668141052126884,
-0.0039930823259055614,
0.04754001274704933,
-0.06365424394607544,
-0.00533659290522337,
-0.012478609569370747,
-0.11242295056581497,
-0.02824942022562027,
-0.01973605342209339,
0.021368540823459625,
0.11267771571874619,
-0.06612660735845566,
-0.060420844703912735,
-0.05759284272789955,
0.016485776752233505,
-0.04386330768465996,
-0.023653531447052956,
-0.008218684233725071,
-0.062028393149375916,
-0.05300090089440346,
-0.11978588998317719,
-0.047146379947662354,
0.0038250985089689493,
0.08964965492486954,
-0.08368862420320511,
0.03440583869814873,
0.07800973206758499,
0.04778335988521576,
-0.04845988750457764,
-0.07652973383665085,
-0.05194399878382683,
-0.05452188476920128,
0.07216522097587585,
0.02975516766309738,
-0.003606984158977866,
-0.04237667843699455,
0.02583869732916355,
0.04016194865107536,
-0.017049238085746765,
0.0801684558391571,
0.024606045335531235,
0.022396240383386612,
-0.022010117769241333,
0.10626543313264847,
0.09616517275571823,
-0.0034628259018063545,
-0.0499558299779892,
0.03004351817071438,
0.049060381948947906,
-0.0759410709142685,
0.020756011828780174,
0.034105461090803146,
-0.0028051109984517097,
-0.08843260258436203,
0.08946793526411057,
0.0687783807516098,
0.0066089634783566,
0.018925482407212257,
0.01599576324224472,
-0.01909477822482586,
0.005050542298704386,
-0.012880704365670681,
0.029786527156829834,
-0.1277351826429367,
0.0074791377410292625,
-0.02214902639389038,
-0.02969130501151085,
-0.05060862377285957,
-0.03521699830889702,
-0.023014280945062637,
-0.030393924564123154,
7.290263976053686e-34,
0.03918062523007393,
0.010515205562114716,
-0.0038722301833331585,
0.043180663138628006,
0.026949558407068253,
-0.07620024681091309,
-0.024264005944132805,
0.0664922222495079,
-0.01596272550523281,
-0.09084607660770416,
0.007099363952875137,
-0.05186663940548897,
-0.04041652753949165,
-0.0303939376026392,
0.03984702378511429,
-0.04839788377285004,
0.002082780236378312,
-0.025125548243522644,
-0.03488074615597725,
0.005403496325016022,
0.004056789446622133,
0.05933447182178497,
-0.04157616198062897,
-0.020408401265740395,
-0.13949541747570038,
0.09620465338230133,
-0.06164926663041115,
0.018423214554786682,
-0.04803328588604927,
0.014285841025412083,
0.023911599069833755,
-0.026578310877084732,
-0.011389319784939289,
-0.03869380056858063,
-0.06928576529026031,
0.02691061794757843,
-0.00023989990586414933,
-0.03033680096268654,
-0.008536696434020996,
0.08085530996322632,
0.11436516046524048,
-0.05755161494016647,
-0.057453565299510956,
-0.04801304265856743,
-0.022733625024557114,
-0.032560620456933975,
-0.08028321713209152,
-0.0016567306593060493,
-0.010285223834216595,
0.04999499395489693,
-0.027064554393291473,
0.004555260296911001,
-0.06417842954397202,
-0.011602976359426975,
0.01957681216299534,
-0.09757862240076065,
-0.0701184868812561,
-0.010571581311523914,
0.040697287768125534,
0.003669302212074399,
-0.06031045317649841,
-0.00165320944506675,
-0.03774570673704147,
0.04207591339945793,
0.04160527512431145,
0.025928920134902,
-0.04499591886997223,
-0.029624316841363907,
0.01120209414511919,
0.011752720922231674,
0.06719239801168442,
-0.04355188459157944,
-0.020083505660295486,
0.023265041410923004,
0.010418474674224854,
0.03440549969673157,
0.044447656720876694,
-0.01851137913763523,
0.04698843136429787,
-0.04763300344347954,
0.014820507727563381,
-0.054789792746305466,
0.044481027871370316,
0.08725007623434067,
0.044084079563617706,
0.11058324575424194,
0.031072568148374557,
0.031640030443668365,
0.039329204708337784,
0.05526949092745781,
-0.004325808025896549,
0.04881658777594566,
0.03879494220018387,
0.038972869515419006,
-0.0330222025513649,
-3.1657425125786176e-8,
-0.10539130866527557,
0.036999791860580444,
-0.037278372794389725,
-0.01697486639022827,
-0.014610502868890762,
0.0016633517807349563,
-0.026221025735139847,
0.07107844948768616,
0.029947340488433838,
-0.004778697621077299,
0.11743058264255524,
0.004500483628362417,
-0.018535086885094643,
-0.002005106769502163,
-0.024394817650318146,
0.053416114300489426,
-0.019693972542881966,
0.029866883531212807,
-0.0011443572584539652,
0.03809460252523422,
0.06802038103342056,
-0.01685321144759655,
0.015983834862709045,
-0.007340666372328997,
-0.013341317884624004,
-0.04494423419237137,
-0.02199331484735012,
0.10053103417158127,
0.04598601534962654,
-0.0026662927120923996,
-0.06870491057634354,
0.00013921094068791717,
0.012364055961370468,
-0.04473046213388443,
0.05863646790385246,
-0.019615918397903442,
-0.021220143884420395,
-0.05301399901509285,
0.02078111469745636,
-0.039171576499938965,
0.037990421056747437,
0.07440464198589325,
-0.028484107926487923,
0.06732425838708878,
0.00098605384118855,
0.06656186282634735,
-0.04415324330329895,
-0.0802597776055336,
-0.01550755463540554,
-0.007135726977139711,
-0.06886524707078934,
0.00858320202678442,
-0.023468755185604095,
0.11531593650579453,
0.029889892786741257,
0.0032033990137279034,
0.01986212655901909,
0.014323672279715538,
0.024679839611053467,
0.04121928662061691,
-0.02978142723441124,
0.010103028267621994,
-0.039609476923942566,
0.008992270566523075
] |
Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors | https://openreview.net/forum?id=RoN6NnHjn4 | [
"Haiyu Wu",
"Jaskirat Singh",
"Sicong Tian",
"Liang Zheng",
"Kevin Bowyer"
] | Poster | This paper studies how to synthesize face images of non-existent persons, to create a dataset that allows effective training of face recognition (FR) models. Besides generating realistic face images, two other important goals are: 1) the ability to generate a large number of distinct identities (inter-class separation), and 2) a proper variation in appearance of the images for each identity (intra-class variation).
However, existing works 1) are typically limited in how many well-separated identities can be generated and 2) either neglect or use an external model for attribute augmentation. We propose Vec2Face, a holistic model that uses only a sampled vector as input and can flexibly generate and control the identity of face images and their attributes. Composed of a feature masked autoencoder and an image decoder, Vec2Face is supervised by face image reconstruction and can be conveniently used in inference. Using vectors with low similarity among themselves as inputs, Vec2Face generates well-separated identities. Randomly perturbing an input identity vector within a small range allows Vec2Face to generate faces of the same identity with proper variation in face attributes. It is also possible to generate images with designated attributes by adjusting vector values with a gradient descent method. Vec2Face has efficiently synthesized as many as 300K identities, whereas 60K is the largest number of identities created in the previous works. As for performance, FR models trained with the generated HSFace datasets, from 10k to 300k identities, achieve state-of-the-art accuracy, from 92\% to 93.52\%, on five real-world test sets (\emph{i.e.}, LFW, CFP-FP, AgeDB-30, CALFW, and CPLFW). For the first time, the FR model trained using our synthetic training set achieves higher accuracy than that trained using a same-scale training set of real face images on the CALFW, IJBB, and IJBC test sets. | Identity privacy, Synthetic face dataset generation, Face recognition, Image generation | We proposed Vec2Face, a model for generating scalable synthetic face recognition datasets. The generated HSFace10K achieves SOTA performance and outperforms same-scale real datasets on CALFW, IJBB, and IJBC test sets. | 807 | 2409.02979 | [
-0.11255300790071487,
-0.03332515060901642,
0.038626883178949356,
0.03385976701974869,
0.084272101521492,
-0.02699321322143078,
0.0020830759312957525,
-0.07484162598848343,
-0.047041330486536026,
-0.0656401515007019,
0.019636936485767365,
-0.13429048657417297,
0.04515881463885307,
-0.01893206126987934,
0.040212638676166534,
-0.0749547928571701,
0.05180615559220314,
0.10504605621099472,
-0.05683024972677231,
0.022075526416301727,
-0.03473411127924919,
-0.04655727744102478,
0.039834585040807724,
-0.059054896235466,
0.029895957559347153,
-0.06200564652681351,
0.09057363867759705,
0.0006396237877197564,
0.0025781409349292517,
-0.04306038096547127,
0.09260573238134384,
-0.03304651752114296,
-0.044717807322740555,
-0.013030493631958961,
-0.09540250152349472,
0.038645029067993164,
-0.026499615982174873,
0.17293737828731537,
-0.08419779688119888,
-0.03197053074836731,
-0.10509813576936722,
-0.00885034166276455,
-0.03893658518791199,
0.07470279932022095,
0.096438467502594,
0.03863196820020676,
0.024955620989203453,
-0.03919849917292595,
0.016313815489411354,
-0.04657112807035446,
0.016543207690119743,
-0.047248516231775284,
-0.07517614960670471,
0.013904384337365627,
-0.06003343686461449,
0.008982750587165356,
0.049920156598091125,
-0.04206813871860504,
-0.01270535308867693,
0.0918947383761406,
-0.020162057131528854,
-0.03543899208307266,
0.027191555127501488,
0.016888821497559547,
-0.036854054778814316,
-0.051640018820762634,
0.026746120303869247,
0.013039981946349144,
-0.009831986390054226,
0.012485327199101448,
0.005227910354733467,
-0.01422810833901167,
-0.07968323677778244,
0.01663067564368248,
0.048788852989673615,
0.051484860479831696,
0.002774504479020834,
-0.002516995882615447,
0.07916749268770218,
0.01209834311157465,
-0.0040644374676048756,
0.000702430959790945,
0.02785956673324108,
0.004012965131551027,
0.029740145429968834,
-0.00521220825612545,
-0.04191870614886284,
0.014346419833600521,
-0.12031598389148712,
0.04629594087600708,
-0.06863736361265182,
0.041288625448942184,
0.07646992802619934,
-0.05590667203068733,
0.08677820861339569,
-0.030347660183906555,
0.03181866183876991,
-0.038123007863759995,
-0.00919112004339695,
0.09350690990686417,
-0.06634212285280228,
-0.0030313441529870033,
-0.02112727053463459,
-0.015669232234358788,
0.06074006110429764,
-0.0348069928586483,
0.03217868134379387,
0.005429436918348074,
0.04867870360612869,
-0.07661288976669312,
0.08063819259405136,
0.013618229888379574,
-0.08596044033765793,
-0.027550986036658287,
0.06908351927995682,
0.010772636160254478,
0.04819941520690918,
-0.06303299963474274,
0.02000678889453411,
-0.02899263985455036,
-0.026827428489923477,
0.005170042160898447,
-0.0008026976720429957,
-0.02753034047782421,
0.012985339388251305,
0.0019964808598160744,
-0.032057054340839386,
3.01153479321512e-33,
0.01654883287847042,
0.06536895036697388,
0.003427589312195778,
0.0815887376666069,
0.03055723011493683,
0.001962419366464019,
-0.04227358475327492,
0.024397719651460648,
-0.04603752866387367,
0.06315448880195618,
-0.05216015502810478,
-0.02580909989774227,
-0.052578266710042953,
0.17026396095752716,
0.02388733997941017,
0.042694151401519775,
0.0010765806073322892,
0.021420173346996307,
-0.015230596996843815,
-0.02501387894153595,
0.05627962946891785,
-0.05465460196137428,
0.02838286943733692,
0.045851584523916245,
-0.04639514535665512,
-0.005890405271202326,
0.07606139034032822,
-0.01840401068329811,
0.004834227729588747,
-0.03274605795741081,
-0.034929778426885605,
0.03963197395205498,
0.10452042520046234,
0.036341145634651184,
0.05168032646179199,
-0.03181612491607666,
0.050298817455768585,
-0.0223207026720047,
0.03423115983605385,
0.021026525646448135,
0.08369774371385574,
0.07794933766126633,
0.018621837720274925,
-0.033445242792367935,
-0.03583014756441116,
0.019060291349887848,
0.06652656197547913,
0.08807437121868134,
-0.06607253849506378,
0.0014013045001775026,
0.09924311190843582,
0.046724818646907806,
-0.06141039729118347,
-0.022036993876099586,
-0.09948069602251053,
0.007531528826802969,
0.009121018461883068,
0.04295692965388298,
0.013729178346693516,
-0.03866055980324745,
-0.031086856499314308,
0.02805189974606037,
-0.0043720765970647335,
0.07388295233249664,
-0.03833884000778198,
-0.06067133694887161,
0.04952108860015869,
-0.03458552062511444,
-0.027605367824435234,
0.026486627757549286,
0.034776050597429276,
0.04730333015322685,
-0.04859379678964615,
-0.05714612826704979,
0.025226350873708725,
0.01566714234650135,
0.08102289587259293,
0.007722978480160236,
-0.03463038057088852,
0.13490283489227295,
-0.001806863583624363,
0.1050320714712143,
-0.020831838250160217,
-0.08949585258960724,
-0.04093218222260475,
-0.04015718027949333,
0.002563390415161848,
-0.08390843868255615,
0.0033803596161305904,
0.011276974342763424,
0.009121398441493511,
-0.03845534101128578,
-0.018835455179214478,
-0.014380743727087975,
-0.030312296003103256,
-1.6328296115593143e-33,
0.0026520290412008762,
0.04425543174147606,
-0.03686600178480148,
-0.05528990924358368,
0.03681355342268944,
-0.034143202006816864,
0.06012411415576935,
0.06730511784553528,
0.0024946073535829782,
-0.07109473645687103,
-0.0008825314580462873,
-0.02126978524029255,
0.09209658950567245,
-0.10153020918369293,
0.03841530531644821,
-0.06223951280117035,
-0.046124931424856186,
-0.0038541466929018497,
-0.05484101548790932,
0.016389694064855576,
-0.010071785189211369,
0.06408030539751053,
-0.00699738971889019,
-0.0035109405871480703,
-0.022965379059314728,
0.054497238248586655,
0.009762521833181381,
0.07680046558380127,
-0.013221179135143757,
-0.008818382397294044,
-0.0017431462183594704,
-0.004734159912914038,
0.0042784446850419044,
0.04355977475643158,
-0.09406694769859314,
0.020552117377519608,
-0.0011665003839880228,
-0.012775301933288574,
-0.04326717555522919,
0.0873192846775055,
0.002393910428509116,
-0.009615461342036724,
-0.13034704327583313,
0.053138747811317444,
-0.045580875128507614,
-0.11407819390296936,
-0.014822880737483501,
-0.05557232350111008,
0.03974686935544014,
0.025806566700339317,
-0.04525226354598999,
-0.005304161459207535,
-0.10327093303203583,
0.02753669209778309,
-0.04160911589860916,
-0.09532544761896133,
0.008367993868887424,
0.010066826827824116,
0.09186296164989471,
0.04398009926080704,
-0.0015224464004859328,
-0.07769707590341568,
-0.018227936699986458,
0.010954421944916248,
0.0025353494565933943,
-0.00399470841512084,
-0.03197243809700012,
0.018522698432207108,
-0.028719287365674973,
-0.01738922856748104,
0.07151626795530319,
-0.0704658031463623,
-0.02670038864016533,
0.00920777302235365,
-0.050377171486616135,
-0.05913218855857849,
-0.004012806341052055,
0.028208190575242043,
-0.004031815100461245,
-0.011371717788279057,
-0.029992442578077316,
-0.08784317970275879,
0.03875431790947914,
0.10298815369606018,
0.05487026274204254,
0.010515161789953709,
0.015744231641292572,
0.02267211675643921,
0.02357654459774494,
0.05253869295120239,
-0.12102093547582626,
0.10157214850187302,
-0.005908667575567961,
0.07819870859384537,
0.03927769884467125,
-3.905994816477687e-8,
-0.058825187385082245,
0.022456107661128044,
0.0029773155692964792,
0.03550711274147034,
-0.06130995601415634,
-0.022035300731658936,
0.0023722974583506584,
0.04055406153202057,
-0.0555134117603302,
-0.001587103703059256,
0.07163184881210327,
0.0402187779545784,
0.022447308525443077,
-0.055229347199201584,
0.02333218790590763,
0.04894436523318291,
-0.021856019273400307,
0.10151299834251404,
0.008923175744712353,
-0.022286050021648407,
-0.03115192987024784,
0.038667429238557816,
-0.023898949846625328,
-0.007024009712040424,
0.0706837922334671,
-0.06555501371622086,
-0.07466251403093338,
0.11828960478305817,
0.028282461687922478,
-0.0024771958123892546,
-0.06899819523096085,
-0.0028673233464360237,
0.06954395771026611,
-0.04617195576429367,
0.049970827996730804,
0.04655512049794197,
-0.024434322491288185,
-0.05099708214402199,
-0.05192898213863373,
-0.06354422122240067,
0.026934869587421417,
0.07361618429422379,
-0.023938369005918503,
0.023336995393037796,
0.016633637249469757,
0.013166745193302631,
0.05223521217703819,
-0.05099285766482353,
0.064571812748909,
-0.006078642327338457,
0.019563941285014153,
-0.07092530280351639,
0.002372652990743518,
0.05461642146110535,
0.021978512406349182,
-0.0008926292066462338,
-0.01709757372736931,
0.056540511548519135,
0.08398070931434631,
0.05864691734313965,
-0.03011917509138584,
0.0638720765709877,
-0.01410355418920517,
-0.029508205130696297
] |
Discrete Codebook World Models for Continuous Control | https://openreview.net/forum?id=lfRYzd8ady | [
"Aidan Scannell",
"Mohammadreza Nakhaeinezhadfard",
"Kalle Kujanpää",
"Yi Zhao",
"Kevin Sebastian Luck",
"Arno Solin",
"Joni Pajarinen"
] | Poster | In reinforcement learning (RL), world models serve as internal simulators, enabling agents to predict environment dynamics and future outcomes in order to make informed decisions. While previous approaches leveraging discrete latent spaces, such as DreamerV3, have demonstrated strong performance in discrete action settings and visual control tasks, their comparative performance in state-based continuous control remains underexplored. In contrast, methods with continuous latent spaces, such as TD-MPC2, have shown notable success in state-based continuous control benchmarks. In this paper, we demonstrate that modeling discrete latent states has benefits over continuous latent states and that discrete codebook encodings are more effective representations for continuous control, compared to alternative encodings, such as one-hot and label-based encodings. Based on these insights, we introduce DCWM: Discrete Codebook World Model, a self-supervised world model with a discrete and stochastic latent space, where latent states are codes from a codebook. We combine DCWM with decision-time planning to get our model-based RL algorithm, named DC-MPC: Discrete Codebook Model Predictive Control, which performs competitively against recent state-of-the-art algorithms, including TD-MPC2 and DreamerV3, on continuous control benchmarks. | reinforcement learning, world model, representation learning, self-supervised learning, model-based reinforcement learning, continuous control | World models with discrete codebook encodings are effective for continuous control | 804 | 2503.00653 | [
-0.04882463440299034,
-0.14391176402568817,
-0.0801345556974411,
0.05119369551539421,
0.02906000427901745,
0.04760847985744476,
-0.04094148054718971,
-0.0753827840089798,
0.04303975775837898,
0.05314990133047104,
-0.02220287173986435,
-0.057515908032655716,
0.025768451392650604,
0.046245355159044266,
0.004406125284731388,
0.03061191737651825,
0.10919266194105148,
0.00773655204102397,
-0.06699881702661514,
-0.08974694460630417,
0.07154804468154907,
0.03347563371062279,
0.05118028447031975,
0.0023490211460739374,
0.013956086710095406,
0.058654461055994034,
0.041988663375377655,
-0.0024939551949501038,
0.00734564708545804,
-0.005189698655158281,
0.04679226502776146,
-0.07246687263250351,
-0.01534394919872284,
-0.05653481185436249,
-0.08515416830778122,
0.04122383892536163,
-0.1346665918827057,
-0.02737616002559662,
-0.0833975076675415,
0.004074173979461193,
-0.002371437381953001,
-0.0038388315588235855,
0.011912740767002106,
0.04132454842329025,
0.0176668893545866,
-0.01490036677569151,
-0.001957976259291172,
-0.049497708678245544,
-0.036544885486364365,
-0.04163047671318054,
-0.022248901426792145,
0.005101834423840046,
0.04984753951430321,
-0.05840505659580231,
-0.011301119811832905,
-0.04081391543149948,
0.10943809151649475,
0.031017811968922615,
0.04385712742805481,
-0.04330359399318695,
0.00447487598285079,
-0.04226386919617653,
0.008700213395059109,
-0.02730952762067318,
-0.0585411973297596,
-0.01160044502466917,
-0.030889319255948067,
0.07634743303060532,
-0.01959419809281826,
-0.07459095120429993,
-0.05049649626016617,
0.04272829741239548,
0.0017808106495067477,
-0.030656620860099792,
0.053250860422849655,
-0.07841327041387558,
0.054243460297584534,
-0.04612496495246887,
0.11668480932712555,
-0.015809232369065285,
-0.01083365548402071,
0.051713768392801285,
-0.013533460907638073,
0.040586017072200775,
0.05621873959898949,
0.03420336917042732,
0.017706478014588356,
0.16828419268131256,
0.0270382072776556,
0.002058290410786867,
0.04404702037572861,
-0.003174843965098262,
0.019808899611234665,
0.01419063750654459,
-0.04805118963122368,
-0.02832423523068428,
0.024897579103708267,
-0.07000207155942917,
-0.01637192629277706,
-0.005808401387184858,
-0.0036321228835731745,
0.02796560898423195,
0.014010616578161716,
0.04187433794140816,
0.010573944076895714,
-0.06644752621650696,
0.08045589923858643,
0.0552881620824337,
0.03404367342591286,
-0.08259071409702301,
-0.03557852283120155,
-0.0003078135196119547,
-0.02487679198384285,
0.020706959068775177,
0.0331709198653698,
0.018010182306170464,
0.05443744733929634,
-0.04543264955282211,
0.05447668954730034,
0.05945221334695816,
-0.006205882877111435,
-0.06761008501052856,
0.03203349560499191,
-0.022604437544941902,
0.03298267722129822,
-0.026990219950675964,
0.007817445322871208,
1.4249990879270645e-33,
-0.0069147739559412,
-0.037846341729164124,
0.009367394261062145,
0.020366355776786804,
0.04195559769868851,
-0.010934008285403252,
0.011318094097077847,
0.01592448353767395,
0.013904738239943981,
-0.0299143698066473,
0.017613036558032036,
0.005224269349128008,
-0.06937935948371887,
0.12839269638061523,
0.036861587315797806,
0.03404604643583298,
-0.10133542120456696,
0.00006953528645681217,
-0.053715381771326065,
-0.022273147478699684,
0.0477108433842659,
0.03777911514043808,
-0.00217310874722898,
-0.03121206909418106,
0.03581324592232704,
0.04092617332935333,
0.03568757697939873,
0.02959020808339119,
0.008150597102940083,
0.0049841743893921375,
-0.05972537770867348,
0.009192368015646935,
-0.051077816635370255,
-0.011546116322278976,
0.01295958086848259,
0.07801422476768494,
-0.0739198625087738,
-0.01987340860068798,
-0.03869409114122391,
-0.04138655588030815,
0.016191516071558,
-0.015824826434254646,
0.00011049584281863645,
0.007819029502570629,
-0.04526452347636223,
-0.10907134413719177,
0.023940060287714005,
0.01132856123149395,
-0.1201125904917717,
0.030319809913635254,
0.06948032975196838,
0.06469384580850601,
0.009523073211312294,
-0.11293407529592514,
-0.061922453343868256,
-0.02247997187077999,
0.02773262932896614,
0.02228553406894207,
0.0146559476852417,
0.057115159928798676,
-0.046429235488176346,
-0.009459828957915306,
0.04263228923082352,
0.0871683731675148,
0.035490017384290695,
0.07696320861577988,
-0.01841164194047451,
0.04365014284849167,
0.08290179818868637,
-0.03125831112265587,
-0.008187339641153812,
0.059712037444114685,
0.027945956215262413,
-0.09423258900642395,
0.06882540136575699,
-0.05121971666812897,
0.07943974435329437,
-0.098581463098526,
-0.05887911841273308,
-0.010116206482052803,
-0.028429223224520683,
-0.01699739880859852,
-0.046370092779397964,
-0.03823922947049141,
-0.02399791218340397,
-0.10200701653957367,
-0.016729459166526794,
-0.06150507181882858,
-0.02327699400484562,
0.018626943230628967,
-0.09278402477502823,
-0.037770286202430725,
0.08044175803661346,
0.056779488921165466,
0.01708478480577469,
-3.065471984160616e-33,
0.009028994478285313,
-0.0048404959961771965,
0.004205713514238596,
0.015559278428554535,
0.03456669673323631,
-0.06314660608768463,
0.00022746194736100733,
-0.023200364783406258,
0.04088684916496277,
-0.08840060979127884,
-0.01727328822016716,
-0.009882204234600067,
0.04902457818388939,
0.0802777111530304,
0.04589086398482323,
-0.03888053074479103,
-0.006077587604522705,
-0.01297963596880436,
0.013387217186391354,
0.016548415645956993,
0.036916714161634445,
0.017425674945116043,
-0.08421645313501358,
0.0045071435160934925,
-0.06770103424787521,
0.044142600148916245,
-0.01580737717449665,
0.057961419224739075,
0.029207902029156685,
-0.0010672035859897733,
-0.008647240698337555,
-0.017219895496964455,
-0.05724416300654411,
-0.005446289665997028,
-0.07597217708826065,
0.012534039095044136,
0.04148067906498909,
0.02734772302210331,
-0.09776932001113892,
0.09437112510204315,
0.079440176486969,
-0.08612846583127975,
-0.04550829529762268,
0.030162127688527107,
-0.02418345771729946,
0.020610971376299858,
-0.057254936546087265,
-0.014305918477475643,
-0.0031092616263777018,
0.05984477326273918,
-0.06727014482021332,
0.03882836550474167,
-0.10351671278476715,
0.02128703147172928,
-0.022181233391165733,
-0.03918914869427681,
-0.043186645954847336,
0.036675166338682175,
0.03973767161369324,
0.02666550688445568,
-0.0834740623831749,
0.032334208488464355,
0.07186459749937057,
0.056027863174676895,
-0.03135523945093155,
-0.09869570285081863,
-0.00958152487874031,
0.037187810987234116,
0.058342933654785156,
-0.0767606720328331,
-0.0057134805247187614,
0.035059910267591476,
0.006392785347998142,
0.025207003578543663,
0.0381278395652771,
-0.07287867367267609,
-0.015083396807312965,
-0.007148562930524349,
0.00756319472566247,
0.01133856363594532,
0.0446070097386837,
0.023336483165621758,
0.023333420976996422,
0.004414810799062252,
-0.0026372959837317467,
0.09589673578739166,
-0.004505806136876345,
0.022720756009221077,
0.02163129858672619,
-0.05101194977760315,
-0.009604605846107006,
0.035967689007520676,
0.009785796515643597,
0.04773428663611412,
-0.08085142076015472,
-3.829632433394181e-8,
-0.09018121659755707,
0.02647191844880581,
0.09596298635005951,
-0.03046657145023346,
-0.008663113228976727,
0.006453400012105703,
0.028056515380740166,
-0.022079944610595703,
0.0450306162238121,
0.06418129056692123,
0.1724557876586914,
0.0066174170933663845,
0.0402267687022686,
-0.10448642820119858,
-0.028378669172525406,
0.04685214161872864,
0.03045487590134144,
-0.09713403135538101,
0.08907951414585114,
0.03414400666952133,
0.012258412316441536,
0.004390378016978502,
0.008244644850492477,
-0.01568644680082798,
0.009193710051476955,
-0.03052869252860546,
-0.04861815646290779,
0.1298535019159317,
0.04347234591841698,
0.027537399902939796,
0.08790300041437149,
0.07699579000473022,
0.09690767526626587,
-0.01984921097755432,
-0.03849145770072937,
-0.041358765214681625,
0.0303871501237154,
-0.02858036383986473,
0.049202147871255875,
-0.00472654914483428,
-0.004013689234852791,
0.060431234538555145,
-0.05890569090843201,
-0.04223913699388504,
-0.009759900160133839,
0.01289817038923502,
-0.011852124705910683,
-0.1018369197845459,
0.03350284695625305,
-0.02229105867445469,
-0.0594249852001667,
-0.05110635608434677,
-0.08569417148828506,
0.07682445645332336,
0.1270146667957306,
0.005852757021784782,
0.006048732902854681,
-0.04446633905172348,
0.0458056665956974,
0.0540446899831295,
-0.042595990002155304,
-0.04919802024960518,
-0.10026828199625015,
0.07622730731964111
] |
Connecting Federated ADMM to Bayes | https://openreview.net/forum?id=ipQrjRsl11 | [
"Siddharth Swaroop",
"Mohammad Emtiyaz Khan",
"Finale Doshi-Velez"
] | Poster | We provide new connections between two distinct federated learning approaches based on (i) ADMM and (ii) Variational Bayes (VB), and propose new variants by combining their complementary strengths. Specifically, we show that the dual variables in ADMM naturally emerge through the "site" parameters used in VB with isotropic Gaussian covariances. Using this, we derive two versions of ADMM from VB that use flexible covariances and functional regularisation, respectively. Through numerical experiments, we validate the improvements obtained in performance. The work shows connection between two fields that are believed to be fundamentally different and combines them to improve federated learning. | federated learning, bayesian, variational inference, admm | We provide connections between federated ADMM methods and Variational Bayes federated learning methods, providing ways to improve them | 801 | 2501.17325 | [
0.025536229833960533,
-0.08980252593755722,
0.07483487576246262,
-0.0020780644845217466,
-0.008026782423257828,
0.01811506412923336,
0.0004102041129954159,
-0.0679703801870346,
-0.028315549716353416,
-0.06550319492816925,
0.0007859440520405769,
0.02855631150305271,
0.0305076502263546,
0.013673938810825348,
0.04083552584052086,
0.03580727055668831,
0.07543398439884186,
0.030244924128055573,
-0.05274937301874161,
0.008878677152097225,
-0.045892998576164246,
0.023669278249144554,
-0.06623688340187073,
-0.07651104032993317,
0.0757506787776947,
-0.053894586861133575,
0.057707346975803375,
0.055204976350069046,
0.0044635748490691185,
-0.0506405346095562,
0.05706815421581268,
0.04214267060160637,
-0.08155084401369095,
-0.04470418021082878,
0.0023604410234838724,
0.02244325540959835,
-0.04377957060933113,
0.049478400498628616,
-0.017469540238380432,
0.054809439927339554,
0.028017938137054443,
0.04115965962409973,
0.02978929877281189,
0.042264245450496674,
0.03066067025065422,
-0.01635631173849106,
-0.0115884430706501,
-0.0939006432890892,
0.024744976311922073,
0.014086381532251835,
-0.013356730341911316,
-0.0019595404155552387,
-0.03625939041376114,
0.061757881194353104,
-0.002897905185818672,
0.03191402927041054,
0.02144722267985344,
0.009424805641174316,
-0.026158533990383148,
-0.01563430391252041,
-0.029783787205815315,
-0.08732419461011887,
-0.06881731003522873,
-0.002599961357191205,
0.033936578780412674,
-0.016016075387597084,
0.01322988048195839,
0.03410883992910385,
0.008406216278672218,
0.0063986158929765224,
-0.051862090826034546,
0.09175530821084976,
-0.11566285043954849,
-0.025439202785491943,
-0.01232279185205698,
0.006695969961583614,
-0.006214407738298178,
-0.03133556991815567,
0.12426026910543442,
-0.06146816536784172,
-0.03181331977248192,
0.07554532587528229,
0.0027304943650960922,
-0.08087179064750671,
0.10485503077507019,
-0.06836995482444763,
-0.05283341184258461,
-0.041396260261535645,
0.006007261108607054,
0.004826557822525501,
-0.012525289319455624,
0.0504910834133625,
0.0209018774330616,
-0.0012855733511969447,
-0.023386023938655853,
-0.04872000589966774,
0.04932577162981033,
-0.04589657112956047,
0.07275138050317764,
0.08842041343450546,
-0.034813445061445236,
0.08434342592954636,
-0.021599559113383293,
0.07264409959316254,
0.07346690446138382,
-0.043505627661943436,
0.016639990732073784,
0.03303956240415573,
0.08488994091749191,
-0.061318498104810715,
-0.017599405720829964,
0.0397261381149292,
0.07322561740875244,
-0.02483745664358139,
-0.1291200965642929,
0.007792576681822538,
0.03379097580909729,
0.02118373103439808,
0.07469413429498672,
-0.01040671207010746,
-0.004780604038387537,
-0.07888909429311752,
0.04261210560798645,
-0.004932221956551075,
0.02798224613070488,
-0.07449779659509659,
-0.06128910556435585,
4.9916114971500844e-33,
0.08465484529733658,
-0.014627627097070217,
0.041501060128211975,
-0.016182083636522293,
0.005740175023674965,
0.033086974173784256,
-0.039032675325870514,
-0.05092840641736984,
-0.08889660239219666,
0.01957780122756958,
0.00016215266077779233,
0.10361948609352112,
-0.0512489415705204,
0.09649091213941574,
0.02499445527791977,
-0.05967404693365097,
-0.02748747169971466,
0.06170625984668732,
0.041848186403512955,
0.01089758425951004,
0.07691770792007446,
0.0024174069985747337,
-0.008512262254953384,
0.01139023806899786,
-0.07996693253517151,
-0.05941513180732727,
0.10325392335653305,
0.04672788083553314,
-0.02648279257118702,
0.03461718186736107,
-0.024444954469799995,
-0.042637892067432404,
-0.04905526712536812,
0.09354966133832932,
-0.028550878167152405,
-0.05526528134942055,
0.0014445208944380283,
-0.05621929094195366,
-0.0049131098203361034,
-0.06829805672168732,
0.07231761515140533,
-0.04398906230926514,
0.1416042000055313,
-0.08272523432970047,
-0.12474735826253891,
-0.008546030148863792,
-0.014300325885415077,
-0.003353850217536092,
0.09602666646242142,
-0.08717318624258041,
-0.0021576285362243652,
-0.060248058289289474,
-0.03418906033039093,
0.0078465286642313,
0.009725121781229973,
0.05072600767016411,
-0.09379662573337555,
0.028014805167913437,
-0.017190631479024887,
-0.01010227296501398,
-0.022281505167484283,
0.0812571570277214,
0.01915515772998333,
0.010089137591421604,
-0.04249429330229759,
0.09733530133962631,
0.05706210806965828,
-0.02368728257715702,
0.09406859427690506,
-0.041753582656383514,
0.01116909645497799,
0.06189261004328728,
-0.00919421948492527,
-0.08134342730045319,
0.07316461950540543,
-0.03869205340743065,
0.07926201075315475,
-0.08789335191249847,
-0.005804740823805332,
-0.00737051572650671,
-0.10966383665800095,
0.06528416275978088,
-0.04809877276420593,
0.0765354186296463,
-0.03509045019745827,
-0.010983253829181194,
-0.05608075112104416,
-0.023797448724508286,
-0.03684153035283089,
-0.004949445370584726,
-0.05505021661520004,
0.014438691549003124,
-0.04621756821870804,
0.05530111491680145,
0.003251677146181464,
-3.9101530209737336e-33,
-0.14129896461963654,
0.015953540802001953,
0.0026912884786725044,
-0.032455574721097946,
-0.004444606602191925,
-0.04059882462024689,
-0.016894258558750153,
0.0032354979775846004,
0.0010992641327902675,
0.0036952258087694645,
-0.04477832093834877,
-0.00880808662623167,
-0.03207135945558548,
-0.006443724036216736,
-0.046861741691827774,
0.03459837660193443,
-0.044372815638780594,
0.023814136162400246,
0.03700815886259079,
-0.009981677867472172,
0.058482661843299866,
0.07046880573034286,
-0.019175292924046516,
0.031751327216625214,
0.12158891558647156,
-0.0013799701118841767,
-0.04063049703836441,
0.1118469163775444,
-0.0044790226966142654,
0.01600254513323307,
0.025426626205444336,
-0.02286146581172943,
-0.07986628264188766,
-0.036967139691114426,
-0.054762665182352066,
0.038598913699388504,
-0.01635938324034214,
0.02446001023054123,
-0.05719143897294998,
0.01548765692859888,
-0.05608106032013893,
-0.04544282704591751,
-0.06688465923070908,
-0.0394725538790226,
0.12804579734802246,
-0.011522703804075718,
0.053139183670282364,
-0.051134251058101654,
0.03864701837301254,
-0.0179840587079525,
-0.019521161913871765,
0.0700392797589302,
-0.051193904131650925,
0.035841599106788635,
-0.13468749821186066,
-0.023107726126909256,
0.05191703513264656,
-0.05816706642508507,
0.07404907792806625,
0.014620671980082989,
0.030431102961301804,
-0.00038380263140425086,
0.02033453993499279,
0.03549510985612869,
0.008908616378903389,
0.03994956240057945,
0.020464632660150528,
0.00019497095490805805,
-0.06458672881126404,
0.03462789207696915,
-0.0029646949842572212,
0.05243493244051933,
-0.0020546820014715195,
0.02839232049882412,
0.10375772416591644,
0.023877901956439018,
0.020416414365172386,
-0.031762998551130295,
-0.05632400885224342,
0.021810173988342285,
0.029280833899974823,
-0.020505830645561218,
0.061709798872470856,
0.07237768173217773,
0.00461162393912673,
0.08473952114582062,
0.00015511506353504956,
0.04127255827188492,
0.05148382857441902,
-0.01008959673345089,
-0.018200786784291267,
0.0017255189595744014,
-0.006147302221506834,
0.02932818979024887,
-0.08823921531438828,
-4.080440163534149e-8,
-0.03924371302127838,
0.07291248440742493,
0.043724119663238525,
-0.038593556731939316,
0.029301831498742104,
-0.006046633701771498,
-0.02276214025914669,
0.038411613553762436,
-0.03171228989958763,
-0.07937479019165039,
-0.03546001389622688,
0.010462022386491299,
-0.017995275557041168,
-0.04940195754170418,
0.011385559104382992,
0.0946083813905716,
0.026854882016777992,
0.017290227115154266,
-0.018046805635094643,
0.03721744194626808,
0.08320633322000504,
-0.012785825878381729,
0.025856196880340576,
-0.0035720490850508213,
-0.027099108323454857,
-0.08885893225669861,
0.02339051477611065,
0.11157595366239548,
0.06183590367436409,
0.04558626562356949,
-0.042296551167964935,
0.05478152260184288,
-0.05415510758757591,
0.014265543781220913,
-0.05781934782862663,
0.08465549349784851,
0.04435703158378601,
-0.08910415321588516,
-0.0538320355117321,
0.004578387830406427,
-0.02419017255306244,
-0.05519649013876915,
-0.019712716341018677,
0.04226713255047798,
0.06525658071041107,
0.052198149263858795,
0.001387518597766757,
-0.09120700508356094,
0.1040038987994194,
-0.008522038348019123,
0.060419950634241104,
-0.006561633199453354,
-0.031048936769366264,
0.043723106384277344,
0.07794930040836334,
-0.013492504134774208,
0.004427136387676001,
-0.03843583166599274,
0.023257149383425713,
0.046354394406080246,
0.0015454440144822001,
0.016582781448960304,
-0.05103900283575058,
-0.019906720146536827
] |
Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets | https://openreview.net/forum?id=yTEwmr1TJb | [
"Guangqi Jiang",
"Yifei Sun",
"Tao Huang",
"Huanyu Li",
"Yongyuan Liang",
"Huazhe Xu"
] | Poster | The pre-training of visual representations has enhanced the efficiency of robot learning. Due to the lack of large-scale in-domain robotic datasets, prior works utilize in-the-wild human videos to pre-train robotic visual representation. Despite their promising results, representations from human videos are inevitably subject to distribution shifts and lack the dynamics information crucial for task completion. We first evaluate various pre-trained representations in terms of their correlation to the downstream robotic manipulation tasks (i.e., manipulation centricity). Interestingly, we find that the ''manipulation centricity'' is a strong indicator of success rates when applied to downstream tasks. Drawing from these findings, we propose **M**anipulation **C**entric **R**epresentation (**MCR**), a foundation representation learning framework capturing both visual features and the dynamics information such as actions and proprioceptions of manipulation tasks to improve manipulation centricity. Specifically, we pre-train a visual encoder on the DROID robotic dataset and leverage motion-relevant data such as robot proprioceptive states and actions. We introduce a novel contrastive loss that aligns visual observations with the robot's proprioceptive state-action dynamics, combined with an action prediction loss and a time contrastive loss during pre-training. Empirical results across four simulation domains with 20 robotic manipulation tasks demonstrate that **MCR** outperforms the strongest baseline by 14.8\%. Additionally, **MCR** significantly boosts the success rate in three real-world manipulation tasks by 76.9\%. Project website: robots-pretrain-robots.github.io | Robot Learning, Foundation Model, Representation Learning | Manipulation-centric robotic representation training on large-scale robot dataset boosts policy performance on manipulation tasks. | 793 | 2410.22325 | [
-0.006541617680341005,
-0.17387692630290985,
0.08198503404855728,
-0.035588253289461136,
0.06277500838041306,
0.07748439162969589,
-0.0406191311776638,
-0.04426753893494606,
0.019981980323791504,
0.005250435788184404,
0.011565649881958961,
-0.058139536529779434,
-0.029798800125718117,
0.11317459493875504,
-0.011836211197078228,
0.007699710316956043,
0.12591633200645447,
0.07981458306312561,
-0.10435261577367783,
0.030223991721868515,
0.05174586549401283,
-0.04261808097362518,
0.10335078835487366,
0.06318584829568863,
0.03532533720135689,
0.02683362178504467,
0.04019436240196228,
-0.06877408921718597,
0.003865302074700594,
-0.09471297264099121,
-0.008701189421117306,
-0.02620270475745201,
-0.00055420066928491,
-0.023863980546593666,
-0.11139120906591415,
0.09372824430465698,
-0.03925453871488571,
0.03231015428900719,
-0.008909893222153187,
-0.08562470227479935,
-0.016197551041841507,
0.019859978929162025,
0.09578701108694077,
0.010794584639370441,
0.027102189138531685,
0.07331474870443344,
0.007038561161607504,
-0.08226490020751953,
0.045634739100933075,
0.02428390458226204,
-0.030944617465138435,
-0.0019209347665309906,
0.030180541798472404,
0.015201488509774208,
0.013056698255240917,
0.07339136302471161,
0.03447546809911728,
-0.05829023942351341,
0.06594985723495483,
0.000863793131429702,
0.04342978820204735,
-0.05277169495820999,
0.02406655251979828,
-0.009643045254051685,
-0.061619799584150314,
-0.05857405066490173,
0.07148836553096771,
-0.010983189567923546,
-0.010981866158545017,
0.009183811023831367,
-0.015226908028125763,
0.011733698658645153,
-0.02087034098803997,
0.037944402545690536,
0.05647925287485123,
0.006161544006317854,
0.006092325784265995,
0.012363641522824764,
0.0060012927278876305,
-0.05695435404777527,
0.06531283259391785,
0.031718507409095764,
0.054998770356178284,
0.037181735038757324,
0.05820897966623306,
0.0026192038785666227,
0.04337310418486595,
0.07058533281087875,
0.0037549270782619715,
0.03605924919247627,
-0.025225166231393814,
-0.025737261399626732,
-0.029901161789894104,
-0.03293502330780029,
-0.0066442666575312614,
-0.021634135395288467,
-0.042072996497154236,
-0.12031907588243484,
0.01128251664340496,
-0.008520313538610935,
0.02811448834836483,
-0.0033174350392073393,
-0.022558625787496567,
0.046414393931627274,
0.03180272504687309,
-0.08007687330245972,
0.03675948455929756,
0.0488789826631546,
0.008637474849820137,
0.01201129611581564,
0.07819907367229462,
-0.06286688148975372,
0.018740136176347733,
0.07520189881324768,
0.03179825842380524,
0.009704322554171085,
-0.04012995958328247,
0.004929839633405209,
-0.014908882789313793,
0.032261453568935394,
0.027585716918110847,
-0.013144835829734802,
-0.014441042207181454,
0.005059063900262117,
0.012353961355984211,
-0.10699041932821274,
-0.03086010180413723,
-3.0257478056135077e-33,
0.05564901977777481,
-0.02714614011347294,
0.07695604115724564,
0.003509111236780882,
-0.03271222487092018,
-0.03194311633706093,
-0.01897643506526947,
0.054837580770254135,
-0.029546579346060753,
-0.031577542424201965,
-0.02429787814617157,
0.013878541067242622,
-0.03899798169732094,
0.13512536883354187,
-0.04675041139125824,
-0.03795181214809418,
0.03413889929652214,
0.04270535334944725,
-0.0686950534582138,
0.031424473971128464,
-0.04576629027724266,
0.03738871216773987,
-0.05350254848599434,
-0.06262753903865814,
0.019265824928879738,
-0.02988266386091709,
0.0010893220314756036,
-0.02192513644695282,
0.015861475840210915,
-0.021347420290112495,
-0.10920415073633194,
0.05596412718296051,
-0.029515612870454788,
0.01196658331900835,
-0.014892068691551685,
0.008779617957770824,
0.009174496866762638,
-0.0678848922252655,
0.016956742852926254,
0.001710789860226214,
-0.01850372739136219,
0.0742785781621933,
-0.07544754445552826,
-0.056361835449934006,
-0.0782899409532547,
-0.11400002241134644,
0.0773155614733696,
0.06788183748722076,
-0.10172159969806671,
-0.07135032117366791,
0.05605665594339371,
0.05751670151948929,
0.006224608514457941,
-0.08500441908836365,
-0.012809704057872295,
0.012407843954861164,
0.05520837754011154,
0.051636774092912674,
0.002007716102525592,
-0.0315793938934803,
0.0033899995032697916,
0.0315869078040123,
-0.0255740974098444,
0.10013895481824875,
-0.04048297181725502,
0.014039585366845131,
-0.001028373371809721,
0.08269620686769485,
0.08913065493106842,
0.0305626317858696,
-0.0902591124176979,
0.02181890234351158,
-0.019825704395771027,
-0.09327259659767151,
0.06596987694501877,
0.0304404404014349,
0.004920127801597118,
-0.08685553818941116,
-0.012496840208768845,
0.07631707191467285,
-0.052208129316568375,
-0.014647587202489376,
-0.02116156928241253,
0.01566222496330738,
0.022441120818257332,
-0.03005511313676834,
0.042357563972473145,
-0.09490464627742767,
0.025982709601521492,
-0.005427999887615442,
-0.015434935688972473,
-0.11190656572580338,
-0.057735685259103775,
0.04784990847110748,
-0.025196319445967674,
2.141564832153984e-33,
0.027755141258239746,
0.049037959426641464,
-0.014731727540493011,
0.05214904993772507,
-0.03263363987207413,
-0.05011306330561638,
0.02241523005068302,
-0.0067615327425301075,
-0.013055712915956974,
-0.08342428505420685,
-0.04772992432117462,
-0.10289394110441208,
-0.0701780840754509,
-0.05251845717430115,
0.029520832002162933,
0.007150579243898392,
0.0019373504910618067,
0.04305536672472954,
-0.03525279462337494,
-0.02445369027554989,
0.014332082122564316,
0.056299835443496704,
0.007675008848309517,
-0.004536591470241547,
-0.06738416105508804,
0.09736593812704086,
0.017749572172760963,
0.04734969139099121,
0.013014620169997215,
0.008300963789224625,
-0.02900596335530281,
-0.05698791891336441,
-0.018681185320019722,
0.010498539544641972,
-0.07212016731500626,
0.05800481140613556,
-0.019759206101298332,
0.015008200891315937,
-0.021427186205983162,
0.05716501921415329,
0.059809282422065735,
-0.02421754226088524,
-0.037604670971632004,
0.054555997252464294,
-0.01317804865539074,
-0.0931529775261879,
-0.007512871641665697,
0.05008123815059662,
0.016423532739281654,
0.04614844173192978,
-0.09195961803197861,
0.1216871440410614,
-0.07559437304735184,
-0.04570915177464485,
-0.039135996252298355,
-0.0686483159661293,
0.010178575292229652,
-0.031626082956790924,
0.11147809028625488,
-0.02232392691075802,
-0.08129239827394485,
0.04069267958402634,
0.009893686510622501,
0.03302201256155968,
-0.0070792618207633495,
0.015313641168177128,
-0.03422956168651581,
0.016183001920580864,
0.07764918357133865,
-0.022692494094371796,
0.015704724937677383,
0.011304497718811035,
0.058558620512485504,
-0.03718256577849388,
0.09492293745279312,
0.016006287187337875,
0.035984158515930176,
-0.10039196163415909,
0.015900354832410812,
0.005030391272157431,
0.006859473884105682,
-0.03312857449054718,
0.04195018857717514,
0.09817873686552048,
-0.007578760851174593,
0.1484098583459854,
-0.005976950284093618,
0.004733486566692591,
0.04584319517016411,
0.016841748729348183,
0.0055237640626728535,
-0.01275397278368473,
-0.015101630240678787,
0.057435691356658936,
-0.01842672750353813,
-3.4383287328410006e-8,
-0.09481000155210495,
0.02197512984275818,
0.05638086795806885,
-0.003398434491828084,
-0.0004833129933103919,
0.03702051192522049,
0.01503799855709076,
0.08821973949670792,
-0.0770026296377182,
-0.02916938252747059,
0.09057201445102692,
-0.032537512481212616,
-0.013914926908910275,
-0.05985724925994873,
0.009491966105997562,
0.02188696525990963,
0.05824961140751839,
0.03240714967250824,
-0.0021716165356338024,
0.08408553153276443,
0.023349184542894363,
-0.02261626347899437,
0.04753583297133446,
-0.07470668107271194,
-0.004562480840831995,
-0.043902602046728134,
-0.09133084863424301,
0.08542772382497787,
0.00931694358587265,
-0.013308031484484673,
-0.00006884024332975969,
-0.038856152445077896,
0.07635367661714554,
-0.011315744370222092,
0.014781511388719082,
0.01015748642385006,
0.02505994215607643,
-0.00764017878100276,
-0.06687167286872864,
-0.012818370945751667,
0.04035695269703865,
0.09443219006061554,
0.006493748631328344,
0.029711978510022163,
-0.06257767230272293,
0.0016537277260795236,
-0.01872749626636505,
-0.06258488446474075,
0.03058624640107155,
-0.06307771801948547,
-0.04092071205377579,
-0.012186817824840546,
-0.009407115168869495,
0.1258062720298767,
0.08967650681734085,
0.05180203542113304,
0.013345284387469292,
-0.10017108917236328,
-0.01960606314241886,
0.04969555139541626,
0.006575542502105236,
0.031818240880966187,
-0.0036663806531578302,
-0.046431224793195724
] |
TimeSuite: Improving MLLMs for Long Video Understanding via Grounded Tuning | https://openreview.net/forum?id=nAVejJURqZ | [
"Xiangyu Zeng",
"Kunchang Li",
"Chenting Wang",
"Xinhao Li",
"Tianxiang Jiang",
"Ziang Yan",
"Songze Li",
"Yansong Shi",
"Zhengrong Yue",
"Yi Wang",
"Yali Wang",
"Yu Qiao",
"Limin Wang"
] | Poster | Multimodal Large Language Models (MLLMs) have demonstrated impressive performance in short video understanding. However, understanding long-form videos still remains challenging for MLLMs. This paper proposes TimeSuite, a collection of new designs to adapt the existing short-form video MLLMs for long video understanding, including a simple yet efficient framework to process long video sequence, a high-quality video dataset for grounded tuning of MLLMs, and a carefully-designed instruction tuning task to explicitly incorporate the grounding supervision in the traditional QA format. Specifically, based on VideoChat, we propose our long-video MLLM, coined as VideoChat-T, by implementing a token shuffling to compress long video tokens and introducing Temporal Adaptive Position Encoding (TAPE) to enhance the temporal awareness of visual representation. Meanwhile, we introduce the TimePro, a comprehensive grounding-centric instruction tuning dataset composed of 9 tasks and 349k high-quality grounded annotations. Notably, we design a new instruction tuning task type, called Temporal Grounded Caption, to peform detailed video descriptions with the corresponding time stamps prediction. This explicit temporal location prediction will guide MLLM to correctly attend on the visual content when generating description, and thus reduce the hallucination risk caused by the LLMs. Experimental results demonstrate that our TimeSuite provides a successful solution to enhance the long video understanding capability of short-form MLLM, achieving improvement of 5.6% and 6.8% on the benchmarks of Egoschema and VideoMME, respectively. In addition, VideoChat-T exhibits robust zero-shot temporal grounding capabilities, significantly outperforming the existing state-of-the-art MLLMs. After fine-tuning, it performs on par with the traditional supervised expert models. | Long Video Understanding; Temporal Grounding; Multimodal Large Language Model | null | 773 | 2410.19702 | [
0.00927787460386753,
-0.1105949878692627,
0.05706598982214928,
-0.09676791727542877,
0.045307062566280365,
0.017425306141376495,
-0.03303046524524689,
0.010693646967411041,
0.06963120400905609,
-0.07919352501630783,
-0.027720337733626366,
-0.054478224366903305,
-0.05552130192518234,
0.0867386981844902,
0.007672710344195366,
-0.058178674429655075,
0.04787536337971687,
0.007564635016024113,
-0.0977102518081665,
-0.0358186811208725,
0.08729071170091629,
-0.016238711774349213,
0.054746367037296295,
0.02085806056857109,
-0.005612274166196585,
-0.007209596689790487,
-0.03926525637507439,
-0.03289076313376427,
0.08393339812755585,
-0.06038569658994675,
-0.009174850769340992,
0.0005166861228644848,
0.07328446209430695,
0.09577788412570953,
-0.07470819354057312,
0.07524235546588898,
0.014688663184642792,
-0.0004532189923338592,
-0.014675676822662354,
-0.04817519336938858,
-0.0206534992903471,
0.006440365221351385,
0.09763819724321365,
0.04449911788105965,
0.07663594186306,
0.03994382545351982,
0.028966471552848816,
-0.14269804954528809,
-0.02286207117140293,
0.019503870978951454,
-0.04506116732954979,
-0.01783916912972927,
0.05389745905995369,
0.019386285915970802,
-0.027368612587451935,
0.0682632103562355,
0.06786252558231354,
0.04002167657017708,
0.06068313866853714,
0.023514706641435623,
-0.04053504019975662,
-0.05925537273287773,
0.01170719787478447,
-0.013494648039340973,
0.01331518404185772,
-0.06330405175685883,
0.04412781447172165,
0.005331336986273527,
0.009161518886685371,
-0.016359549015760422,
-0.11681223660707474,
-0.014371433295309544,
-0.04491972550749779,
0.055778536945581436,
-0.05022670701146126,
0.07812345027923584,
0.004516527522355318,
-0.014000550843775272,
-0.004061794839799404,
-0.13112172484397888,
0.09752096235752106,
-0.02574341930449009,
0.0547301284968853,
0.06072289124131203,
0.05098726227879524,
-0.024630894884467125,
0.06565118581056595,
0.03400466963648796,
-0.015572148375213146,
-0.02830241434276104,
-0.11999491602182388,
-0.05477195978164673,
0.02674432471394539,
-0.03941427171230316,
0.027606302872300148,
0.038871947675943375,
0.014497100375592709,
-0.10332180559635162,
-0.028248362243175507,
0.055956386029720306,
0.03148428350687027,
0.03688963130116463,
-0.02688879333436489,
0.004229255951941013,
0.019461356103420258,
-0.02133914828300476,
0.061124037951231,
0.06992940604686737,
0.005144661292433739,
-0.04384685307741165,
0.03105783462524414,
0.026054751127958298,
-0.002214762382209301,
-0.012953351251780987,
0.03875919431447983,
0.01928890496492386,
-0.014432313852012157,
-0.019350718706846237,
0.016776803880929947,
0.033891621977090836,
0.05369780212640762,
0.0004904282395727932,
-0.03190748766064644,
-0.047309815883636475,
0.04246700182557106,
-0.05414624884724617,
0.05548820644617081,
1.0383334455355197e-33,
0.0625183954834938,
0.06964734941720963,
0.03291197121143341,
0.008984037674963474,
0.08846885710954666,
-0.004370272625237703,
-0.009228278882801533,
0.06939297914505005,
-0.034007154405117035,
0.0009482012246735394,
0.04948490485548973,
-0.09032168984413147,
-0.051342010498046875,
0.0843224972486496,
-0.0203787200152874,
0.02620309591293335,
-0.05961312726140022,
-0.03290902078151703,
-0.02116517722606659,
-0.02376672253012657,
0.07376094162464142,
-0.015371911227703094,
-0.010080007836222649,
0.0007389067905023694,
0.09086550772190094,
-0.002909629372879863,
0.07700379192829132,
-0.06197143346071243,
-0.012065855786204338,
-0.029679322615265846,
-0.1414874643087387,
0.0093788867816329,
-0.020081745460629463,
-0.01962793432176113,
0.10382331162691116,
-0.0691535621881485,
-0.014214559458196163,
-0.03394657373428345,
0.024977849796414375,
-0.0020999531261622906,
0.015034612268209457,
-0.010133115574717522,
-0.0490063801407814,
-0.02901182882487774,
-0.045628853142261505,
-0.07714247703552246,
0.016130631789565086,
0.05052986741065979,
-0.06123530492186546,
0.03998386859893799,
0.0225367434322834,
0.043162401765584946,
-0.007648620288819075,
-0.08392351865768433,
0.008523802272975445,
-0.03266718238592148,
0.07079759985208511,
0.03595757856965065,
-0.0099055590108037,
0.006709256209433079,
0.019946128129959106,
0.06555209308862686,
-0.002072604838758707,
0.04822487756609917,
-0.0014545113081112504,
-0.021736102178692818,
0.01566154509782791,
0.13232649862766266,
0.08128195255994797,
0.022270403802394867,
-0.0324440561234951,
0.01652846671640873,
0.0843958631157875,
-0.012446838431060314,
-0.03787920996546745,
0.0009831703500822186,
0.04634249955415726,
-0.05254904553294182,
0.013883091509342194,
0.033866435289382935,
0.047882240265607834,
-0.06456314027309418,
0.019654106348752975,
-0.04371955245733261,
0.022863810881972313,
-0.13912740349769592,
0.05474751070141792,
-0.08775541186332703,
-0.0173728559166193,
-0.06575339287519455,
-0.06291117519140244,
-0.04025164619088173,
0.0018577422015368938,
-0.009754769504070282,
-0.010147427208721638,
-2.496444361135888e-35,
0.06877248734235764,
0.04649290814995766,
-0.06076973304152489,
0.02034665457904339,
0.04660683870315552,
-0.09790697693824768,
0.02786928415298462,
0.0718013271689415,
0.02165450155735016,
-0.014248336665332317,
-0.08362133800983429,
-0.06589852273464203,
-0.05169796198606491,
0.003108561271801591,
-0.00840746983885765,
0.0061060176230967045,
-0.0370287150144577,
-0.00736901443451643,
-0.017541320994496346,
0.01205967552959919,
0.0193406343460083,
0.03988314047455788,
-0.08568912744522095,
0.03404567763209343,
0.03886040672659874,
0.03931909054517746,
0.04335373267531395,
0.0836801677942276,
-0.010912520810961723,
-0.019665515050292015,
0.007455749902874231,
-0.04414042830467224,
0.059877909719944,
-0.04653973877429962,
-0.041921526193618774,
0.012253236956894398,
0.07319173961877823,
-0.021129878237843513,
-0.024113815277814865,
0.0483110174536705,
0.11679273843765259,
-0.019631410017609596,
-0.010496190749108791,
-0.04407253488898277,
-0.009293890558183193,
0.002358626341447234,
-0.12395565956830978,
-0.02381022833287716,
-0.014717396348714828,
0.04101491719484329,
-0.019045554101467133,
0.03796729817986488,
-0.079123355448246,
-0.000875400030054152,
-0.015759168192744255,
-0.09533939510583878,
-0.05440283566713333,
0.04144814983010292,
0.060558516532182693,
-0.009882859885692596,
-0.028761480003595352,
-0.054897282272577286,
0.04133930802345276,
-0.05039731785655022,
0.06630711257457733,
-0.029379326850175858,
-0.007944806478917599,
-0.03397227078676224,
-0.02279215306043625,
-0.021421262994408607,
0.04338520020246506,
-0.07436928898096085,
-0.04012481868267059,
0.025231637060642242,
-0.014734744094312191,
0.014551489613950253,
0.02069849520921707,
0.010044821538031101,
-0.006813155021518469,
-0.049145374447107315,
-0.008236275985836983,
-0.075105220079422,
0.012101391330361366,
0.09501656144857407,
0.05059457942843437,
0.141859769821167,
0.0338149331510067,
0.14201922714710236,
-0.01097293570637703,
0.02583419345319271,
-0.04850419983267784,
0.06617970019578934,
0.03806570917367935,
0.08529292792081833,
-0.02021842636168003,
-3.685726213120688e-8,
-0.0358012393116951,
0.011243348009884357,
-0.029062751680612564,
0.04048904404044151,
0.0020537306554615498,
0.028205543756484985,
-0.010189973749220371,
0.010207947343587875,
0.07393743842840195,
-0.0458553247153759,
0.03787573426961899,
0.0024621067568659782,
-0.011059243232011795,
-0.04599553719162941,
0.0009230914292857051,
0.0858517438173294,
-0.028446028009057045,
0.007063515949994326,
-0.013504278846085072,
0.03195301070809364,
0.024233423173427582,
0.0263666994869709,
-0.005388086196035147,
-0.06465364247560501,
-0.10359688103199005,
-0.011300279758870602,
0.028770674020051956,
0.16487948596477509,
0.043989431113004684,
-0.06468411535024643,
0.035994984209537506,
0.04724142700433731,
0.01258925348520279,
-0.12580183148384094,
-0.04183214157819748,
-0.0297373179346323,
0.0351044237613678,
-0.025031698867678642,
0.03495234623551369,
0.0012143142521381378,
-0.00527219008654356,
0.014274250715970993,
-0.04573332518339157,
0.005495603196322918,
-0.021356284618377686,
0.00320832384750247,
-0.06882166117429733,
-0.10439692437648773,
0.001882052980363369,
0.00810026191174984,
-0.025181634351611137,
-0.016553614288568497,
-0.008403606712818146,
0.08423411101102829,
0.10671904683113098,
0.051523931324481964,
0.0486266054213047,
-0.018579306080937386,
0.030571987852454185,
0.09448480606079102,
-0.05643323436379433,
0.02449803613126278,
0.012464016675949097,
0.0013018520548939705
] |
CLIPure: Purification in Latent Space via CLIP for Adversarially Robust Zero-Shot Classification | https://openreview.net/forum?id=TQ2ZOy6miT | [
"Mingkun Zhang",
"Keping Bi",
"Wei Chen",
"Jiafeng Guo",
"Xueqi Cheng"
] | Poster | In this paper, we aim to build an adversarially robust zero-shot image classifier that can accurately and efficiently classify unseen examples while defending against unforeseen adversarial attacks, addressing critical challenges in real-world safety-sensitive scenarios. To achieve this, we focus on two key challenges: zero-shot classification and defense against unforeseen attacks. We ground our work on CLIP, a vision-language pre-trained model to perform zero-shot classification.
To defend against unforeseen attacks, we adopt a purification approach, as it is independent of specific attack types.
We then define a purification risk as the KL divergence between the joint distributions of the purification and attack process.
The derived lower bound of purification risk inspires us to explore purification in CLIP's multi-modal latent space.
We propose a CLIP-based purification method called CLIPure, which has two variants: _CLIPure-Diff_, which models image likelihood with a generative process of its latent vector, and _CLIPure-Cos_, which models the likelihood based on the similarity between embeddings of the image and a blank template. As far as we know, CLIPure is the first purification method in latent space and _CLIPure-Cos_ is the first purification method not relying on generative models, substantially improving defense efficiency. Extensive experimental results show that the robustness achieved by CLIPure is within a small gap of clean accuracy, outperforming SOTA robustness by a large margin, e.g., from 71.7\% to **91.1\%** on CIFAR10, from 59.6\% to **72.6\%** on ImageNet, and **108\%** relative improvements of average robustness on the 13 datasets over previous SOTA, with only 14\% extra inference cost and no additional training. | adversarial robustness, purification, CLIP | null | 772 | 2502.18176 | [
-0.0654301717877388,
-0.059490494430065155,
0.022119048982858658,
0.0031120143830776215,
0.13234658539295197,
-0.0007573905168101192,
0.05593055114150047,
-0.08406153321266174,
0.08577442169189453,
-0.052828967571258545,
0.06293360888957977,
-0.06854849308729172,
0.04456258937716484,
0.03096429444849491,
-0.04893039911985397,
-0.03790911287069321,
0.04958343505859375,
0.10230625420808792,
-0.12342598289251328,
0.004372099880129099,
0.04206731915473938,
-0.007285117171704769,
0.029536820948123932,
-0.00214722310192883,
0.05161888897418976,
-0.017192073166370392,
0.0014756915625184774,
-0.05356241390109062,
0.009727650322020054,
-0.05388687178492546,
0.038601066917181015,
-0.06809616833925247,
-0.018961172550916672,
-0.004680689889937639,
-0.06573090702295303,
0.056102387607097626,
0.023584190756082535,
-0.009628605097532272,
0.022943703457713127,
0.0035160258412361145,
-0.04542084038257599,
0.00953381322324276,
-0.03474609926342964,
0.03660983219742775,
-0.038935597985982895,
-0.013828354887664318,
0.002316713100299239,
-0.041618358343839645,
-0.006041188724339008,
-0.07436645776033401,
-0.06264878809452057,
0.01315959170460701,
-0.06286957859992981,
0.04095263406634331,
-0.03984776511788368,
-0.12236503511667252,
0.025208618491888046,
0.02925228886306286,
0.02530115284025669,
0.06346917152404785,
0.0007590975728817284,
-0.12192871421575546,
0.007188382558524609,
0.009930291213095188,
0.053290173411369324,
-0.04028918221592903,
0.016562649980187416,
0.031462814658880234,
0.0043363748118281364,
0.028287773951888084,
0.01137497927993536,
0.0777064636349678,
-0.08410470932722092,
0.04746786132454872,
-0.008635924197733402,
0.07665019482374191,
-0.0199831984937191,
-0.04281425103545189,
0.028640057891607285,
-0.03164055943489075,
0.09436313062906265,
-0.010702378116548061,
0.08030549436807632,
0.02731155790388584,
0.09668752551078796,
0.06232711672782898,
-0.08703586459159851,
0.04220319539308548,
0.0514528863132,
0.00008473862544633448,
-0.09989303350448608,
0.004804666619747877,
0.058574676513671875,
0.0578179731965065,
0.11740651726722717,
0.016346896067261696,
-0.058492232114076614,
-0.05890456959605217,
-0.02694891206920147,
0.10171462595462799,
-0.04941027984023094,
0.004147575702518225,
-0.006335882935672998,
-0.02095501311123371,
0.09746205061674118,
0.044894490391016006,
0.03525812551379204,
-0.0033446396701037884,
0.009799073450267315,
-0.014670311473309994,
0.025076959282159805,
0.06662332266569138,
0.01186976209282875,
-0.03837593272328377,
0.026426397264003754,
-0.017732983455061913,
-0.020794425159692764,
-0.03695986419916153,
0.009602466598153114,
0.0007931084255687892,
0.00242276256904006,
0.011817336082458496,
-0.014821099117398262,
-0.0144297294318676,
0.05769430845975876,
0.00466287462040782,
-0.017948398366570473,
-1.6785324645773002e-34,
0.04447036609053612,
0.030547920614480972,
0.027631321921944618,
0.004063616506755352,
0.025676269084215164,
-0.04140781983733177,
-0.019482240080833435,
0.03372974693775177,
-0.022809863090515137,
-0.04194989427924156,
0.013641705736517906,
-0.039989106357097626,
-0.06214986741542816,
0.05158088356256485,
0.022800104692578316,
0.03721817210316658,
0.03468620777130127,
0.014967814087867737,
-0.002258829539641738,
0.05301764979958534,
0.023125138133764267,
-0.027380557730793953,
-0.03618099167943001,
0.016645807772874832,
0.04076697304844856,
-0.0569288544356823,
0.03363845869898796,
-0.07973718643188477,
-0.02126764878630638,
0.03567345067858696,
-0.15638825297355652,
0.027037285268306732,
0.0454380102455616,
0.0036167644429951906,
0.0433061383664608,
-0.03574042767286301,
-0.03358905017375946,
0.0044419546611607075,
-0.009701648727059364,
-0.01538506243377924,
-0.0158994197845459,
0.04675532877445221,
-0.0398922823369503,
-0.044926855713129044,
-0.05740974098443985,
-0.06600692123174667,
-0.055295154452323914,
0.06585785746574402,
-0.12612982094287872,
-0.034860651940107346,
0.04986279085278511,
0.02766750566661358,
-0.0030309888534247875,
-0.05827367305755615,
-0.1339973509311676,
-0.013629480265080929,
0.01461920514702797,
-0.04657505452632904,
0.010848929174244404,
0.015824532136321068,
-0.014320296235382557,
0.07652328908443451,
-0.0278092660009861,
0.07320769876241684,
-0.034262821078300476,
-0.025043562054634094,
-0.01730125956237316,
0.0011150302598252892,
0.023694848641753197,
0.01110992394387722,
-0.10050398111343384,
0.05221738666296005,
-0.07229310274124146,
-0.01711733639240265,
0.005165723152458668,
0.023892801254987717,
0.06933113187551498,
-0.05807577073574066,
0.04365108907222748,
0.06677722930908203,
-0.010347871109843254,
-0.06657186150550842,
0.09459485113620758,
-0.020185407251119614,
-0.14725469052791595,
-0.019542621448636055,
0.04587368667125702,
-0.042827170342206955,
-0.04790789261460304,
-0.016374070197343826,
-0.04865359514951706,
-0.050179123878479004,
-0.05356194078922272,
0.019092483446002007,
0.01606086641550064,
-2.6468399667599092e-34,
-0.003311754437163472,
0.04076198488473892,
-0.06583321839570999,
0.028933655470609665,
-0.04430761933326721,
0.009953254833817482,
0.006656324025243521,
0.055579930543899536,
0.018813416361808777,
0.009478924795985222,
-0.03817499428987503,
0.028777558356523514,
0.020874852314591408,
-0.029040753841400146,
0.00011234894918743521,
-0.025259794667363167,
-0.017830688506364822,
0.09422680735588074,
-0.07298406213521957,
0.03581954538822174,
0.10571596771478653,
0.065443255007267,
-0.012634191662073135,
0.041928499937057495,
-0.01762491650879383,
0.038237154483795166,
-0.03727525845170021,
0.032128795981407166,
0.0068586221896111965,
0.01371665671467781,
-0.006721345707774162,
0.00036749549326486886,
0.03513644263148308,
-0.026985350996255875,
-0.08105596154928207,
0.031684212386608124,
0.05428289994597435,
0.032802000641822815,
0.046402785927057266,
0.039367616176605225,
0.09467940032482147,
-0.027759846299886703,
-0.13264797627925873,
-0.006953416392207146,
-0.06037433072924614,
-0.090370774269104,
0.06201912462711334,
0.0024168335366994143,
0.03708302974700928,
-0.01475900411605835,
-0.0257673729211092,
0.04070862755179405,
-0.05603870376944542,
0.10011599212884903,
-0.02130208909511566,
-0.08933447301387787,
0.04757944121956825,
-0.03541930392384529,
0.06221974641084671,
0.06648506224155426,
-0.030956823378801346,
0.0077711790800094604,
-0.05135750770568848,
-0.043964918702840805,
0.05419806018471718,
-0.07000115513801575,
-0.039422668516635895,
0.08601725101470947,
-0.001183894812129438,
-0.06160201504826546,
0.08538509160280228,
0.06306733936071396,
0.01745181530714035,
0.030198005959391594,
-0.02578684873878956,
-0.016352763399481773,
-0.01001681573688984,
-0.0011777676409110427,
0.013772526755928993,
0.05086477845907211,
0.046442218124866486,
-0.07921155542135239,
0.02948300912976265,
0.029765324667096138,
0.11114893108606339,
0.07320880144834518,
0.07066958397626877,
-0.03936922177672386,
0.030777480453252792,
-0.015586281195282936,
-0.05985763296484947,
-0.003645950695499778,
-0.03088374435901642,
0.06910105049610138,
0.016900310292840004,
-3.85715672734932e-8,
-0.05233842134475708,
0.02356286533176899,
0.007950956001877785,
0.010770408436655998,
0.008890059776604176,
0.003556732553988695,
-0.06953707337379456,
0.037760838866233826,
0.03519199416041374,
-0.04814665764570236,
0.09908802062273026,
0.013851100578904152,
-0.031109973788261414,
-0.13684357702732086,
-0.07275427877902985,
0.04472555220127106,
-0.06757936626672745,
0.07419971376657486,
-0.012485741637647152,
0.04347778856754303,
-0.009736735373735428,
-0.05245019868016243,
0.026737138628959656,
-0.04423848167061806,
0.03237761929631233,
-0.0974244624376297,
-0.009839404374361038,
0.02842067740857601,
0.08802292495965958,
0.02243770845234394,
-0.07712561637163162,
0.006038835272192955,
0.012399665080010891,
-0.011370746418833733,
-0.06066904962062836,
0.13508117198944092,
0.001971118850633502,
-0.03826169669628143,
-0.03530123084783554,
0.0023348163813352585,
-0.03220851719379425,
-0.032256800681352615,
0.014488734304904938,
-0.054182421416044235,
-0.024579832330346107,
0.03229352831840515,
0.036098819226026535,
-0.14305606484413147,
0.0821172371506691,
0.0184959527105093,
0.027219850569963455,
0.003437392646446824,
0.015407819300889969,
0.1669982671737671,
-0.01458169799298048,
0.002586214803159237,
0.0059528956189751625,
-0.032314371317625046,
0.027310967445373535,
0.0536988228559494,
0.04793684557080269,
0.02433946542441845,
0.027749134227633476,
-0.020323945209383965
] |
Solving Video Inverse Problems Using Image Diffusion Models | https://openreview.net/forum?id=TRWxFUzK9K | [
"Taesung Kwon",
"Jong Chul Ye"
] | Poster | Recently, diffusion model-based inverse problem solvers (DIS) have emerged as state-of-the-art approaches for addressing inverse problems, including image super-resolution, deblurring, inpainting, etc.
However, their application to video inverse problems arising from spatio-temporal degradation remains largely unexplored due to the challenges in training video diffusion models.
To address this issue, here we introduce an innovative video inverse solver that leverages only image diffusion models.
Specifically, by
drawing inspiration from the success of the recent decomposed diffusion sampler (DDS),
our method treats the time dimension of a video as the batch dimension of image diffusion models and solves spatio-temporal optimization problems within denoised spatio-temporal batches derived from each image diffusion model.
Moreover, we introduce a batch-consistent diffusion sampling strategy that encourages consistency across batches by synchronizing the stochastic noise components in image diffusion models.
Our approach synergistically combines batch-consistent sampling with simultaneous optimization of denoised spatio-temporal batches at each reverse diffusion step, resulting in a novel and efficient diffusion sampling strategy for video inverse problems.
Experimental results demonstrate that our method effectively addresses various spatio-temporal degradations in video inverse problems, achieving state-of-the-art reconstructions.
Project page: https://svi-diffusion.github.io/ | Image diffusion models, Video inverse problems, Batch-consistent sampling | We propose a video inverse problem solver using an image diffusion model with a batch-consistent sampling strategy. | 760 | 2409.02574 | [
-0.051311835646629333,
-0.10172976553440094,
0.09145226329565048,
0.03249296545982361,
0.05161454528570175,
-0.10644382238388062,
-0.004354901146143675,
-0.05085214227437973,
0.052771274000406265,
-0.040619611740112305,
-0.04224231094121933,
0.024229571223258972,
-0.04805552959442139,
0.07212348282337189,
-0.04681539908051491,
-0.013625099323689938,
0.02068890817463398,
0.06051535904407501,
-0.09952607750892639,
-0.04210250824689865,
-0.050131916999816895,
-0.07035636901855469,
-0.04446522518992424,
-0.019690416753292084,
0.04989558085799217,
-0.03689272329211235,
0.055694472044706345,
-0.06306161731481552,
-0.06181738153100014,
0.006306889466941357,
0.08002254366874695,
-0.0026022514794021845,
0.05636528134346008,
-0.02177261933684349,
-0.0364878885447979,
0.01894534006714821,
-0.027499979361891747,
0.008243924006819725,
0.029317254200577736,
-0.018542716279625893,
0.005703340750187635,
0.06851783394813538,
0.011556953191757202,
-0.04333411157131195,
0.03949352726340294,
-0.02759174443781376,
-0.02060627192258835,
-0.05301900580525398,
0.04893926903605461,
0.05336706340312958,
0.03252382203936577,
0.012215952388942242,
-0.09522491693496704,
0.013591056689620018,
-0.021033689379692078,
0.06138714402914047,
0.0793277770280838,
0.0024659528862684965,
0.015457331202924252,
0.01709282025694847,
0.010999917984008789,
-0.07551804929971695,
-0.04834960401058197,
0.008034484460949898,
-0.01354129333049059,
0.014103244990110397,
0.09128111600875854,
-0.026657352223992348,
-0.01774105429649353,
0.0852082371711731,
-0.07734378427267075,
0.052045125514268875,
-0.03903282806277275,
0.025929804891347885,
-0.044391606003046036,
-0.02599368244409561,
0.0006233007297851145,
-0.02243809401988983,
0.018356608226895332,
-0.04586920142173767,
0.1299213021993637,
-0.003975829109549522,
0.11667808890342712,
-0.011419511400163174,
0.09991679340600967,
-0.014250023290514946,
0.01558684278279543,
0.06859375536441803,
0.04162902757525444,
-0.007805448025465012,
-0.13070182502269745,
0.03977823629975319,
-0.062490034848451614,
-0.01717088185250759,
0.05292341858148575,
-0.0038002859801054,
0.03566865250468254,
-0.011840924620628357,
0.04846827685832977,
0.06573671102523804,
-0.014900042675435543,
-0.02445153146982193,
-0.046875279396772385,
-0.01666364073753357,
0.04180777072906494,
0.013244659639894962,
0.05233503505587578,
0.05550818517804146,
0.02093471586704254,
0.01566128619015217,
0.056678131222724915,
-0.032282959669828415,
0.11104807257652283,
0.02393483929336071,
0.08673516660928726,
-0.06196817010641098,
0.07452094554901123,
-0.03157933056354523,
-0.0461559034883976,
0.003141235327348113,
0.01155826635658741,
-0.05177691951394081,
-0.07549616694450378,
-0.08447254449129105,
0.05608231574296951,
0.02205634117126465,
0.08146902173757553,
2.780722436669333e-33,
-0.05874050408601761,
-0.021169204264879227,
0.060043156147003174,
0.026685569435358047,
0.08013471961021423,
0.008885861374437809,
0.07189112901687622,
-0.024547744542360306,
0.014452637173235416,
-0.053144317120313644,
0.0038406422827392817,
-0.08110859245061874,
-0.08175644278526306,
0.028827840462327003,
-0.021901188418269157,
0.014324666000902653,
-0.037040457129478455,
0.08278503268957138,
-0.024947723373770714,
-0.047616466879844666,
0.046877071261405945,
-0.01011679321527481,
-0.05101187899708748,
-0.031159358099102974,
-0.006271852646023035,
-0.0610530860722065,
0.0640544667840004,
-0.05121227353811264,
0.05761697515845299,
-0.021404331550002098,
-0.0911799967288971,
-0.030932458117604256,
-0.016296811401844025,
0.07107199728488922,
0.046074360609054565,
0.04247526824474335,
0.06464262306690216,
0.015217688865959644,
-0.024909373372793198,
-0.029463190585374832,
0.013869477435946465,
0.09538383781909943,
-0.061332836747169495,
-0.03432206064462662,
-0.11159352958202362,
-0.025557134300470352,
0.04270368069410324,
0.01950567215681076,
-0.007803092245012522,
-0.013806242495775223,
0.022262457758188248,
-0.04012064263224602,
0.005738561507314444,
-0.14951038360595703,
-0.04840882867574692,
0.027167484164237976,
0.11552656441926956,
-0.048958130180835724,
0.050067637115716934,
0.028506511822342873,
0.08364535868167877,
-0.014401480555534363,
-0.06069420650601387,
0.02355489693582058,
0.035492829978466034,
-0.03506924957036972,
0.01795775070786476,
0.016914241015911102,
-0.024918507784605026,
0.03163105621933937,
-0.0006491696112789214,
0.0416133850812912,
-0.0243516955524683,
-0.07460429519414902,
0.033104535192251205,
0.041367895901203156,
0.053365495055913925,
0.036812376230955124,
-0.025291994214057922,
0.031453557312488556,
-0.07113374024629593,
-0.04362579435110092,
-0.025811817497015,
-0.07464267313480377,
0.00448898458853364,
-0.03044774942100048,
0.056715380400419235,
-0.07905883342027664,
-0.0359790064394474,
-0.08276654034852982,
-0.014694929122924805,
0.009843519888818264,
-0.05847128853201866,
0.03338039293885231,
0.08642897009849548,
-6.237766741945719e-34,
0.04124440997838974,
0.10790327191352844,
0.021487358957529068,
0.06409025192260742,
-0.013222337700426579,
-0.02072424255311489,
0.023473374545574188,
0.025144197046756744,
0.12477800995111465,
-0.11399458348751068,
-0.08414564281702042,
-0.047155074775218964,
-0.03323187679052353,
-0.023551998659968376,
-0.01795113831758499,
-0.04476526007056236,
0.032310549169778824,
-0.026108596473932266,
-0.11174602806568146,
0.020996831357479095,
-0.03209362179040909,
0.08598538488149643,
-0.01731147989630699,
0.0031813804525882006,
-0.08185840398073196,
0.05423224717378616,
-0.004839928820729256,
0.08831636607646942,
0.07650730758905411,
-0.04116931930184364,
-0.021798256784677505,
0.003094361163675785,
0.04422354698181152,
-0.02537686750292778,
-0.11140484362840652,
0.14183636009693146,
0.02985173650085926,
0.0001097401836887002,
-0.03876487538218498,
0.04421424865722656,
0.0234844908118248,
-0.03793812915682793,
-0.060968801379203796,
0.061045851558446884,
0.051891572773456573,
-0.06810210645198822,
-0.019455987960100174,
-0.012667613103985786,
0.0096026211977005,
-0.025006650015711784,
-0.02330506034195423,
-0.0002796124608721584,
-0.07563912868499756,
0.07596087455749512,
0.014701752923429012,
-0.028923384845256805,
-0.07171545922756195,
-0.002924993634223938,
0.09262674301862717,
0.03476379066705704,
0.02961949072778225,
-0.018647246062755585,
0.016535423696041107,
-0.057303838431835175,
-0.07785173505544662,
0.04617951437830925,
-0.02822125516831875,
-0.01694268360733986,
-0.029459528625011444,
0.057866767048835754,
-0.024799199774861336,
-0.0010492029832676053,
0.010259523056447506,
0.023865504190325737,
0.014232493937015533,
-0.07864072173833847,
0.0041189431212842464,
0.0533638671040535,
-0.027828751131892204,
-0.03709454834461212,
-0.028358610346913338,
-0.019265206530690193,
-0.006049514748156071,
0.019121680408716202,
0.020729627460241318,
0.035968709737062454,
0.030210742726922035,
0.02091159299015999,
0.015365486964583397,
0.01932479441165924,
-0.042513974010944366,
-0.01847977563738823,
0.0742654949426651,
0.02122681587934494,
0.004554992076009512,
-3.6138445125288854e-8,
-0.09301618486642838,
0.06367310136556625,
0.05993488430976868,
0.01380407065153122,
-0.02055058814585209,
-0.01365503016859293,
0.10726118087768555,
0.04486680030822754,
0.07595804333686829,
-0.0712900310754776,
0.06429287791252136,
-0.004881205968558788,
0.008476434275507927,
-0.01512714009732008,
-0.10347974300384521,
0.06255210936069489,
-0.010013210587203503,
0.008716044947504997,
0.002592409960925579,
0.05484144762158394,
-0.08045632392168045,
0.00990994367748499,
0.05060410872101784,
-0.049161601811647415,
0.04126182943582535,
0.0385747030377388,
-0.04414079710841179,
-0.010369079187512398,
0.1256895810365677,
-0.026160327717661858,
-0.005770286079496145,
0.02171674743294716,
0.07643890380859375,
-0.021964233368635178,
-0.021845415234565735,
0.008712130598723888,
0.011094476096332073,
0.008639787323772907,
-0.06888505071401596,
-0.10053550451993942,
-0.05174145847558975,
-0.011568957008421421,
-0.005210149567574263,
-0.014312622137367725,
0.002598251448944211,
0.07220727205276489,
0.03620421141386032,
0.002488820580765605,
0.022010931745171547,
0.01631035842001438,
0.05499577894806862,
-0.06860652565956116,
-0.009327445179224014,
0.040374934673309326,
0.09961660951375961,
-0.04839090630412102,
0.05594257265329361,
0.03159130737185478,
0.06949371844530106,
0.06521840393543243,
0.006155515089631081,
0.02487598918378353,
-0.011782071553170681,
0.014133080840110779
] |
Jailbreaking as a Reward Misspecification Problem | https://openreview.net/forum?id=uBnM3EFovQ | [
"Zhihui Xie",
"Jiahui Gao",
"Lei Li",
"Zhenguo Li",
"Qi Liu",
"Lingpeng Kong"
] | Poster | The widespread adoption of large language models (LLMs) has raised concerns about their safety and reliability, particularly regarding their vulnerability to adversarial attacks. In this paper, we propose a new perspective that attributes this vulnerability to reward misspecification during the alignment process. This misspecification occurs when the reward function fails to accurately capture the intended behavior, leading to misaligned model outputs. We introduce a metric ReGap to quantify the extent of reward misspecification and demonstrate its effectiveness and robustness in detecting harmful backdoor prompts. Building upon these insights, we present ReMiss, a system for automated red teaming that generates adversarial prompts in a reward-misspecified space. ReMiss achieves state-of-the-art attack success rates on the AdvBench benchmark against various target aligned LLMs while preserving the human readability of the generated prompts. Furthermore, these attacks on open-source models demonstrate high transferability to closed-source models like GPT-4o and out-of-distribution tasks from HarmBench. Detailed analysis highlights the unique advantages of the proposed reward misspecification objective compared to previous methods, offering new insights for improving LLM safety and robustness. | Large language models, alignment, jailbreaking | This paper reframes jailbreaking as a reward misspecification problem and introduces novel methods to quantify and exploit these misspecifications in aligned language models. | 756 | 2406.14393 | [
-0.08659680187702179,
-0.06764122098684311,
-0.04745291545987129,
0.018235111609101295,
0.03773525729775429,
0.08001435548067093,
0.004169668070971966,
-0.02742386981844902,
0.046602100133895874,
-0.02768288552761078,
-0.03291301056742668,
-0.10287874937057495,
0.0708247646689415,
-0.01494542695581913,
0.026915980502963066,
-0.029444940388202667,
0.04146016761660576,
0.1253870278596878,
-0.08218652009963989,
-0.034350208938121796,
-0.028370056301355362,
0.020332489162683487,
0.00011152693332405761,
0.018338682129979134,
-0.013025629334151745,
-0.03611142188310623,
-0.049593985080718994,
-0.054678648710250854,
-0.06101124361157417,
-0.034114353358745575,
-0.014623661525547504,
-0.047802455723285675,
0.04671807214617729,
0.026371506974101067,
-0.043815381824970245,
0.049840234220027924,
-0.06850889325141907,
-0.018918471410870552,
0.0004017733735963702,
-0.12153983116149902,
-0.04936916381120682,
0.009250444360077381,
0.0038088327273726463,
0.03703972324728966,
0.017736347392201424,
-0.08381038904190063,
0.006153056398034096,
0.03995543718338013,
0.00541734928265214,
-0.0665566474199295,
-0.020377837121486664,
-0.013085956685245037,
0.020445851609110832,
0.09704133123159409,
-0.07551824301481247,
0.04680093750357628,
0.07658464461565018,
0.07100150734186172,
0.012334326282143593,
0.02074800617992878,
-0.028666196390986443,
-0.031363990157842636,
-0.04502430558204651,
0.033013854175806046,
-0.0154966339468956,
-0.030007818713784218,
0.039921097457408905,
0.03659169748425484,
-0.0016238248208537698,
0.05873927101492882,
-0.0015066280029714108,
0.019480949267745018,
-0.04394855722784996,
-0.017499525099992752,
0.04268626868724823,
0.14193806052207947,
-0.046945467591285706,
-0.000592189549934119,
0.03207077085971832,
-0.06488610059022903,
0.021190518513321877,
-0.026383953168988228,
0.06639528274536133,
0.04587167128920555,
0.09936632961034775,
0.02463175356388092,
-0.014495933428406715,
0.02319258637726307,
0.13801905512809753,
0.0470941886305809,
-0.08450579643249512,
0.004289952106773853,
0.06922828406095505,
0.04220378398895264,
0.03562168404459953,
0.03006627783179283,
0.004983808379620314,
-0.05595037713646889,
-0.059614069759845734,
0.0950435921549797,
-0.007814818993210793,
0.04489292576909065,
0.03742605447769165,
-0.07745855301618576,
0.06887510418891907,
0.006951313465833664,
0.013163452968001366,
-0.028390754014253616,
0.02958519011735916,
-0.07139487564563751,
0.04307069256901741,
0.038916900753974915,
0.08544066548347473,
0.00559869222342968,
0.006365828216075897,
0.010953327640891075,
-0.08471027761697769,
-0.028638355433940887,
-0.01514419261366129,
0.008709505200386047,
0.007062107790261507,
0.019072867929935455,
0.044470835477113724,
-0.022030644118785858,
0.06277398765087128,
-0.01670042797923088,
-0.06614922732114792,
2.753024483686098e-33,
0.03742636740207672,
0.05485750734806061,
-0.01198284700512886,
0.03311033174395561,
0.08952928334474564,
-0.00240315985865891,
-0.03219277039170265,
0.08841748535633087,
-0.05821320042014122,
-0.024842509999871254,
-0.031219303607940674,
-0.04698896035552025,
-0.013695474714040756,
0.07433216273784637,
0.07868804037570953,
0.0283046867698431,
-0.011838861741125584,
0.033301662653684616,
-0.07223927974700928,
0.013827390968799591,
0.02696968987584114,
-0.05486619099974632,
0.0016062522772699594,
-0.025974443182349205,
0.015967780724167824,
0.030046643689274788,
-0.01642596907913685,
-0.0337710864841938,
-0.028629634529352188,
0.030973033979535103,
-0.12419349700212479,
-0.039677154272794724,
0.10193406790494919,
0.02043493278324604,
0.013656122609972954,
-0.027962051331996918,
-0.022953391075134277,
-0.08185376226902008,
-0.011582489125430584,
0.005067160818725824,
-0.019656136631965637,
0.012907575815916061,
-0.029472853988409042,
0.011912309564650059,
0.0037395793478935957,
-0.025164812803268433,
-0.005868204403668642,
0.003946294542402029,
-0.019087286666035652,
-0.012641961686313152,
-0.014766902662813663,
0.08021823316812515,
0.01546727865934372,
0.0034513878636062145,
-0.1457795351743698,
-0.011062788777053356,
-0.011472621001303196,
0.016978278756141663,
-0.019340718165040016,
0.09881875663995743,
-0.003931844141334295,
0.04821271449327469,
-0.01804455928504467,
0.029040178284049034,
0.014129864983260632,
-0.06181635335087776,
0.01766863279044628,
0.002642173320055008,
-0.011525748297572136,
0.043863292783498764,
-0.05367831513285637,
-0.030028963461518288,
-0.03832900524139404,
0.02419092319905758,
-0.036360591650009155,
-0.022955933585762978,
0.10223418474197388,
0.0020822470542043447,
0.0911090150475502,
-0.09992243349552155,
0.030529264360666275,
-0.051301758736371994,
-0.0021683729719370604,
-0.015576859004795551,
-0.07651589065790176,
-0.06204311177134514,
0.014421189203858376,
-0.09063106775283813,
-0.028893444687128067,
0.08343157172203064,
-0.019260738044977188,
-0.036442436277866364,
-0.02852761000394821,
0.029202958568930626,
0.010519945994019508,
-3.300461202742661e-33,
-0.06554272025823593,
-0.02089715749025345,
-0.02073122374713421,
0.04087182879447937,
-0.0355033315718174,
-0.048129599541425705,
-0.01364658772945404,
0.03545986860990524,
-0.01176756527274847,
-0.0003451010852586478,
-0.07267117500305176,
0.018069101497530937,
0.04987267032265663,
0.025223372504115105,
0.02703835815191269,
-0.041842468082904816,
-0.01306842640042305,
0.0558360330760479,
-0.023463761433959007,
0.0447578988969326,
0.10609791427850723,
0.11431416124105453,
-0.10194136202335358,
0.06443841755390167,
0.027623621746897697,
-0.03782159090042114,
-0.014131832867860794,
0.02507799305021763,
0.02980368584394455,
0.02740885131061077,
0.0013716082321479917,
0.0573674812912941,
-0.07374366372823715,
-0.019200656563043594,
-0.060298819094896317,
0.05747023597359657,
0.09393181651830673,
-0.019269384443759918,
-0.0454711988568306,
0.11786038428544998,
0.07991079986095428,
-0.018305586650967598,
-0.14733116328716278,
-0.052895065397024155,
0.016324320808053017,
0.02790912240743637,
-0.08808610588312149,
0.004239926114678383,
-0.0071660494431853294,
-0.07870014756917953,
0.015592270530760288,
0.013264940120279789,
-0.018046362325549126,
0.030940186232328415,
-0.09208256751298904,
-0.10847210139036179,
-0.03833896294236183,
-0.038763999938964844,
0.04948893561959267,
0.013600446283817291,
-0.05941690877079964,
-0.017347902059555054,
0.026248695328831673,
-0.030529677867889404,
0.026445960626006126,
-0.07434438169002533,
0.02437708154320717,
0.08374287933111191,
-0.0013914462178945541,
0.0036501430440694094,
0.07691961526870728,
-0.03040730208158493,
0.0028733632061630487,
0.06745985895395279,
-0.04103511944413185,
-0.00035578309325501323,
-0.07499164342880249,
-0.08511415123939514,
-0.033518236130476,
-0.040332794189453125,
0.09788160771131516,
0.016837874427437782,
0.08114316314458847,
0.046191830188035965,
0.055799685418605804,
0.030885111540555954,
0.1071612611413002,
0.11540487408638,
-0.025912277400493622,
0.04465485364198685,
-0.04167412593960762,
0.015849115327000618,
0.003507328452542424,
0.1044350415468216,
-0.029799340292811394,
-4.792720176283183e-8,
-0.00539811048656702,
0.05655869469046593,
-0.03838195279240608,
0.048612918704748154,
-0.02968546189367771,
-0.003703532973304391,
-0.10551337897777557,
-0.011763370595872402,
0.013477538712322712,
-0.038351643830537796,
0.021753771230578423,
-0.03674374520778656,
-0.037687890231609344,
-0.05585106089711189,
-0.05894853547215462,
0.10696296393871307,
-0.08803807199001312,
0.038416750729084015,
-0.04674803838133812,
-0.033885955810546875,
0.008807292208075523,
0.0867137610912323,
-0.016397858038544655,
-0.05167068541049957,
0.01105074305087328,
-0.07167109102010727,
-0.049386829137802124,
0.06718221306800842,
0.021653279662132263,
0.02605857513844967,
0.005247344262897968,
-0.05414818227291107,
0.025173915550112724,
0.011073174886405468,
0.024362990632653236,
0.13766086101531982,
-0.036804430186748505,
-0.05313332378864288,
0.04984934628009796,
0.02934223599731922,
0.006062157917767763,
0.041710153222084045,
-0.028416164219379425,
0.00034025489003397524,
-0.053352050483226776,
-0.0012306319549679756,
-0.02617688477039337,
-0.10040552169084549,
0.07569465786218643,
-0.0641896054148674,
0.05021721124649048,
-0.05953839048743248,
-0.03444986045360565,
0.03153609111905098,
0.02137100324034691,
0.0278889462351799,
-0.040255289524793625,
-0.01580836996436119,
0.05778888240456581,
0.037187837064266205,
0.05679486319422722,
0.013552815653383732,
-0.006406423635780811,
-0.044816430658102036
] |
Streaming Video Understanding and Multi-round Interaction with Memory-enhanced Knowledge | https://openreview.net/forum?id=JbPb6RieNC | [
"Haomiao Xiong",
"Zongxin Yang",
"Jiazuo Yu",
"Yunzhi Zhuge",
"Lu Zhang",
"Jiawen Zhu",
"Huchuan Lu"
] | Poster | Recent advances in Large Language Models (LLMs) have enabled the development of Video-LLMs, advancing multimodal learning by bridging video data with language tasks. However, current video understanding models struggle with processing long video sequences, supporting multi-turn dialogues, and adapting to real-world dynamic scenarios. To address these issues, we propose StreamChat, a training-free framework for streaming video reasoning and conversational interaction. StreamChat leverages a novel hierarchical memory system to efficiently process and compress video features over extended sequences, enabling real-time, multi-turn dialogue. Our framework incorporates a parallel system scheduling strategy that enhances processing speed and reduces latency, ensuring robust performance in real-world applications. Furthermore, we introduce StreamBench, a versatile benchmark that evaluates streaming video understanding across diverse media types and interactive scenarios, including multi-turn interactions and complex reasoning tasks. Extensive evaluations on StreamBench and other public benchmarks demonstrate that StreamChat significantly outperforms existing
state-of-the-art models in terms of accuracy and response times, confirming its effectiveness for streaming video understanding. Code is available at StreamChat. | Streaming Video Understanding; Video MLLM; Hierarchical Memory System | A Vertasile Approach and Benchmark for Streaming Video Understanding and Multi-round Interaction | 743 | 2501.13468 | [
-0.003918366972357035,
-0.11746517568826675,
0.028497077524662018,
-0.001304015051573515,
0.10134665668010712,
-0.012754212133586407,
-0.013658997602760792,
-0.0013638971140608191,
0.04809384047985077,
-0.04827246069908142,
-0.07968773692846298,
-0.052214935421943665,
-0.07586070895195007,
0.1025003120303154,
0.012497271411120892,
-0.016458772122859955,
0.02719532512128353,
0.0006619662162847817,
-0.0892506018280983,
-0.04099590703845024,
0.0214132871478796,
-0.06712556630373001,
0.08976270258426666,
0.022169306874275208,
0.029650993645191193,
0.01599549502134323,
0.001512355636805296,
-0.035688869655132294,
0.05390579253435135,
-0.0036889442708343267,
0.0476541742682457,
-0.015266699716448784,
-0.0025709769688546658,
0.06280315667390823,
-0.07322079688310623,
0.024479180574417114,
0.014402167871594429,
-0.007293323054909706,
-0.09697245061397552,
-0.01454362366348505,
-0.0073334514163434505,
0.017093276605010033,
0.07341914623975754,
-0.008039990440011024,
0.03618678078055382,
0.03513015806674957,
-0.020713645964860916,
-0.06572052836418152,
0.03916430100798607,
0.057649657130241394,
-0.08851136267185211,
-0.028245598077774048,
0.005714959464967251,
0.10791407525539398,
-0.05794661119580269,
0.13040585815906525,
0.004589152988046408,
0.021980388090014458,
0.04871794953942299,
0.01892700046300888,
-0.018398558720946312,
-0.15262068808078766,
0.0046228570863604546,
0.025857238098978996,
-0.07814133912324905,
-0.008939981460571289,
0.061676204204559326,
0.0005247412482276559,
-0.001201991573907435,
-0.007926988415420055,
-0.07544192671775818,
0.0523216687142849,
-0.04397423565387726,
0.0879589095711708,
-0.08098315447568893,
-0.030590727925300598,
0.04667162522673607,
-0.028637362644076347,
0.01440933346748352,
-0.046933796256780624,
0.03679918870329857,
-0.045895881950855255,
0.01987140066921711,
-0.006707541644573212,
0.12094230204820633,
-0.047328170388936996,
-0.0026327811647206545,
0.04791807383298874,
-0.05387627333402634,
0.06300245970487595,
-0.07888147234916687,
-0.018349971622228622,
0.08574970811605453,
-0.04420351982116699,
0.05940702557563782,
0.09460832178592682,
0.03917481377720833,
-0.08846691250801086,
-0.05295043811202049,
0.04858056455850601,
-0.00019926103414036334,
0.04596354067325592,
-0.00920808408409357,
-0.06258901953697205,
0.035086795687675476,
0.010676476173102856,
0.03919796273112297,
0.06832322478294373,
0.0034500753972679377,
-0.013155920431017876,
0.01937689259648323,
0.024787727743387222,
0.07088210433721542,
-0.013547861017286777,
0.04564102366566658,
0.05280506610870361,
-0.016828780993819237,
-0.0899643823504448,
-0.017853781580924988,
0.04324718192219734,
0.04757080599665642,
-0.037109993398189545,
-0.045241229236125946,
-0.028854485601186752,
0.03358728066086769,
-0.020659897476434708,
0.033441103994846344,
2.434836346346744e-34,
0.07926953583955765,
0.013054346665740013,
0.035669565200805664,
0.09027496725320816,
0.03411717712879181,
0.036278679966926575,
-0.0000082191791079822,
0.09387367218732834,
-0.08945228904485703,
-0.0686156153678894,
0.029490413144230843,
-0.10048235207796097,
-0.04075893387198448,
0.021476861089468002,
-0.011316617019474506,
-0.0232414361089468,
-0.07289579510688782,
0.01439471822232008,
-0.01857612654566765,
-0.04013485088944435,
0.02372652105987072,
0.0029244772158563137,
-0.024659698829054832,
0.03451601043343544,
0.028390366584062576,
0.019685139879584312,
0.0842321440577507,
-0.07625443488359451,
0.0556897334754467,
-0.02993258647620678,
-0.101984903216362,
-0.01365690864622593,
-0.02938809059560299,
0.020938806235790253,
0.10428477078676224,
-0.06527817994356155,
-0.060187432914972305,
-0.08219388127326965,
0.007294568698853254,
-0.0189006756991148,
-0.04794083163142204,
-0.04038436338305473,
-0.03776971995830536,
-0.045489951968193054,
-0.07755257934331894,
-0.08261114358901978,
0.019715948030352592,
0.05518857389688492,
-0.116266168653965,
0.04400024935603142,
0.01929524913430214,
0.09144622832536697,
-0.03763722628355026,
0.0035405424423515797,
-0.03267110511660576,
-0.03787190467119217,
0.04356896132230759,
0.014155573211610317,
0.0012231022119522095,
0.020395681262016296,
0.00301060127094388,
0.042427752166986465,
-0.017289141193032265,
0.07879321277141571,
0.04976646229624748,
-0.005096153821796179,
0.014168044552206993,
0.13167960941791534,
0.05286097899079323,
-0.041146472096443176,
0.0037713232450187206,
0.0007729676435701549,
0.04412917047739029,
-0.01635543629527092,
0.0006033044192008674,
0.0847020223736763,
0.017720630392432213,
-0.07290279120206833,
0.033595044165849686,
0.048367030918598175,
0.03331209719181061,
-0.04859548062086105,
0.02300305850803852,
-0.04813665896654129,
0.01832299306988716,
-0.062186431139707565,
0.036799680441617966,
-0.11767493933439255,
-0.025671038776636124,
0.0009093991830013692,
-0.06023452430963516,
-0.011732980608940125,
0.032882485538721085,
0.0011877452488988638,
0.029889259487390518,
1.566987785246858e-33,
0.07078749686479568,
0.011941169388592243,
-0.08646242320537567,
0.052794199436903,
0.08893571048974991,
-0.05126062408089638,
0.022666819393634796,
0.04029657691717148,
0.03595874831080437,
-0.07311450690031052,
-0.042794253677129745,
-0.024415384978055954,
-0.0362095907330513,
-0.044672008603811264,
0.05114852637052536,
-0.024937408044934273,
-0.010452630929648876,
-0.0792296826839447,
-0.05611779913306236,
0.03203919529914856,
-0.02228824608027935,
0.023822356015443802,
-0.04253989830613136,
-0.007488803938031197,
-0.08319876343011856,
0.028428135439753532,
0.015503025613725185,
0.05951915308833122,
0.008217751048505306,
0.04245932027697563,
0.004363167565315962,
-0.025692835450172424,
0.004621332045644522,
-0.03197060152888298,
-0.05264889821410179,
0.032784219831228256,
0.09407062828540802,
-0.014172956347465515,
-0.05916468799114227,
0.04224814847111702,
0.15390826761722565,
-0.0907931849360466,
-0.0798189714550972,
-0.0030850598122924566,
-0.0034001749008893967,
0.019718416035175323,
-0.15329903364181519,
-0.006036027800291777,
-0.07304764539003372,
0.027806652709841728,
-0.014949491247534752,
0.01791287772357464,
-0.014057000167667866,
0.002363649196922779,
0.024090634658932686,
-0.13126710057258606,
-0.050501927733421326,
0.01874719373881817,
0.04271412640810013,
0.0026331832632422447,
-0.022004347294569016,
-0.00417432002723217,
0.0167689248919487,
-0.07283496111631393,
0.06776826083660126,
-0.014201406389474869,
-0.029128529131412506,
-0.01876145787537098,
-0.020706309005618095,
0.01113162375986576,
0.05295681208372116,
-0.04712752625346184,
0.003628780832514167,
-0.007749653421342373,
0.05980772152543068,
0.017427008599042892,
0.025567572563886642,
-0.024634800851345062,
0.0019007375231012702,
-0.01706813834607601,
-0.0019438824383541942,
-0.01114827860146761,
0.05799921602010727,
0.049720533192157745,
0.06243624910712242,
0.07453393936157227,
0.04191659018397331,
0.10602176189422607,
-0.023512208834290504,
0.05488748475909233,
-0.010077105835080147,
0.06482064723968506,
0.00401346804574132,
0.043089479207992554,
0.000521355657838285,
-3.0855808574870025e-8,
-0.1269753873348236,
0.000857987382914871,
-0.03589659184217453,
0.026561614125967026,
-0.008440329693257809,
-0.05448635667562485,
0.014933326281607151,
0.044737812131643295,
0.05469660833477974,
-0.005541743710637093,
0.07374929636716843,
-0.02622346207499504,
0.06705252826213837,
-0.019585907459259033,
-0.026039768010377884,
0.030302975326776505,
-0.02487947605550289,
-0.05240553617477417,
-0.004731609486043453,
0.042807869613170624,
0.09074710309505463,
-0.003767658257856965,
0.0077307685278356075,
0.040324341505765915,
-0.045896440744400024,
-0.025732068344950676,
0.049773141741752625,
0.1450852006673813,
0.034378163516521454,
-0.05138591304421425,
-0.06325828284025192,
0.014768103137612343,
0.020657943561673164,
-0.06791334599256516,
0.0460871197283268,
0.004922162275761366,
0.01031459216028452,
-0.04370184615254402,
0.03755264729261398,
-0.03609684482216835,
0.02295367605984211,
0.022827301174402237,
-0.016735969111323357,
0.016338378190994263,
0.0006285327835939825,
0.05703386291861534,
-0.06080828979611397,
-0.06567532569169998,
0.016320636495947838,
0.032347843050956726,
-0.07201412320137024,
0.006330504547804594,
0.041185349225997925,
0.06448613107204437,
0.07430820912122726,
0.01632669009268284,
0.009316460229456425,
0.01246851496398449,
0.04683059826493263,
0.09103237837553024,
0.009428629651665688,
0.04409397020936012,
-0.017627347260713577,
0.05388779938220978
] |
Beyond the convexity assumption: Realistic tabular data generation under quantifier-free real linear constraints | https://openreview.net/forum?id=rx0TCew0Lj | [
"Mihaela C. Stoian",
"Eleonora Giunchiglia"
] | Poster | Synthetic tabular data generation has traditionally been a challenging problem due to the high complexity of the underlying distributions that characterise this type of data. Despite recent advances in deep generative models (DGMs), existing methods often fail to produce realistic datapoints that are well-aligned with available background knowledge.
In this paper, we address this limitation by introducing Disjunctive Refinement Layer (DRL), a novel layer designed
to enforce the alignment of generated data with the background knowledge specified in user-defined constraints.
DRL is the first method able to automatically make deep learning models inherently compliant with constraints as expressive as quantifier-free linear formulas, which can define non-convex and even disconnected spaces.
Our experimental analysis shows that DRL not only guarantees constraint satisfaction but also improves efficacy in downstream tasks. Notably, when applied to DGMs that frequently violate constraints, DRL eliminates violations entirely. Further, it improves performance metrics by up to 21.4\% in F1-score and 20.9\% in Area Under the ROC Curve, thus demonstrating its practical impact on data generation. | tabular data generation, neuro-symbolic AI, informed machine learning, safe AI | null | 742 | 2502.18237 | [
-0.07637017965316772,
-0.07197479903697968,
0.04726911708712578,
0.03392189368605614,
0.011758237145841122,
0.04122297093272209,
-0.05893716216087341,
-0.10200124979019165,
-0.01537913829088211,
-0.01755325123667717,
-0.0675891637802124,
-0.09984344989061356,
0.008028320036828518,
0.03735121712088585,
-0.05967012792825699,
0.008263392373919487,
0.06535118818283081,
0.06627900898456573,
-0.13389739394187927,
0.02147839218378067,
0.061697859317064285,
0.01577918604016304,
-0.002736498136073351,
-0.023034105077385902,
-0.010230648331344128,
-0.016402000561356544,
0.051122765988111496,
-0.05465816706418991,
0.04845960810780525,
-0.027163278311491013,
0.034149907529354095,
0.008621417917311192,
-0.05603596568107605,
0.04155755788087845,
0.013434948399662971,
0.09275708347558975,
-0.08282661437988281,
0.031382810324430466,
-0.012826620601117611,
-0.03384148329496384,
-0.06803030520677567,
0.009004946798086166,
0.03019317239522934,
0.0475953184068203,
0.092317134141922,
0.012160778045654297,
0.013637027703225613,
0.0184309184551239,
0.006991216912865639,
0.024333039298653603,
-0.06818466633558273,
0.053619883954524994,
-0.030382828786969185,
0.07416924834251404,
0.02191823348402977,
0.06581755727529526,
0.028433820232748985,
-0.0009531269315630198,
0.003404089016839862,
0.01682974584400654,
-0.047034744173288345,
-0.08762693405151367,
-0.013582074083387852,
-0.06142525002360344,
-0.06603232026100159,
-0.03572388365864754,
-0.07481738924980164,
0.06565459817647934,
0.0406544953584671,
0.03348100185394287,
-0.0012982190819457173,
0.08662453293800354,
-0.05801142007112503,
0.0003153103752993047,
0.041379209607839584,
0.07964122295379639,
0.047912634909152985,
-0.022471608594059944,
0.06123734265565872,
-0.05732676014304161,
0.0483168363571167,
0.05743001401424408,
0.0422637052834034,
0.07255646586418152,
0.13181057572364807,
-0.0427628830075264,
0.013098495081067085,
0.03531244024634361,
0.04116564616560936,
0.03792145848274231,
-0.03560796007514,
-0.021445628255605698,
0.026460440829396248,
0.017503036186099052,
0.022638026624917984,
0.010269660502672195,
0.04967775195837021,
-0.07753264158964157,
-0.05817834660410881,
0.06472702324390411,
0.002185577992349863,
0.06425900012254715,
-0.014633982442319393,
-0.03975917398929596,
0.002972730901092291,
-0.0551140159368515,
0.05871497094631195,
0.0398171991109848,
-0.06395655125379562,
-0.12099450826644897,
0.11389302462339401,
0.056697916239500046,
0.05161627382040024,
-0.06158151850104332,
-0.014963377267122269,
-0.06339247524738312,
-0.036356959491968155,
-0.07590469717979431,
0.035795655101537704,
-0.025378171354532242,
-0.10363800823688507,
0.02479070983827114,
-0.0036793332546949387,
0.1213851049542427,
-0.0047491746954619884,
-0.09757997840642929,
-0.07317483425140381,
5.21757971494453e-33,
-0.0065676807425916195,
0.01527958270162344,
0.02316134236752987,
0.032899435609579086,
0.0911189541220665,
0.015473115257918835,
-0.05864904448390007,
0.05525103583931923,
-0.04459688067436218,
0.05634496361017227,
-0.03402943164110184,
-0.047333840280771255,
-0.1398654729127884,
0.07116592675447464,
0.11415364593267441,
0.0195626188069582,
-0.07450100034475327,
0.05371566116809845,
-0.08319836854934692,
-0.016195932403206825,
-0.04104088619351387,
-0.09120591729879379,
-0.06522087007761002,
0.05590524896979332,
0.013843200169503689,
0.025663524866104126,
0.0237808246165514,
0.022397175431251526,
0.011783395893871784,
-0.02642318420112133,
-0.0953194722533226,
-0.05784262716770172,
0.11832430213689804,
0.02183939330279827,
-0.027722597122192383,
0.055251337587833405,
-0.037844959646463394,
-0.017531054094433784,
0.005073556210845709,
0.03817395493388176,
0.032107602804899216,
0.07101286947727203,
0.00767560163512826,
0.03994593769311905,
-0.05597103387117386,
-0.07282788306474686,
0.06376943737268448,
0.04580343887209892,
-0.0531073734164238,
-0.05328666791319847,
0.03898267447948456,
0.060347024351358414,
0.04015453904867172,
-0.05719780921936035,
-0.030320949852466583,
-0.055051229894161224,
-0.005987694952636957,
0.006143555976450443,
0.030667567625641823,
0.08867576718330383,
-0.03688811510801315,
-0.006812483537942171,
-0.03811010345816612,
0.056234344840049744,
-0.03992200270295143,
0.016383487731218338,
0.008590840734541416,
0.04232284054160118,
0.08287501335144043,
-0.0008368062553927302,
-0.08007572591304779,
-0.018942128866910934,
-0.03736230731010437,
-0.04164976254105568,
0.03854280337691307,
0.034370072185993195,
0.03299233689904213,
-0.021462390199303627,
0.02543235756456852,
0.012933134101331234,
0.03321642056107521,
0.12237212806940079,
-0.015427573584020138,
0.02671104483306408,
-0.048059314489364624,
-0.06350491940975189,
0.03271333500742912,
-0.03596663475036621,
-0.012756656855344772,
0.016453133895993233,
-0.056127943098545074,
-0.05328657105565071,
-0.03634491562843323,
0.024853890761733055,
0.030127694830298424,
-4.2247873082827656e-33,
-0.059514887630939484,
0.032284896820783615,
-0.0454106330871582,
0.016909202560782433,
-0.01059713028371334,
-0.023210981860756874,
-0.042677007615566254,
-0.011368812993168831,
0.06524310261011124,
-0.08792372792959213,
-0.012692267075181007,
-0.08623174577951431,
0.038017407059669495,
-0.00841809157282114,
-0.03891206905245781,
0.016621839255094528,
-0.10842831432819366,
0.018227821215987206,
-0.058532144874334335,
0.035678714513778687,
0.0662599429488182,
0.08867926150560379,
-0.06697440892457962,
0.03239056095480919,
-0.017887236550450325,
0.04194718599319458,
-0.046315036714076996,
0.08737824112176895,
-0.013865665532648563,
0.06295470893383026,
0.003317134454846382,
-0.056311652064323425,
-0.07298393547534943,
-0.06682674586772919,
-0.02268056571483612,
0.003535724012181163,
0.01880173198878765,
0.025556357577443123,
-0.018069300800561905,
0.06072141230106354,
0.04643867164850235,
-0.023155905306339264,
-0.09155783802270889,
0.05932745337486267,
0.0007308379863388836,
0.03810993954539299,
0.023901892825961113,
-0.05092243477702141,
0.04653936251997948,
0.05424134060740471,
-0.031500183045864105,
0.01576799340546131,
-0.046004537492990494,
-0.0012842308497056365,
-0.03573407232761383,
-0.09519778937101364,
-0.10594425350427628,
0.0004072762094438076,
0.11140633374452591,
-0.0033510494977235794,
-0.09196675568819046,
-0.0839783102273941,
0.040285006165504456,
-0.036581333726644516,
0.01324336789548397,
-0.05418214574456215,
0.008227714337408543,
0.024881407618522644,
0.0053314813412725925,
0.011310771107673645,
0.02854381687939167,
-0.039772115647792816,
-0.004321719985455275,
0.06643728166818619,
-0.026369886472821236,
-0.0070830280892550945,
0.028619127348065376,
-0.009106063283979893,
0.01203157752752304,
-0.031153544783592224,
0.07137591391801834,
0.016710761934518814,
0.09112237393856049,
0.012603440321981907,
0.0497913658618927,
0.03455231338739395,
0.0676645040512085,
0.054803408682346344,
0.01400727964937687,
0.08821874111890793,
-0.0656251311302185,
-0.027561603114008904,
-0.06345275044441223,
0.0460851676762104,
-0.032317861914634705,
-4.6613049420329844e-8,
-0.04075567424297333,
0.10987237840890884,
-0.0039484528824687,
-0.001034602290019393,
0.012479162774980068,
-0.03929608687758446,
0.010641590692102909,
0.06030059978365898,
-0.011577854864299297,
-0.018415404483675957,
0.07458093017339706,
0.0024269884452223778,
-0.04288175702095032,
-0.05619693547487259,
-0.026584554463624954,
0.09022349119186401,
0.003908954560756683,
-0.012844567187130451,
0.004751840140670538,
0.003641013987362385,
0.010125726461410522,
0.08003220707178116,
-0.054274775087833405,
0.004998679272830486,
0.06364668160676956,
-0.07800453901290894,
-0.011194084770977497,
0.03731885924935341,
0.03386981040239334,
0.0016700313426554203,
0.05391903594136238,
-0.029716428369283676,
0.08117149025201797,
-0.029345057904720306,
-0.02333071455359459,
0.08739995956420898,
-0.03604508191347122,
-0.043812479823827744,
-0.04307999089360237,
-0.10758597403764725,
-0.004911255091428757,
0.1197132095694542,
-0.04783052206039429,
0.003555885748937726,
0.009616331197321415,
0.032921891659498215,
0.006381395272910595,
-0.008453572168946266,
0.011816529557108879,
0.0505843348801136,
-0.0031010680831968784,
-0.0049721356481313705,
0.06319011747837067,
-0.024081604555249214,
0.015714609995484352,
0.03737847879528999,
0.020458858460187912,
-0.010086561553180218,
0.02752213180065155,
-0.005479973275214434,
-0.011448047123849392,
0.04457838088274002,
-0.01450157817453146,
-0.03960174694657326
] |
Large Scale Knowledge Washing | https://openreview.net/forum?id=dXCpPgjTtd | [
"Yu Wang",
"Ruihan Wu",
"Zexue He",
"Xiusi Chen",
"Julian McAuley"
] | Poster | Large language models show impressive abilities in memorizing world knowledge, which leads to concerns regarding memorization of private information, toxic or sensitive knowledge, and copyrighted content. We introduce the problem of Large Scale Knowledge Washing, focusing on unlearning an extensive amount of factual knowledge. Previous unlearning methods usually define the reverse loss and update the model via backpropagation, which may affect the model's fluency and reasoning ability or even destroy the model due to extensive training with the reverse loss. Existing works introduce additional data from downstream tasks to prevent the model from losing capabilities, which requires downstream task awareness. Controlling the tradeoff of unlearning existing knowledge while maintaining existing capabilities is also challenging. To this end, we propose LaW (Large Scale Washing), where we update the MLP layers in decoder-only large language models to perform knowledge washing, as inspired by model editing methods. We derive a new objective with the knowledge to be unlearned to update the weights of certain MLP layers. Experimental results demonstrate the effectiveness of LaW in forgetting target knowledge while maximally maintaining reasoning ability. The code will be open-sourced. | knowledge unlearning, large language models | Washing the knowledge in large language models in a large scale | 740 | 2405.16720 | [
-0.06145568564534187,
-0.08454644680023193,
0.0041997600346803665,
0.0503668375313282,
0.054710108786821365,
-0.009023631922900677,
-0.06911390274763107,
-0.03154950588941574,
-0.008115110918879509,
0.037951093167066574,
-0.006638890132308006,
0.015630023553967476,
0.06544940918684006,
-0.0005999479908496141,
-0.006717343349009752,
0.05006478726863861,
0.0577688105404377,
0.13476119935512543,
-0.13329502940177917,
-0.08286141604185104,
0.014051157049834728,
0.022354260087013245,
0.03523964062333107,
0.0565970204770565,
0.01390878576785326,
-0.047923844307661057,
0.015600083395838737,
-0.04866966977715492,
0.038763634860515594,
-0.05371677130460739,
0.07958230376243591,
0.02205405756831169,
0.0077810161747038364,
0.04076038673520088,
-0.022457577288150787,
0.06984596699476242,
-0.04280022159218788,
0.02264861762523651,
0.008641982451081276,
-0.049081623554229736,
-0.005666190292686224,
-0.04039385914802551,
-0.08564849942922592,
0.00842321291565895,
0.02835148386657238,
0.03352421894669533,
0.04232647642493248,
0.013647706247866154,
0.010396643541753292,
0.03541603311896324,
-0.028832953423261642,
-0.039129726588726044,
-0.03872891888022423,
0.01195528358221054,
-0.009782159700989723,
-0.05588605999946594,
0.0471024215221405,
-0.0010025096125900745,
-0.09440463781356812,
-0.059347327798604965,
-0.05505993962287903,
-0.04990160092711449,
-0.050369538366794586,
-0.04822700098156929,
-0.03569716587662697,
-0.024081094190478325,
-0.01011302974075079,
0.11232331395149231,
-0.023871829733252525,
0.032819949090480804,
0.01912532187998295,
0.07248491048812866,
-0.03270392864942551,
0.04015825688838959,
0.018538912758231163,
0.02034860849380493,
0.002800958463922143,
-0.03264070674777031,
0.010695422068238258,
-0.07228381931781769,
0.02322910539805889,
0.03093196637928486,
0.015650451183319092,
0.00831456296145916,
0.057005878537893295,
0.002644079737365246,
0.055845439434051514,
0.04316972568631172,
0.11602135002613068,
0.011732716113328934,
-0.009456980973482132,
-0.04023069515824318,
0.12306036800146103,
0.009120810776948929,
0.057132527232170105,
0.03782288357615471,
0.005222754552960396,
-0.041634079068899155,
-0.04294801875948906,
0.08984222263097763,
-0.0288305152207613,
0.10292793810367584,
0.0016959833446890116,
-0.05777404457330704,
0.045077256858348846,
0.0001065065007423982,
0.02741801179945469,
0.08154833316802979,
0.01679350808262825,
-0.07785904407501221,
0.05408760532736778,
0.08399111032485962,
0.009030966088175774,
-0.01852698065340519,
0.0832669660449028,
-0.011750653386116028,
0.0073897927068173885,
0.019388480111956596,
-0.051085464656353,
0.08526939153671265,
-0.0033116613049060106,
0.05213459953665733,
0.015548030845820904,
0.06986799091100693,
-0.057474080473184586,
-0.04097001627087593,
-0.07692927122116089,
5.758296830747212e-34,
0.07615640014410019,
0.09759867936372757,
-0.007064477074891329,
0.020105179399251938,
0.042784761637449265,
0.038523584604263306,
0.02439400926232338,
0.05594807118177414,
-0.05345051735639572,
-0.024927403777837753,
0.016841942444443703,
-0.008146463893353939,
-0.06621893495321274,
0.03241562843322754,
0.06842545419931412,
0.03023313730955124,
-0.04412408918142319,
0.06840268522500992,
-0.002852465957403183,
-0.06793765723705292,
0.10981419682502747,
-0.050793107599020004,
-0.018123531714081764,
-0.06990572810173035,
-0.0783272385597229,
-0.01631365530192852,
0.04182744398713112,
-0.03967062756419182,
0.017625275999307632,
0.022237282246351242,
-0.031548622995615005,
-0.0482894666492939,
0.020562225952744484,
0.04007982835173607,
-0.0026668026112020016,
-0.056147366762161255,
-0.04433169215917587,
-0.09756156802177429,
0.09118790924549103,
-0.050242628902196884,
-0.015971219167113304,
0.024686483666300774,
-0.024172481149435043,
0.015362417325377464,
-0.07841609418392181,
-0.08582376688718796,
-0.03792143240571022,
-0.004921541083604097,
-0.02565082162618637,
-0.06859029084444046,
0.12562444806098938,
0.033787403255701065,
-0.048590660095214844,
-0.09321685135364532,
-0.05149155482649803,
-0.01345806010067463,
-0.0014888477744534612,
0.004442053381353617,
0.028243936598300934,
0.04897093400359154,
0.007881173864006996,
0.061348725110292435,
0.012343620881438255,
0.1776880919933319,
0.042149823158979416,
0.032886579632759094,
-0.03131377324461937,
0.047430600970983505,
0.06766480207443237,
-0.09902580082416534,
-0.06427491456270218,
-0.023602167144417763,
-0.02776356413960457,
-0.04608076065778732,
-0.04448791220784187,
-0.07307978719472885,
0.010152646340429783,
-0.10682246834039688,
0.054286327213048935,
0.009188381023705006,
0.027361247688531876,
-0.0010730642825365067,
-0.025037258863449097,
-0.0023409512359648943,
-0.036792196333408356,
-0.046269483864307404,
0.021720699965953827,
-0.0649838075041771,
0.018891122192144394,
0.055689819157123566,
-0.014021039940416813,
-0.09930375963449478,
-0.029701070860028267,
-0.037994202226400375,
0.003712961683049798,
-6.7376025519497024e-34,
-0.03449229523539543,
-0.04636469483375549,
-0.08143199980258942,
0.060454580932855606,
-0.06880932301282883,
-0.07921001315116882,
-0.038403913378715515,
0.09694437682628632,
-0.07570680975914001,
-0.04087316244840622,
-0.019172802567481995,
-0.09212014079093933,
-0.023309575393795967,
-0.022769588977098465,
0.03934415057301521,
-0.02483978681266308,
0.028965990990400314,
0.03270494192838669,
-0.058912839740514755,
0.05179397389292717,
-0.04408947378396988,
0.051564447581768036,
-0.04232605919241905,
0.09331489354372025,
0.017316965386271477,
-0.019803347066044807,
-0.125748410820961,
0.06295488774776459,
0.06794600188732147,
0.0371454618871212,
-0.06215377151966095,
0.06622432172298431,
-0.036129698157310486,
-0.056691449135541916,
-0.0935753583908081,
-0.008606585673987865,
0.04652930796146393,
0.00530404644086957,
-0.02378571592271328,
0.04493892937898636,
0.08673594892024994,
-0.037104930728673935,
-0.07627279311418533,
-0.08647465705871582,
-0.0031815827824175358,
-0.04651657119393349,
-0.13071155548095703,
-0.0235117357224226,
0.05004914849996567,
-0.021127672865986824,
0.020493727177381516,
-0.033272262662649155,
-0.0058414870873093605,
-0.000597381207626313,
0.032357439398765564,
-0.04168405011296272,
0.0132460230961442,
-0.02032584324479103,
0.045603878796100616,
-0.009184799157083035,
-0.034832265228033066,
-0.009412495419383049,
-0.0028107191901654005,
-0.01565387286245823,
0.004113075323402882,
-0.042116373777389526,
-0.03336286544799805,
0.08409082889556885,
0.0028490331023931503,
-0.006464706268161535,
0.031119057908654213,
-0.03552921861410141,
-0.019605491310358047,
-0.05436789616942406,
0.01584060862660408,
-0.04754497483372688,
-0.01622648909687996,
-0.002360708313062787,
-0.0476890429854393,
-0.032468345016241074,
0.0227748304605484,
-0.03146975114941597,
0.07095053046941757,
0.037160053849220276,
0.09768952429294586,
-0.00843384675681591,
0.03177943080663681,
0.08176372200250626,
-0.019665012136101723,
0.04868366941809654,
-0.04699324816465378,
-0.030578957870602608,
-0.02800649031996727,
0.05265500396490097,
-0.047910504043102264,
-4.057317326555676e-8,
-0.06738357990980148,
0.039024703204631805,
0.04974968358874321,
0.08469754457473755,
0.026303831487894058,
-0.04913339018821716,
-0.08543901890516281,
0.0785914808511734,
-0.0015372157795354724,
-0.03228897601366043,
0.07104196399450302,
-0.019265640527009964,
-0.03418998792767525,
-0.0353044718503952,
-0.03830178827047348,
0.12473368644714355,
0.025746852159500122,
0.0006046959315426648,
0.021072406321763992,
0.0026200865395367146,
0.02193000540137291,
0.0040513500571250916,
0.004866460338234901,
0.05631455406546593,
0.039795517921447754,
-0.06285101920366287,
0.002648591063916683,
0.08073963969945908,
0.09790412336587906,
0.06837669014930725,
0.009375829249620438,
-0.0717504546046257,
0.03223482891917229,
0.025169072672724724,
0.07615148276090622,
0.08121632039546967,
0.03153878077864647,
-0.014921692200005054,
-0.04666145145893097,
0.027105918154120445,
0.06801984459161758,
0.08040529489517212,
0.00012136001168983057,
0.044709980487823486,
0.05030561238527298,
-0.03341289982199669,
-0.03112424537539482,
-0.088860422372818,
0.019673442468047142,
-0.0077956197783350945,
0.03607858344912529,
0.03895196318626404,
0.0019509142730385065,
0.04211215674877167,
0.025573665276169777,
0.08515007048845291,
-0.01178684737533331,
0.030864302068948746,
-0.01375541090965271,
0.02732493355870247,
0.05561896413564682,
0.018176613375544548,
0.000447128462838009,
0.006004678551107645
] |
Self-Updatable Large Language Models by Integrating Context into Model Parameters | https://openreview.net/forum?id=aCPFCDL9QY | [
"Yu Wang",
"Xinshuang Liu",
"Xiusi Chen",
"Sean O'Brien",
"Junda Wu",
"Julian McAuley"
] | Poster | Despite significant advancements in large language models (LLMs), the rapid and frequent integration of small-scale experiences, such as interactions with sur- rounding objects, remains a substantial challenge. Two critical factors in assimilating these experiences are (1) **Efficacy**: the ability to accurately remember recent events; (2) **Retention**: the capacity to recall long-past experiences. Current methods either embed experiences within model parameters using continual learning, model editing, or knowledge distillation techniques, which often struggle with rapid updates and complex interactions, or rely on external storage to achieve long-term retention, thereby increasing storage requirements. In this paper, we propose **SELF-PARAM** (Self-Updatable Large Language Models with Parameter Integration). SELF-PARAM requires no extra parameters while ensuring near-optimal efficacy and long-term retention. Our method employs a training objective that minimizes the Kullback-Leibler (KL) divergence between the predictions of an original model (with access to contextual information) and a target model (without such access). By generating diverse question-answer pairs related to the knowledge and minimizing the KL divergence across this dataset, we update the target model to internalize the knowledge seamlessly within its parameters. Evaluations on question-answering and conversational recommendation tasks demonstrate that SELF-PARAM significantly outperforms existing methods, even when accounting for non-zero storage requirements. This advancement paves the way for more efficient and scalable integration of experiences in large language models by embedding knowledge directly into model parameters. | Large Language Models, Knowledge Injection, Self-Updatable LLMs | This paper proposes a simple method to absorb contexts of any form into model parameters. | 738 | 2410.00487 | [
0.016638850793242455,
-0.061155132949352264,
0.0022092682775110006,
0.08952773362398148,
-0.00806002039462328,
0.06732949614524841,
-0.015406095422804356,
0.04438503831624985,
0.06904654204845428,
-0.05684962868690491,
-0.02101474441587925,
0.011755921877920628,
0.04239499568939209,
0.004700689110904932,
-0.0026578716933727264,
0.008962271735072136,
0.04370005056262016,
0.070917047560215,
-0.1406073123216629,
-0.04810740053653717,
0.02210880257189274,
0.00698334164917469,
0.05532928183674812,
0.03342354670166969,
0.023025749251246452,
-0.024384744465351105,
0.003522463608533144,
-0.030815279111266136,
0.05919812619686127,
-0.01306886039674282,
0.049030374735593796,
0.05550721660256386,
-0.05606172978878021,
0.040601521730422974,
0.0024375177454203367,
0.032616883516311646,
-0.08791831880807877,
0.03831963241100311,
0.027349237352609634,
-0.037548668682575226,
-0.02165195345878601,
-0.009569087065756321,
-0.0017470030579715967,
0.03989556059241295,
0.09649275988340378,
-0.05192437767982483,
-0.03391094133257866,
-0.005250662565231323,
0.0041922894306480885,
0.061803460121154785,
-0.06353027373552322,
-0.03565463796257973,
-0.050238195806741714,
0.05777418985962868,
0.0023383735679090023,
-0.004260885063558817,
-0.015288542956113815,
0.06490866094827652,
-0.034485332667827606,
-0.03893143683671951,
-0.05846748501062393,
-0.12323411554098129,
-0.03975969925522804,
0.006816815119236708,
-0.12438970059156418,
-0.01928633078932762,
0.010963315144181252,
0.0684899240732193,
-0.03472986817359924,
0.09051137417554855,
-0.03555617853999138,
0.06334660202264786,
-0.05413822457194328,
0.07192785292863846,
0.05471302941441536,
0.017674734815955162,
0.028203239664435387,
-0.047766298055648804,
0.03429902717471123,
0.004092163406312466,
0.018522335216403008,
0.01172896008938551,
0.03734442591667175,
-0.050590865314006805,
0.08160936832427979,
0.012372057884931564,
0.08462315052747726,
0.04259538650512695,
0.02048538625240326,
-0.03640339523553848,
0.01998627372086048,
-0.049733150750398636,
0.03343242034316063,
-0.03495366871356964,
0.048840731382369995,
0.03934112936258316,
0.0364723801612854,
-0.05585227534174919,
-0.04633432626724243,
0.0553797110915184,
0.037457384169101715,
0.11198591440916061,
-0.01902911812067032,
-0.03363009914755821,
-0.05640025436878204,
-0.0389443002641201,
-0.01529886294156313,
0.06312175840139389,
0.004381902515888214,
-0.05899060517549515,
-0.027665583416819572,
0.05418015643954277,
-0.012026341632008553,
0.01785757765173912,
0.006916445679962635,
-0.0026004223618656397,
0.04567977041006088,
-0.052292171865701675,
0.021276358515024185,
0.06938815116882324,
-0.0021115944255143404,
0.04163794964551926,
0.023959195241332054,
0.017030686140060425,
-0.07218749076128006,
-0.038022030144929886,
-0.005958326160907745,
-4.684264172790197e-34,
0.05877521261572838,
0.0578431598842144,
-0.006965941283851862,
0.041839756071567535,
-0.008628877811133862,
0.01688423752784729,
-0.0038528828881680965,
0.03980465233325958,
-0.026844192296266556,
-0.057628877460956573,
0.007433341350406408,
0.10172467678785324,
-0.04837166517972946,
0.018589571118354797,
0.09580573439598083,
0.059201717376708984,
-0.10146086663007736,
0.1268167495727539,
-0.018558930605649948,
-0.06295179575681686,
0.04968233034014702,
0.0006678622448816895,
0.03287957236170769,
-0.03883441165089607,
0.04035303369164467,
0.03731442987918854,
0.11957306414842606,
0.002389806555584073,
-0.04755116254091263,
0.01822633482515812,
-0.08583895117044449,
-0.036238692700862885,
-0.03012095019221306,
-0.03314526751637459,
0.041821207851171494,
-0.028634244576096535,
0.011368406936526299,
-0.07309433817863464,
0.015225931070744991,
-0.07925008982419968,
-0.05361029878258705,
0.03132367506623268,
0.011685743927955627,
0.009002325125038624,
-0.1311519742012024,
-0.13490261137485504,
0.03686548396945,
-0.005045664496719837,
-0.08360549807548523,
-0.05098022148013115,
0.009452726691961288,
0.008775507099926472,
-0.08566395938396454,
-0.03897188603878021,
-0.04624698683619499,
0.018400011584162712,
0.02283954992890358,
0.017496077343821526,
0.010156868025660515,
0.05994771048426628,
0.03904806077480316,
0.020713455975055695,
0.027301864698529243,
0.13160672783851624,
0.07487662136554718,
0.04796489700675011,
-0.0542307011783123,
0.09001440554857254,
0.06761044263839722,
-0.05097905546426773,
-0.014106232672929764,
-0.03872637078166008,
-0.04027222469449043,
-0.08424152433872223,
0.021356496959924698,
-0.06339147686958313,
0.03918223828077316,
-0.07340764999389648,
-0.0044759041629731655,
0.005156193859875202,
0.07322944700717926,
-0.09672226011753082,
0.009760439395904541,
-0.05668075010180473,
-0.022117309272289276,
-0.04037949815392494,
0.06641839444637299,
-0.07455804944038391,
-0.0345696322619915,
-0.033624328672885895,
-0.033537715673446655,
0.014925867319107056,
0.018468812108039856,
-0.035987239331007004,
0.012552384287118912,
2.749028078399791e-34,
0.021682944148778915,
-0.046261101961135864,
-0.03297019377350807,
0.06539219617843628,
0.02592587284743786,
-0.05699346214532852,
-0.030835110694169998,
0.09287945926189423,
-0.016144972294569016,
-0.14945298433303833,
-0.07081520557403564,
-0.03653600066900253,
0.09147639572620392,
0.027236590161919594,
-0.02203788235783577,
0.01633957214653492,
-0.016264289617538452,
0.02279140241444111,
-0.02329663746058941,
0.028271278366446495,
-0.053709521889686584,
-0.017628218978643417,
-0.045654479414224625,
0.033013034611940384,
-0.05897195264697075,
0.01056875940412283,
-0.06086581572890282,
0.04028545320034027,
-0.05629339814186096,
0.021747827529907227,
-0.07551147788763046,
-0.037479426711797714,
-0.008604354225099087,
-0.05523121729493141,
-0.07297523319721222,
0.02023519016802311,
0.05354388803243637,
-0.059692367911338806,
-0.03976763039827347,
0.08572950214147568,
0.08212345838546753,
0.0069403997622430325,
-0.06766613572835922,
-0.051550690084695816,
-0.030018316581845284,
-0.053607258945703506,
-0.12130539119243622,
0.031712811440229416,
0.07464630901813507,
-0.0013797493884339929,
-0.03417832404375076,
0.015250577591359615,
-0.057204470038414,
-0.004068886395543814,
-0.029319027438759804,
-0.014299233444035053,
-0.030373627319931984,
-0.06494290381669998,
0.022869424894452095,
-0.01973382569849491,
-0.06739446520805359,
-0.019323376938700676,
0.072417251765728,
-0.09614836424589157,
0.027230268344283104,
-0.05557210370898247,
-0.014399752020835876,
-0.00042878606473095715,
-0.042353305965662,
-0.02000850811600685,
-0.006570437923073769,
0.020486561581492424,
0.01309246476739645,
0.0320315957069397,
-0.02683039754629135,
-0.0538003146648407,
-0.012662074528634548,
-0.04398127272725105,
0.01603081449866295,
-0.06280054152011871,
-0.03555365651845932,
0.03286818787455559,
-0.0038235930260270834,
0.0511360727250576,
0.03346312791109085,
0.06825505197048187,
0.023045016452670097,
0.13791237771511078,
0.006710370536893606,
0.014836929738521576,
-0.025336768478155136,
0.023625832051038742,
-0.0032745227217674255,
0.12223859131336212,
-0.0458373948931694,
-4.330190606083306e-8,
-0.0609552226960659,
0.09635566920042038,
0.007554761599749327,
0.10119292885065079,
0.06428281962871552,
-0.028186369687318802,
-0.006937166675925255,
0.08351880311965942,
-0.013193704187870026,
-0.024758238345384598,
0.051039621233940125,
0.051384225487709045,
-0.04410642012953758,
-0.03569212555885315,
-0.035062532871961594,
0.09631341695785522,
0.05377746745944023,
0.013959230855107307,
0.006779938004910946,
-0.016339829191565514,
0.0866936594247818,
0.003977742977440357,
0.02633856050670147,
0.01285921037197113,
0.06315042823553085,
0.028171256184577942,
0.0018080363515764475,
0.15413245558738708,
0.05796051397919655,
0.0010727145709097385,
0.0009199815685860813,
-0.024663619697093964,
-0.058764565736055374,
-0.03184381127357483,
-0.04129352793097496,
0.04214097559452057,
0.07157128304243088,
-0.04506739601492882,
0.010434326715767384,
-0.02953995019197464,
0.08317001163959503,
0.02137773111462593,
-0.04967057704925537,
0.051255688071250916,
0.043550800532102585,
0.034218233078718185,
-0.009465641342103481,
-0.08656477928161621,
0.02466452307999134,
0.006204077508300543,
-0.008478745818138123,
-0.05867091566324234,
-0.007132211234420538,
0.0763469934463501,
0.05333777517080307,
0.03372831642627716,
-0.023278314620256424,
0.04782234877347946,
0.015615857206285,
-0.00735113862901926,
0.03477255627512932,
0.023215554654598236,
-0.04071108251810074,
0.08650051057338715
] |
Process Reward Model with Q-value Rankings | https://openreview.net/forum?id=wQEdh2cgEk | [
"Wendi Li",
"Yixuan Li"
] | Poster | Process Reward Modeling (PRM) is critical for complex reasoning and decision-making tasks where the accuracy of intermediate steps significantly influences the overall outcome. Existing PRM approaches, primarily framed as classification problems, employ cross-entropy loss to independently evaluate each step's correctness. This method can lead to suboptimal reward distribution and does not adequately address the interdependencies among steps. To address these limitations, we introduce the Process Q-value Model (PQM), a novel framework that redefines PRM in the context of a Markov Decision Process. PQM optimizes Q-value rankings based on a novel comparative loss function, enhancing the model's ability to capture the intricate dynamics among sequential decisions. This approach provides a more granular and theoretically grounded methodology for process rewards. Our extensive empirical evaluations across various sampling policies, language model backbones, and multi-step reasoning benchmarks show that PQM outperforms classification-based PRMs. The effectiveness of the comparative loss function is highlighted in our comprehensive ablation studies, confirming PQM’s practical efficacy and theoretical advantage. | process reward model, reasoning | null | 730 | null | [
-0.09075044840574265,
-0.01966584101319313,
-0.00805167481303215,
-0.03886830434203148,
-0.018767880275845528,
0.04535933956503868,
0.05035075545310974,
0.042404789477586746,
0.08719008415937424,
0.0635436400771141,
-0.09586763381958008,
-0.06006300821900368,
0.07007044553756714,
0.09968122839927673,
0.03152598440647125,
0.007097533904016018,
0.015126784332096577,
-0.025563737377524376,
-0.020401692017912865,
-0.07950253784656525,
0.0027975088451057673,
0.019598856568336487,
0.07752061635255814,
0.020875979214906693,
-0.014975309371948242,
-0.04607035592198372,
0.013237378560006618,
0.05120430886745453,
0.01881134696304798,
-0.026197094470262527,
0.010734017938375473,
-0.0040467954240739346,
0.05490770563483238,
-0.10176631063222885,
-0.03375522419810295,
0.06664843112230301,
-0.13894124329090118,
-0.08312536776065826,
-0.048421457409858704,
0.01246323436498642,
-0.01898149959743023,
-0.03993723541498184,
-0.010867627337574959,
0.05480991676449776,
0.06006637215614319,
0.00927817728370428,
-0.008525271899998188,
-0.03095104731619358,
-0.04498361051082611,
-0.00991163868457079,
-0.023630265146493912,
0.031623534858226776,
-0.017568159848451614,
0.07072755694389343,
-0.04105532914400101,
0.06876841932535172,
0.04009326547384262,
-0.027165884152054787,
-0.050433967262506485,
-0.06764304637908936,
-0.0955950990319252,
-0.12527577579021454,
-0.03750928118824959,
-0.0022568334825336933,
-0.03273544833064079,
0.000929716625250876,
-0.01089696679264307,
0.023605456575751305,
-0.04280567169189453,
0.001135467435233295,
-0.027180388569831848,
-0.0013468838296830654,
-0.013824818655848503,
0.00547365378588438,
0.002306200796738267,
0.04274410009384155,
-0.00463560875505209,
0.0023488050792366266,
-0.08612366765737534,
-0.001597569789737463,
-0.036087766289711,
-0.0002622341853566468,
-0.00204125652089715,
-0.017176596447825432,
0.1512574851512909,
-0.011680973693728447,
0.03632289543747902,
0.004772020038217306,
0.04921581596136093,
0.010060780681669712,
-0.029609939083456993,
0.0060216584242880344,
0.041630905121564865,
-0.018505731597542763,
0.015806078910827637,
0.0911598950624466,
-0.055160582065582275,
-0.06499726325273514,
-0.0003797273675445467,
0.013388796709477901,
0.02840210124850273,
0.11224779486656189,
-0.054873306304216385,
-0.06145726516842842,
0.1146971732378006,
-0.06659483164548874,
-0.04237167164683342,
0.04296502470970154,
0.113680899143219,
-0.10943753272294998,
-0.04949456825852394,
0.029649311676621437,
0.09211188554763794,
0.044289808720350266,
-0.01565624587237835,
0.07104802876710892,
0.040071483701467514,
0.005759902764111757,
0.012942650355398655,
0.038526248186826706,
0.0003283550904598087,
-0.012724384665489197,
0.019100919365882874,
-0.001587758306413889,
-0.020756833255290985,
0.01905851438641548,
-0.01724434085190296,
2.9677926280401072e-33,
0.030450761318206787,
-0.012703681364655495,
0.10011377185583115,
-0.009385207667946815,
0.061877794563770294,
0.005184316076338291,
0.04575987160205841,
0.004338504280894995,
-0.017417892813682556,
-0.04751014709472656,
-0.031512003391981125,
0.04755993187427521,
-0.018166687339544296,
0.049272019416093826,
0.023253802210092545,
-0.02382498048245907,
-0.12661013007164001,
0.05431709438562393,
0.00976231787353754,
-0.07247927039861679,
0.04965401068329811,
-0.0570233017206192,
0.005778511054813862,
-0.006255400832742453,
0.040628623217344284,
0.09012608230113983,
-0.009053695015609264,
-0.012183512561023235,
-0.03672757372260094,
-0.04058727249503136,
-0.033350877463817596,
0.004034156445413828,
0.03754571080207825,
-0.10682471096515656,
-0.035640135407447815,
0.0020150879863649607,
-0.04416352137923241,
0.022137142717838287,
0.03762771934270859,
-0.06382062286138535,
-0.11835627257823944,
-0.06102747470140457,
-0.006582103203982115,
0.0648944154381752,
-0.09348125010728836,
-0.05769683048129082,
-0.010731622576713562,
-0.05374656990170479,
-0.03964123874902725,
0.007239700760692358,
0.06579554826021194,
-0.04526036977767944,
0.11323841661214828,
-0.008227099664509296,
-0.0675034150481224,
0.013650142587721348,
0.06348594278097153,
-0.014806576073169708,
-0.08386754989624023,
0.009385996498167515,
0.04047086462378502,
0.053462885320186615,
-0.02553209476172924,
0.05289319530129433,
-0.02524319291114807,
0.03049667738378048,
0.0016496864845976233,
0.027702342718839645,
0.07886145263910294,
0.0688023492693901,
0.04321396350860596,
0.05557848513126373,
0.041849829256534576,
-0.06499232351779938,
0.08201141655445099,
0.02238757722079754,
0.02505745179951191,
-0.05028359964489937,
0.043385833501815796,
-0.049470700323581696,
0.006477338261902332,
-0.03579891473054886,
-0.02463967353105545,
-0.02453736402094364,
0.06738902628421783,
-0.02668345719575882,
-0.0005091413040645421,
-0.0996258407831192,
-0.03051229752600193,
0.027445824816823006,
-0.04411979392170906,
0.04270116239786148,
0.050287019461393356,
0.07018277049064636,
-0.00805858988314867,
-3.184166607879935e-33,
-0.032650694251060486,
-0.027540257200598717,
0.0077503761276602745,
0.02305179089307785,
0.009266282431781292,
-0.06586399674415588,
-0.02498566545546055,
-0.05748875066637993,
0.030555006116628647,
-0.017917107790708542,
-0.03339853510260582,
0.02969319000840187,
0.022817740216851234,
0.0482097864151001,
0.023958148434758186,
-0.007570259738713503,
-0.02291278913617134,
0.05294281989336014,
0.019687239080667496,
0.03962801769375801,
0.06978528946638107,
0.1740194857120514,
-0.16117234528064728,
0.043517835438251495,
-0.033171165734529495,
0.016237134113907814,
0.0962456613779068,
-0.019381791353225708,
0.03889175131917,
-0.030038131400942802,
-0.0632404312491417,
-0.11433141678571701,
-0.025697967037558556,
-0.016282640397548676,
-0.023058779537677765,
-0.022844061255455017,
0.08363527804613113,
-0.07548331469297409,
-0.0755046084523201,
0.09709667414426804,
0.0015009701019153,
-0.01703864522278309,
-0.09764578193426132,
-0.00048056195373646915,
-0.006815725471824408,
-0.00901787355542183,
-0.039138320833444595,
-0.020780133083462715,
-0.022001728415489197,
-0.001255164504982531,
-0.006856320425868034,
0.019610021263360977,
-0.05840625613927841,
0.14024996757507324,
-0.04464826360344887,
-0.03193281963467598,
-0.09479264914989471,
-0.05303611978888512,
0.009887071326375008,
0.0682646632194519,
-0.040343523025512695,
0.08976545929908752,
0.07893471419811249,
-0.03515307232737541,
0.02420353703200817,
0.035660531371831894,
0.030721241608262062,
0.019344279542565346,
-0.029733426868915558,
-0.0015463522868230939,
0.015002621337771416,
0.07516469061374664,
0.05285801738500595,
0.009855597279965878,
0.06551303714513779,
-0.018614178523421288,
-0.0388263463973999,
-0.029825374484062195,
-0.061646994203329086,
-0.039063919335603714,
-0.010264778509736061,
-0.023998243734240532,
0.0065476661548018456,
0.07816307246685028,
-0.060071635991334915,
0.05100439861416817,
0.07726791501045227,
0.023352906107902527,
0.012494008056819439,
-0.06141249090433121,
0.001000008312985301,
0.00590946851298213,
0.09741638600826263,
0.07240056246519089,
-0.06884413212537766,
-4.6784890628259745e-8,
-0.02130439318716526,
0.03908698633313179,
0.03375162556767464,
0.05827276036143303,
0.056592907756567,
-0.012450499460101128,
-0.04861636087298393,
0.03488357737660408,
-0.018394937738776207,
-0.04084017500281334,
0.056113455444574356,
-0.004357475321739912,
-0.0050714267417788506,
0.008764713071286678,
-0.017755424603819847,
0.054385047405958176,
0.04526341333985329,
0.03684273362159729,
0.026831129565835,
-0.0504019558429718,
0.016163423657417297,
0.02395564131438732,
0.02668657898902893,
-0.00587256345897913,
-0.0028655685018748045,
-0.07218290120363235,
0.03385556489229202,
0.08825966715812683,
0.0891728401184082,
-0.03457513451576233,
0.021939314901828766,
-0.019932318478822708,
0.018897246569395065,
0.08175577223300934,
0.05718596652150154,
-0.012836012057960033,
0.009398759342730045,
-0.05336344242095947,
-0.037081703543663025,
0.006024823524057865,
-0.017384033650159836,
0.03479443117976189,
-0.03112337365746498,
0.06626193970441818,
-0.05111982300877571,
0.03203584998846054,
-0.11253907531499863,
-0.05927865579724312,
0.04234633222222328,
-0.035002242773771286,
0.06988854706287384,
-0.037783119827508926,
-0.021370835602283478,
0.03963801637291908,
0.05306350439786911,
-0.010097104124724865,
0.011843863874673843,
-0.02584298700094223,
0.006240159273147583,
-0.00117787707131356,
0.09922496974468231,
-0.06870730966329575,
-0.029453856870532036,
0.05993185192346573
] |
PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling | https://openreview.net/forum?id=v2zcCDYMok | [
"Junchao Gong",
"Siwei Tu",
"Weidong Yang",
"Ben Fei",
"Kun Chen",
"zhangwenlong",
"Xiaokang Yang",
"Wanli Ouyang",
"LEI BAI"
] | Poster | Precipitation nowcasting plays a pivotal role in socioeconomic sectors, especially in severe convective weather warnings. Although notable progress has been achieved by approaches mining the spatiotemporal correlations with deep learning, these methods still suffer severe blurriness as the lead time increases, which hampers accurate predictions for extreme precipitation. To alleviate blurriness, researchers explore generative methods conditioned on blurry predictions. However, the pairs of blurry predictions and corresponding ground truth need to be given in advance, making the training pipeline cumbersome and limiting the generality of generative models within blurry modes that appear in training data. By rethinking the blurriness in precipitation nowcasting as a blur kernel acting on predictions, we propose an unsupervised postprocessing method to eliminate the blurriness without the requirement of training with the pairs of blurry predictions and corresponding ground truth. Specifically, we utilize blurry predictions to guide the generation process of a pre-trained unconditional denoising diffusion probabilistic model (DDPM) to obtain high-fidelity predictions with eliminated blurriness. A zero-shot blur kernel estimation mechanism and an auto-scale denoise guidance strategy are introduced to adapt the unconditional DDPM to any blurriness modes varying from datasets and lead times in precipitation nowcasting. Extensive experiments are conducted on 7 precipitation radar datasets, demonstrating the generality and superiority of our method. | AI for Science; Precipitation Nowcasting; Diffusion Model; Zero-shot Blurriness Kernel; Auto-scale Denoise Guidance | null | 728 | 2410.05805 | [
-0.07912438362836838,
-0.1471041887998581,
0.10733295232057571,
0.09762126952409744,
0.044035904109478,
0.004437615629285574,
-0.01647055707871914,
-0.07474032044410706,
-0.06998103857040405,
-0.04780808836221695,
-0.034871868789196014,
-0.044380784034729004,
0.01982187107205391,
0.01295210886746645,
-0.08233339339494705,
-0.029903177171945572,
0.06632781773805618,
0.020383775234222412,
-0.06639156490564346,
-0.08605745434761047,
0.07111065089702606,
0.04399135708808899,
-0.10203774273395538,
0.08703973889350891,
0.09253653138875961,
-0.003467662027105689,
0.03435607627034187,
-0.002839920111000538,
-0.018601752817630768,
0.05119172856211662,
0.03593401238322258,
0.009081737138330936,
0.054821908473968506,
-0.034418992698192596,
-0.03970694541931152,
0.022569820284843445,
-0.017621196806430817,
0.024116965010762215,
0.01530681923031807,
-0.033629074692726135,
0.04067584499716759,
-0.02954704500734806,
-0.02052990160882473,
0.027273356914520264,
0.0723225325345993,
0.04614820331335068,
0.07809658348560333,
-0.006336844060570002,
0.012020333670079708,
0.02454652637243271,
0.02553047426044941,
0.057975467294454575,
-0.003778172191232443,
0.027842123061418533,
-0.008400517515838146,
0.011315732263028622,
0.002975114621222019,
0.049016427248716354,
0.00719458656385541,
-0.003349920269101858,
-0.08227690309286118,
-0.07869105786085129,
-0.04644070193171501,
-0.019430415704846382,
0.06431320309638977,
-0.008952969685196877,
0.020333873108029366,
0.0937584638595581,
-0.02687893807888031,
0.007172878831624985,
0.00045035642688162625,
0.13503214716911316,
-0.02113623172044754,
-0.0032305705826729536,
0.015077383257448673,
0.03569546341896057,
0.0711163878440857,
-0.04389755800366402,
0.0556475967168808,
-0.0716792568564415,
0.08975815027952194,
0.01639869250357151,
0.07832900434732437,
-0.02250632271170616,
0.026099493727087975,
-0.009695122018456459,
0.032146237790584564,
0.0793236717581749,
-0.006041010841727257,
-0.003992327954620123,
-0.0672418400645256,
-0.04795365780591965,
-0.009432562626898289,
0.01370591763406992,
-0.04149230569601059,
0.06754568964242935,
0.024408206343650818,
-0.08274589478969574,
-0.014827459119260311,
0.0810900405049324,
0.013766338117420673,
-0.05699707940220833,
0.036392729729413986,
-0.07674827426671982,
0.0843641459941864,
-0.04526545852422714,
0.02345294877886772,
0.02176090143620968,
0.05926494672894478,
0.018413685262203217,
0.011730242520570755,
-0.0006467968341894448,
0.03514251485466957,
-0.031555820256471634,
0.05571522191166878,
0.04323283210396767,
-0.0031849343795329332,
-0.020628249272704124,
-0.020351599901914597,
-0.044438425451517105,
-0.04050032049417496,
0.018611455336213112,
-0.013717238791286945,
0.07542053610086441,
0.057015832513570786,
-0.041868504136800766,
-0.08963985741138458,
4.524462695293615e-33,
0.07010942697525024,
-0.02388857863843441,
0.061749715358018875,
0.04091986268758774,
0.009700641967356205,
-0.040546223521232605,
-0.053970400243997574,
0.020578186959028244,
0.015154308639466763,
-0.04984812065958977,
-0.028919659554958344,
-0.04505106434226036,
-0.08344747126102448,
0.07526623457670212,
0.020364927127957344,
-0.03973438963294029,
-0.046559497714042664,
0.07342474907636642,
-0.008943089284002781,
0.04014507308602333,
-0.05021107569336891,
0.014497466385364532,
-0.07295074313879013,
-0.00987174455076456,
0.018116047605872154,
0.005900390911847353,
0.12310720980167389,
-0.018127314746379852,
0.018370041623711586,
0.007328129839152098,
-0.08923256397247314,
0.013374553062021732,
0.03532278910279274,
-0.0071837459690868855,
0.027777692303061485,
-0.08195855468511581,
-0.046310875564813614,
-0.0026515780482441187,
0.010618765838444233,
-0.026400305330753326,
-0.027876710519194603,
0.018484029918909073,
-0.0465029701590538,
-0.03431866690516472,
0.0014718020102009177,
-0.07759195566177368,
0.021817775443196297,
0.031046824529767036,
-0.09132258594036102,
-0.025053270161151886,
0.056187957525253296,
0.024850983172655106,
-0.006801173090934753,
-0.08668442070484161,
-0.0592559315264225,
-0.008747238665819168,
0.09448706358671188,
-0.008963828906416893,
0.09960464388132095,
-0.09371834993362427,
0.05179929733276367,
-0.0021623570937663317,
-0.021169258281588554,
-0.03510977700352669,
-0.020142031833529472,
0.03435460478067398,
-0.018953274935483932,
0.11471041291952133,
0.003958968445658684,
-0.04882359504699707,
0.0000447942984465044,
0.052076809108257294,
-0.06418147683143616,
-0.05511470511555672,
0.06866179406642914,
0.026899702847003937,
0.05025578290224075,
-0.036493849009275436,
0.0030933350790292025,
0.016984933987259865,
-0.035047177225351334,
-0.007161818910390139,
-0.06882859766483307,
-0.06525039672851562,
-0.06106114760041237,
0.012999767437577248,
-0.025461044162511826,
0.010901819914579391,
-0.06205170229077339,
-0.025703132152557373,
-0.0694175735116005,
-0.03958912938833237,
0.010913574136793613,
-0.0033298747148364782,
-0.01468500867486,
-4.342192377965034e-33,
-0.044176504015922546,
0.08223304897546768,
-0.0863240584731102,
0.046537406742572784,
-0.0351090282201767,
-0.02372371405363083,
0.0037067111115902662,
-0.0046074241399765015,
0.07404845207929611,
-0.06399974226951599,
0.015941284596920013,
-0.014324384741485119,
0.03646228462457657,
-0.0060229673981666565,
-0.027528854086995125,
-0.019054176285862923,
-0.03358226642012596,
0.026237383484840393,
-0.10491577535867691,
0.020351910963654518,
-0.0208330936729908,
0.008938242681324482,
-0.044812120497226715,
0.06321349740028381,
-0.04008030891418457,
0.01651911996304989,
0.006134513299912214,
0.16765546798706055,
-0.01528017595410347,
0.020229225978255272,
-0.03228806331753731,
-0.011388377286493778,
-0.04923005774617195,
-0.0568508543074131,
-0.12117442488670349,
0.11500002443790436,
0.06931920349597931,
-0.04936295747756958,
-0.09649934619665146,
-0.008888845331966877,
0.0824754536151886,
-0.0760975256562233,
-0.04302053898572922,
0.0200316421687603,
-0.03845702111721039,
-0.042490627616643906,
0.036047376692295074,
-0.00915070716291666,
0.0741463303565979,
0.053469061851501465,
0.00502982921898365,
0.061172619462013245,
-0.12589147686958313,
0.08054746687412262,
-0.019435787573456764,
-0.1097082868218422,
-0.0684957504272461,
-0.06867202371358871,
0.00949820689857006,
-0.013461645692586899,
-0.09558068960905075,
-0.044962625950574875,
-0.00941103883087635,
-0.036010682582855225,
-0.04519287869334221,
-0.10449808090925217,
-0.03730704262852669,
0.02593230828642845,
0.07411923259496689,
0.042620979249477386,
0.03792193531990051,
0.005091662053018808,
0.031547751277685165,
0.00309946876950562,
0.03884386271238327,
-0.021963292732834816,
0.06674584746360779,
0.05762602388858795,
0.0025662248954176903,
0.010902989655733109,
0.030697112902998924,
-0.02704755589365959,
0.015312956646084785,
0.04986750707030296,
0.06563300639390945,
0.010250922292470932,
0.05989142134785652,
-0.03229638934135437,
0.026503801345825195,
0.055247869342565536,
-0.08228807151317596,
0.03849996253848076,
-0.037848006933927536,
0.014250429347157478,
-0.09213065356016159,
-4.117942609127567e-8,
-0.017902234569191933,
0.06725360453128815,
0.021747304126620293,
0.009228739887475967,
0.056915853172540665,
-0.04289762303233147,
0.023877624422311783,
0.06532352417707443,
0.008973849937319756,
-0.09077528119087219,
0.05253348872065544,
-0.02013249509036541,
-0.040549006313085556,
-0.0654119923710823,
0.06616009026765823,
-0.01394951343536377,
0.041198763996362686,
0.008643959648907185,
-0.05575079098343849,
-0.006648617330938578,
0.05113194137811661,
-0.051884591579437256,
-0.07168416678905487,
0.04178555682301521,
0.115594282746315,
-0.042797792702913284,
-0.031154952943325043,
0.02147015370428562,
0.06913335621356964,
0.04550643637776375,
-0.013142399489879608,
0.0025558460038155317,
-0.01673969440162182,
-0.05089861527085304,
0.0011103887809440494,
0.0624387226998806,
0.07620822638273239,
0.10302039980888367,
-0.006187296472489834,
-0.0007862173952162266,
-0.02778521738946438,
0.043925993144512177,
-0.029005076736211777,
0.012635289691388607,
-0.009917830117046833,
0.0190426092594862,
0.04854055866599083,
-0.019680513069033623,
0.05701455473899841,
0.011828583665192127,
-0.04085136204957962,
0.011414428241550922,
0.02305341325700283,
0.11650342494249344,
0.07538726925849915,
-0.034493882209062576,
0.02001281827688217,
-0.03525124862790108,
0.01315028965473175,
0.059364672750234604,
-0.016737019643187523,
-0.013436544686555862,
-0.08308446407318115,
0.045929498970508575
] |
ChartMimic: Evaluating LMM's Cross-Modal Reasoning Capability via Chart-to-Code Generation | https://openreview.net/forum?id=sGpCzsfd1K | [
"Cheng Yang",
"Chufan Shi",
"Yaxin Liu",
"Bo Shui",
"Junjie Wang",
"Mohan Jing",
"Linran XU",
"Xinyu Zhu",
"Siheng Li",
"Yuxiang Zhang",
"Gongye Liu",
"Xiaomei Nie",
"Deng Cai",
"Yujiu Yang"
] | Poster | We introduce a new benchmark, ChartMimic, aimed at assessing the visually-grounded code generation capabilities of large multimodal models (LMMs). ChartMimic utilizes information-intensive visual charts and textual instructions as inputs, requiring LMMs to generate the corresponding code for chart rendering.
ChartMimic includes $4,800$ human-curated (figure, instruction, code) triplets, which represent the authentic chart use cases found in scientific papers across various domains (e.g., Physics, Computer Science, Economics, etc). These charts span $18$ regular types and $4$ advanced types, diversifying into $201$ subcategories.
Furthermore, we propose multi-level evaluation metrics to provide an automatic and thorough assessment of the output code and the rendered charts.
Unlike existing code generation benchmarks, ChartMimic places emphasis on evaluating LMMs' capacity to harmonize a blend of cognitive capabilities, encompassing visual understanding, code generation, and cross-modal reasoning. The evaluation of $3$ proprietary models and $14$ open-weight models highlights the substantial challenges posed by ChartMimic. Even the advanced GPT-4o, InternVL2-Llama3-76B only achieved an average score across Direct Mimic and Customized Mimic tasks of $82.2$ and $61.6$, respectively, indicating significant room for improvement.
We anticipate that ChartMimic will inspire the development of LMMs, advancing the pursuit of artificial general intelligence. | Dataset and Benchmark, Code generation, Chart Understand and Reasoning | A benchmark for evaluating LMM’s cross-modal reasoning capability via chart-to-code generation | 723 | 2406.09961 | [
-0.028552129864692688,
-0.09680141508579254,
-0.02588244527578354,
0.005030239932239056,
0.0896327793598175,
-0.061264291405677795,
-0.09208551049232483,
0.08333847671747208,
-0.004626656416803598,
0.015064973384141922,
-0.03422097861766815,
-0.11722781509160995,
0.017849750816822052,
0.018429776653647423,
0.04069118574261665,
0.02578173577785492,
-0.0141165005043149,
-0.013026704080402851,
-0.0505191795527935,
-0.07340923696756363,
0.03581630066037178,
-0.03208767622709274,
-0.006782558746635914,
-0.04326404631137848,
0.077833391726017,
0.04763804376125336,
-0.03255457803606987,
-0.014214732684195042,
0.062380362302064896,
-0.016294829547405243,
0.005036329384893179,
0.0328197106719017,
0.008033069781959057,
0.06511833518743515,
-0.043090205639600754,
0.05851425603032112,
0.02205623872578144,
-0.01994732953608036,
0.006333822384476662,
-0.0019807994831353426,
-0.094356968998909,
0.03268421068787575,
0.05168962478637695,
0.009442005306482315,
0.03136363625526428,
-0.06896316260099411,
-0.09710367769002914,
-0.09091648459434509,
-0.03620157018303871,
0.06132176145911217,
-0.12480519711971283,
-0.03897668421268463,
0.0005765150999650359,
-0.01937316171824932,
0.019296247512102127,
0.035089023411273956,
0.05959584191441536,
-0.08693274855613708,
0.02496248297393322,
-0.030498573556542397,
-0.09773418307304382,
-0.014170963317155838,
-0.033057499676942825,
-0.002303588902577758,
0.06456057727336884,
0.0209465604275465,
-0.0025416919961571693,
-0.0036981564480811357,
0.010194730944931507,
0.020156130194664,
-0.0914800614118576,
-0.0012887999182567,
-0.045742467045784,
0.014257622882723808,
-0.015631509944796562,
0.038983121514320374,
0.00979539379477501,
-0.055673182010650635,
0.00034106997190974653,
-0.15092025697231293,
0.035999976098537445,
0.028918854892253876,
-0.04026837646961212,
0.06657050549983978,
0.04209623485803604,
0.014705234207212925,
0.022638652473688126,
0.0676039382815361,
0.03418392688035965,
0.05216627195477486,
-0.020114146173000336,
-0.03290034085512161,
0.049089331179857254,
0.021150674670934677,
0.04114867374300957,
0.09076289087533951,
0.04281033203005791,
-0.10889790207147598,
-0.001184814260341227,
0.05964365601539612,
0.046000365167856216,
0.014654545113444328,
0.029288578778505325,
-0.05052107572555542,
0.050915464758872986,
-0.04996918514370918,
0.1210566982626915,
0.03611278533935547,
-0.008594703860580921,
-0.05984785780310631,
0.036481261253356934,
0.06470763683319092,
-0.016065960749983788,
-0.024257035925984383,
-0.01723557524383068,
-0.002752389060333371,
0.01903597079217434,
-0.0035071515012532473,
0.09992080181837082,
0.05743095651268959,
0.07439781725406647,
0.030270026996731758,
0.009529105387628078,
0.002963343635201454,
0.04323085770010948,
-0.09012274444103241,
-0.062369246035814285,
7.403440663986352e-34,
0.048058003187179565,
0.05008454620838165,
0.0752633661031723,
0.07657133042812347,
0.072724349796772,
-0.023688606917858124,
-0.08300492912530899,
-0.007014885079115629,
-0.06940198689699173,
0.03158783167600632,
-0.017082534730434418,
0.11589246243238449,
-0.063348688185215,
0.1695253551006317,
0.06570718437433243,
0.030428921803832054,
-0.007705433759838343,
0.023638039827346802,
-0.09294937551021576,
-0.028581934049725533,
-0.03574614226818085,
-0.05060754716396332,
0.018508901819586754,
0.02078736573457718,
0.03341979905962944,
0.10861009359359741,
0.035181667655706406,
-0.023908087983727455,
-0.04839397221803665,
0.017513107508420944,
-0.09524764865636826,
-0.02088930457830429,
-0.024914544075727463,
0.005852098111063242,
0.01696883700788021,
-0.03820308670401573,
-0.02684410661458969,
-0.09601590037345886,
0.025594562292099,
0.01041072141379118,
-0.043884046375751495,
-0.011878186836838722,
-0.0325516052544117,
0.022104328498244286,
-0.040047790855169296,
-0.03542504832148552,
-0.008395851589739323,
0.0348944254219532,
-0.028562031686306,
-0.0010177859803661704,
0.003237382508814335,
0.060390766710042953,
0.025617867708206177,
-0.047522250562906265,
-0.0038985649589449167,
-0.07174549251794815,
0.03161626681685448,
-0.0007373785483650863,
0.0794990286231041,
0.0758078545331955,
-0.010740339756011963,
0.0890575647354126,
0.03153037279844284,
0.011549941264092922,
-0.05536573752760887,
0.0560145303606987,
-0.05185861885547638,
0.03673042729496956,
0.07270990312099457,
0.053077224642038345,
-0.10213629901409149,
0.02110351249575615,
-0.009393547661602497,
-0.026804279536008835,
0.026787064969539642,
0.016835154965519905,
0.012497148476541042,
-0.07950018346309662,
0.04048008844256401,
-0.047200169414281845,
-0.003907039295881987,
0.009463940747082233,
0.009298932738602161,
-0.025179870426654816,
0.013783073984086514,
-0.11074833571910858,
-0.015892814844846725,
-0.039972517639398575,
-0.038145240396261215,
-0.07564341276884079,
-0.049993738532066345,
-0.0711495578289032,
-0.01993371918797493,
-0.05559529736638069,
-0.021430959925055504,
-1.740688014073975e-33,
-0.012832125648856163,
0.04161003977060318,
-0.0365854948759079,
0.051084164530038834,
-0.027839798480272293,
-0.0824575424194336,
0.01885729841887951,
0.012945803813636303,
-0.01727379485964775,
0.023638581857085228,
-0.0007340380689129233,
-0.047395866364240646,
-0.043238863348960876,
-0.01360801886767149,
0.08947501331567764,
-0.0059389108791947365,
-0.09163730591535568,
0.0026343839708715677,
-0.02501840703189373,
-0.02069581113755703,
0.04480621963739395,
0.07721154391765594,
-0.036021240055561066,
0.030018312856554985,
-0.02741946652531624,
0.09430526942014694,
-0.06808380037546158,
-0.06695802509784698,
-0.01517342310398817,
-0.03257664665579796,
-0.021900588646531105,
-0.03169672563672066,
0.021132130175828934,
-0.04931078851222992,
0.034510642290115356,
0.030301004648208618,
0.16427984833717346,
-0.024084078148007393,
-0.04256149008870125,
0.09043075144290924,
0.04618271067738533,
-0.026900017634034157,
0.016675006598234177,
-0.0023819594644010067,
-0.0048047760501503944,
0.00426581921055913,
-0.058698542416095734,
-0.04519336670637131,
0.020668700337409973,
0.01904585026204586,
-0.01594841107726097,
-0.00666580256074667,
0.045931700617074966,
-0.040984854102134705,
-0.014161177910864353,
-0.11928451806306839,
-0.05596226826310158,
0.01484774798154831,
0.037955593317747116,
0.0032337962184101343,
-0.060416094958782196,
-0.045730382204055786,
0.012458846904337406,
-0.03225797042250633,
-0.007253498304635286,
-0.07479555904865265,
-0.025589948520064354,
-0.06702364981174469,
-0.07170899957418442,
0.0023543124552816153,
0.0049248915165662766,
-0.05915147811174393,
-0.023669743910431862,
-0.03618953004479408,
0.023959681391716003,
0.034152690321207047,
0.0422012098133564,
0.005676544737070799,
-0.0009076701244339347,
-0.010410687886178493,
0.08212585002183914,
-0.05938337370753288,
0.10754231363534927,
0.07144521921873093,
0.02966940775513649,
0.036608919501304626,
-0.003511484945192933,
0.0747576430439949,
-0.00479111960157752,
0.03765205666422844,
-0.062074318528175354,
0.054062049835920334,
0.005648287478834391,
0.11135294288396835,
0.012393241748213768,
-3.993066144403201e-8,
-0.039334412664175034,
0.04784531891345978,
0.004727551247924566,
-0.01079901959747076,
0.011529509909451008,
0.027005746960639954,
-0.0646124854683876,
0.021516378968954086,
0.011578400619328022,
0.038066573441028595,
0.10604041814804077,
-0.04723340645432472,
-0.014507490210235119,
-0.036106448620557785,
0.04654688760638237,
0.07673925161361694,
-0.04033864662051201,
-0.029756689444184303,
0.00012908366625197232,
0.017769644036889076,
0.047372229397296906,
0.048526618629693985,
0.007055233232676983,
0.011145644821226597,
-0.060333266854286194,
-0.08113229274749756,
-0.034358132630586624,
0.14867152273654938,
0.01831541769206524,
0.024010134860873222,
0.019301163032650948,
0.04282892867922783,
0.02118978276848793,
-0.002755607943981886,
-0.002442172495648265,
-0.035849738866090775,
-0.030266769230365753,
-0.007602959871292114,
0.03519594296813011,
0.00798512902110815,
0.006607338320463896,
0.07849400490522385,
-0.10927175730466843,
0.03730788826942444,
0.1056501567363739,
-0.0018812958151102066,
-0.04376142844557762,
-0.05867549404501915,
-0.026889566332101822,
-0.02634294144809246,
-0.10067678987979889,
-0.084668830037117,
-0.033093325793743134,
0.0883215144276619,
0.031992122530937195,
0.018168332055211067,
-0.047754742205142975,
0.07151327282190323,
0.048988498747348785,
0.06998779624700546,
0.029353003948926926,
-0.02048349380493164,
-0.013079538941383362,
-0.05362121760845184
] |
Does Training with Synthetic Data Truly Protect Privacy? | https://openreview.net/forum?id=C8niXBHjfO | [
"Yunpeng Zhao",
"Jie Zhang"
] | Poster | As synthetic data becomes increasingly popular in machine learning tasks, numerous methods---without formal differential privacy guarantees---use synthetic data for training. These methods often claim, either explicitly or implicitly, to protect the privacy of the original training data.
In this work, we explore four different training paradigms: coreset selection, dataset distillation, data-free knowledge distillation, and synthetic data generated from diffusion models. While all these methods utilize synthetic data for training, they lead to vastly different conclusions regarding privacy preservation. We caution that empirical approaches to preserving data privacy require careful and rigorous evaluation; otherwise, they risk providing a false sense of privacy. | ML privacy, membership inference | We rigorously evaluate privacy leakage across various methods based on training with synthetic data, and none of these methods achieve a better trade-off than the differential privacy baselines. | 722 | 2502.12976 | [
-0.13124385476112366,
-0.041350286453962326,
0.003362265881150961,
0.04131932556629181,
0.07031314820051193,
0.018193678930401802,
-0.01912771537899971,
-0.06603754311800003,
0.011848535388708115,
-0.02044307067990303,
0.0299236960709095,
0.010068562813103199,
0.002282383618876338,
-0.06610266119241714,
-0.027451222762465477,
0.002369857858866453,
0.04188646376132965,
0.07839664816856384,
-0.08296389877796173,
-0.03459306061267853,
-0.049863528460264206,
-0.035499438643455505,
0.021787159144878387,
0.07749509066343307,
0.08766419440507889,
-0.024701012298464775,
0.13095732033252716,
-0.03881445899605751,
-0.006057420279830694,
-0.04054964333772659,
0.08750008791685104,
-0.017242740839719772,
-0.027454175055027008,
0.02590528503060341,
-0.02240150049328804,
0.01430005207657814,
-0.029635503888130188,
0.0004301322333049029,
0.008333062753081322,
-0.013003760948777199,
-0.018698930740356445,
-0.0765044242143631,
-0.07850025594234467,
0.027799086645245552,
0.024184931069612503,
0.07167787104845047,
0.05853268504142761,
-0.021896852180361748,
-0.007512582466006279,
-0.012847119010984898,
-0.07881022989749908,
-0.0379975289106369,
-0.05890851095318794,
0.059663329273462296,
-0.027598273009061813,
-0.09099222719669342,
0.05039740353822708,
0.011057278141379356,
-0.03679533302783966,
0.04325259104371071,
-0.04103168845176697,
-0.07177919149398804,
-0.04856786131858826,
0.025797225534915924,
-0.019503215327858925,
0.01500707771629095,
-0.0008349541458301246,
0.10410383343696594,
0.03490469232201576,
-0.04578715190291405,
-0.011759120039641857,
0.11215471476316452,
-0.04157782718539238,
-0.013195320032536983,
0.07735449820756912,
0.029258081689476967,
-0.015355785377323627,
-0.01476042065769434,
0.02279445342719555,
-0.08986571431159973,
0.03216417506337166,
-0.010052359662950039,
0.06893385201692581,
-0.013568594120442867,
0.016429634764790535,
0.019036054611206055,
0.021577982231974602,
0.009187523275613785,
0.03757719695568085,
0.005613876041024923,
-0.05075879395008087,
0.040665604174137115,
0.009184911847114563,
-0.0013495530001819134,
-0.0108680110424757,
0.012269618920981884,
-0.03403617814183235,
0.0026130154728889465,
0.03577122837305069,
0.02140118181705475,
-0.02777002938091755,
0.06495741754770279,
-0.04997178167104721,
0.007331929635256529,
0.028867514804005623,
0.007305252365767956,
0.0620291642844677,
-0.018953757360577583,
0.06518687307834625,
0.013874763622879982,
0.026123683899641037,
0.05992170050740242,
-0.03213925287127495,
-0.003901137737557292,
-0.009539222344756126,
-0.04427780956029892,
-0.05013991892337799,
0.018899064511060715,
-0.010188506916165352,
0.04375727102160454,
0.004603947978466749,
0.01335538737475872,
0.04556755721569061,
0.014658155851066113,
-0.04985526576638222,
-0.06381651759147644,
-0.0965651348233223,
1.5232460792493258e-33,
0.008580229245126247,
-0.0003826917090918869,
0.008935441263020039,
-0.03913208097219467,
-0.040897123515605927,
-0.0027658664621412754,
-0.04842140153050423,
-0.024276740849018097,
-0.041240718215703964,
0.04493359848856926,
-0.018497515469789505,
0.051358044147491455,
-0.041779085993766785,
0.07413647323846817,
0.08251766115427017,
0.07052890211343765,
-0.010875437408685684,
0.08112350851297379,
0.06084351986646652,
-0.024956315755844116,
0.16150331497192383,
-0.007758719380944967,
-0.00898438785225153,
-0.02438393235206604,
0.010328958742320538,
-0.012413005344569683,
0.04867097735404968,
0.003157402388751507,
0.06607026606798172,
0.025340694934129715,
-0.11216074228286743,
0.042979128658771515,
0.051887284964323044,
0.04081271216273308,
0.0949934870004654,
0.07028495520353317,
-0.019600510597229004,
-0.037677742540836334,
0.053242314606904984,
0.05126895383000374,
0.07921512424945831,
0.01531832292675972,
0.06630825251340866,
0.04269177466630936,
-0.034424539655447006,
0.02408941462635994,
-0.060379184782505035,
0.017387084662914276,
-0.0024598033633083105,
-0.04320395737886429,
-0.03201789781451225,
0.034646157175302505,
-0.06603094190359116,
-0.09807952493429184,
-0.06175203248858452,
0.018015732988715172,
0.027479171752929688,
-0.034381646662950516,
0.008348137140274048,
-0.016387106850743294,
-0.03342853859066963,
0.08286353945732117,
0.01198219507932663,
0.04000134393572807,
-0.0723571628332138,
-0.006316885352134705,
-0.05171927809715271,
-0.05913931503891945,
0.06857026368379593,
-0.0308852419257164,
-0.0248220507055521,
0.028468970209360123,
-0.08423197269439697,
-0.0035866580437868834,
0.022290654480457306,
-0.10370449721813202,
0.03694993257522583,
0.03819521889090538,
0.020684512332081795,
-0.052071671932935715,
0.07890507578849792,
-0.01501466054469347,
-0.06992699950933456,
-0.012019453570246696,
-0.08916954696178436,
-0.007821695879101753,
-0.04334276542067528,
0.03425372391939163,
0.004319456871598959,
0.04653215408325195,
-0.07509075105190277,
-0.055873069912195206,
-0.07241281867027283,
-0.014599828980863094,
-0.058714523911476135,
-2.5490159697329516e-33,
-0.10818377882242203,
0.05891815572977066,
-0.06491050869226456,
0.09226325154304504,
-0.022673258557915688,
0.02420479618012905,
-0.015556718222796917,
0.025637825950980186,
0.003836359828710556,
0.003402163041755557,
-0.012225523591041565,
-0.005113479681313038,
0.040767669677734375,
-0.07292967289686203,
-0.030204594135284424,
0.039754245430231094,
-0.020607521757483482,
-0.010026083327829838,
-0.11607523262500763,
-0.011926176957786083,
-0.045677244663238525,
0.05300856754183769,
-0.06982819736003876,
0.07680457830429077,
0.06006745249032974,
-0.016679715365171432,
-0.0983113944530487,
0.12312386929988861,
0.10362913459539413,
0.002722884062677622,
-0.06605019420385361,
-0.03859768062829971,
-0.056534431874752045,
-0.13583822548389435,
-0.11413152515888214,
-0.013092965818941593,
0.0480879582464695,
0.05934300273656845,
-0.006633559241890907,
0.06262867897748947,
-0.01256053987890482,
0.017867106944322586,
-0.14381003379821777,
0.0388682521879673,
0.0014955569058656693,
-0.038780685514211655,
-0.04782538861036301,
-0.03625347465276718,
0.05736083909869194,
0.013476729393005371,
0.07322410494089127,
0.04532536864280701,
0.0292907003313303,
-0.027152221649885178,
0.018988821655511856,
-0.06428521126508713,
0.019264433532953262,
-0.009754505939781666,
-0.0393000990152359,
0.023569883778691292,
-0.020413357764482498,
0.013069278560578823,
0.009201730601489544,
0.024909693747758865,
-0.011118486523628235,
-0.06594064086675644,
-0.0683741644024849,
0.04915723577141762,
-0.027773385867476463,
-0.004043491557240486,
0.06365358829498291,
-0.042156390845775604,
-0.013519104570150375,
0.010041158646345139,
-0.01628740318119526,
-0.020289985463023186,
0.011604592204093933,
-0.008914663456380367,
-0.04760320112109184,
0.1306271106004715,
0.08410154283046722,
-0.06134895980358124,
0.09243721514940262,
-0.041011255234479904,
0.11675646156072617,
0.024404579773545265,
0.09784597158432007,
-0.0810181051492691,
-0.05687807872891426,
-0.005835725460201502,
-0.009329280816018581,
-0.026424000039696693,
-0.12411007285118103,
0.04899746552109718,
-0.07695850729942322,
-3.52974467432432e-8,
0.02403932996094227,
-0.012359172105789185,
0.08377022296190262,
0.05284160375595093,
0.005685348995029926,
0.00014073667989578098,
-0.01834067888557911,
0.06970477849245071,
-0.11494843661785126,
-0.007039867341518402,
0.03599056974053383,
0.027262305840849876,
-0.02641000971198082,
-0.023027995601296425,
0.03502247482538223,
0.09750965237617493,
0.017294811084866524,
-0.04020804539322853,
-0.00844720657914877,
0.035147201269865036,
-0.03000415302813053,
0.03477514907717705,
-0.06947532296180725,
-0.02089032530784607,
0.11137920618057251,
0.029818404465913773,
0.007205714471638203,
0.007060414645820856,
0.05999328941106796,
0.15994666516780853,
-0.023365763947367668,
-0.06051837280392647,
0.009315866976976395,
-0.01875857636332512,
-0.023085014894604683,
0.10375187546014786,
0.009585049003362656,
0.007835425436496735,
-0.00763320317491889,
-0.03149912506341934,
-0.009314554743468761,
0.036562178283929825,
-0.025753295049071312,
0.022619538009166718,
-0.03250817209482193,
-0.023278718814253807,
0.04520365968346596,
-0.08871419727802277,
0.01211122702807188,
0.03281296417117119,
0.04275813326239586,
-0.02491834945976734,
0.024974210187792778,
0.04902509227395058,
0.0629814937710762,
0.0055464389733970165,
-0.00691982451826334,
-0.010119793005287647,
0.03213953226804733,
0.03763757646083832,
0.03299856558442116,
-0.03489384055137634,
0.009186891838908195,
0.023332716897130013
] |
PT-T2I/V: An Efficient Proxy-Tokenized Diffusion Transformer for Text-to-Image/Video-Task | https://openreview.net/forum?id=lTrrnNdkOX | [
"Jing Wang",
"Ao Ma",
"Jiasong Feng",
"Dawei Leng",
"Yuhui Yin",
"Xiaodan Liang"
] | Poster | The global self-attention mechanism in diffusion transformers involves redundant computation due to the sparse and redundant nature of visual information, and the attention map of tokens within a spatial window shows significant similarity. To address this redundancy, we propose the Proxy-Tokenized Diffusion Transformer (PT-DiT), which employs sparse representative token attention (where the number of representative tokens is much smaller than the total number of tokens) to efficiently model global visual information. Specifically, within each transformer block, we compute an averaging token from each spatial-temporal window to serve as a proxy token for that region. The global semantics are captured through the self-attention of these proxy tokens and then injected into all latent tokens via cross-attention. Simultaneously, we introduce window and shift window attention to address the limitations in detail modeling caused by the sparse attention mechanism. Building on the well-designed PT-DiT, we further develop the PT-T2I/V family, which includes a variety of models for T2I, T2V, and T2MV tasks. Experimental results show that PT-DiT achieves competitive performance while reducing computational complexity in image and video generation tasks (e.g., a reduction 59\% compared to DiT and a reduction 34\% compared to PixArt-$\alpha$). The visual exhibition of and code are available at https://360cvgroup.github.io/Qihoo-T2X/. | Diffusion Transformer, Image Generation, Video Generation | null | 715 | null | [
0.005406452342867851,
-0.054181940853595734,
0.06606274098157883,
-0.004529602825641632,
0.05876108258962631,
-0.053019046783447266,
-0.043597523123025894,
0.060892511159181595,
0.09339849650859833,
-0.05640028044581413,
-0.032942965626716614,
-0.025137821212410927,
-0.01609056256711483,
0.08378975093364716,
0.013513973914086819,
-0.03062107414007187,
0.08892187476158142,
0.050972532480955124,
-0.09817201644182205,
-0.108352892100811,
0.022478818893432617,
-0.10541962087154388,
0.040185026824474335,
-0.02873966470360756,
0.004827149678021669,
0.10555315762758255,
-0.011350066401064396,
-0.07563074678182602,
0.014497529715299606,
-0.013492853380739689,
-0.012039512395858765,
-0.0136629743501544,
-0.07918178290128708,
0.055135149508714676,
-0.10423839092254639,
0.03925439715385437,
-0.07556729018688202,
0.024635110050439835,
-0.0000021422779354907107,
-0.10489505529403687,
0.07884129881858826,
0.026924515143036842,
-0.050328779965639114,
0.0394786112010479,
0.08220690488815308,
0.0313851535320282,
0.04881269857287407,
-0.02207387052476406,
0.02540777251124382,
-0.035173896700143814,
-0.045429181307554245,
0.004628396593034267,
-0.0393255278468132,
0.023503368720412254,
-0.04058937355875969,
0.04904747009277344,
0.06272606551647186,
0.001965575385838747,
0.029851652681827545,
-0.009152273647487164,
-0.03622400760650635,
-0.03448307886719704,
0.07017597556114197,
-0.0016598172951489687,
-0.010823599994182587,
-0.03417520597577095,
0.0436498261988163,
0.013540389947593212,
-0.018057629466056824,
-0.01501513086259365,
-0.03726647049188614,
0.08233008533716202,
0.015431532636284828,
-0.005870533175766468,
0.016787763684988022,
-0.030949506908655167,
-0.015094224363565445,
-0.032589543610811234,
0.03085191547870636,
-0.07516034692525864,
0.13377614319324493,
0.014842874370515347,
0.07068703323602676,
0.002764717675745487,
0.1494520604610443,
0.003414943115785718,
0.00022980499488767236,
0.023931745439767838,
-0.02296443283557892,
0.012508029118180275,
-0.07627708464860916,
-0.07836349308490753,
0.0395200252532959,
-0.038209155201911926,
0.0246951375156641,
-0.05744326114654541,
-0.010593890212476254,
-0.046544160693883896,
-0.017725428566336632,
0.0784357488155365,
-0.007428516633808613,
-0.03581630811095238,
0.03567468747496605,
0.003934470005333424,
0.04443772882223129,
-0.019038446247577667,
0.05372269079089165,
0.027766071259975433,
0.016385741531848907,
-0.026183104142546654,
0.03693978115916252,
0.02590196579694748,
0.007979127578437328,
-0.01758231408894062,
0.01118351612240076,
-0.04398263618350029,
0.004648789297789335,
0.006884376052767038,
0.058798979967832565,
0.046638499945402145,
0.03867122158408165,
-0.035591427236795425,
-0.08912673592567444,
-0.0006459170836023986,
-0.03783426061272621,
-0.04478279501199722,
0.03579272702336311,
-4.3296481772439553e-35,
-0.0501197911798954,
0.012896692380309105,
0.021337421610951424,
0.06083369255065918,
0.03526638448238373,
-0.005434446502476931,
0.010464934632182121,
-0.034017663449048996,
-0.04009212180972099,
-0.07269603759050369,
0.01663936860859394,
-0.055456094443798065,
-0.03147139027714729,
0.1557609587907791,
-0.028417536988854408,
-0.04912281408905983,
-0.010149354115128517,
0.0763295441865921,
0.014698957093060017,
-0.024306967854499817,
-0.006905882153660059,
0.03915879502892494,
-0.02016466110944748,
-0.046002548187971115,
0.00729278102517128,
-0.043065834790468216,
0.07578323036432266,
-0.1383543312549591,
0.020766356959939003,
0.0008303752983920276,
-0.13476991653442383,
0.02203686721622944,
0.025247815996408463,
-0.004667856264859438,
0.04389061778783798,
-0.028025513514876366,
-0.007798096630722284,
-0.08431610465049744,
0.04942105710506439,
-0.03240715712308884,
-0.004085903521627188,
0.08786340802907944,
-0.0401187501847744,
-0.07416800409555435,
-0.11288898438215256,
-0.034492127597332,
0.03757113218307495,
0.07700052857398987,
-0.03485538065433502,
-0.0011814081808552146,
0.07761193066835403,
0.001182279083877802,
-0.07266043871641159,
-0.08008729666471481,
0.035842057317495346,
-0.05006832629442215,
0.07054372876882553,
-0.008959716185927391,
0.057689473032951355,
0.07161320745944977,
0.04796707257628441,
0.045957401394844055,
-0.03390543535351753,
0.11988428980112076,
0.03523503243923187,
0.00001553996480652131,
-0.008359655737876892,
0.09041213244199753,
0.05080191791057587,
0.07006369531154633,
-0.042129527777433395,
0.07726404070854187,
-0.002070247894152999,
-0.06646214425563812,
0.08074529469013214,
0.0392255000770092,
0.12455006688833237,
-0.04769562557339668,
-0.026066329330205917,
0.04738115519285202,
-0.02141048014163971,
-0.0778658539056778,
0.030348794534802437,
-0.06264171749353409,
-0.02469559758901596,
0.0036977017298340797,
0.0023583879228681326,
-0.09492145478725433,
-0.06202326714992523,
-0.09151484072208405,
0.042486440390348434,
-0.012194258160889149,
-0.026558928191661835,
-0.00461895065382123,
-0.019079316407442093,
-3.3655080550826257e-34,
0.014844775199890137,
0.07544829696416855,
0.04063060134649277,
0.0081197340041399,
-0.05458786338567734,
-0.03631880134344101,
-0.027734242379665375,
-0.017028145492076874,
0.043059904128313065,
0.01780087873339653,
-0.08156275749206543,
-0.03231159970164299,
-0.0047362735494971275,
-0.0073834676295518875,
-0.006230001803487539,
0.04166921600699425,
0.049113038927316666,
-0.00886586494743824,
-0.09618359059095383,
0.004635539371520281,
0.017641587182879448,
0.06172485277056694,
-0.029463164508342743,
0.033886488527059555,
-0.07979712635278702,
0.06331788748502731,
-0.02135750837624073,
0.07524555176496506,
0.04523728787899017,
-0.06428461521863937,
-0.020191777497529984,
-0.005448726937174797,
0.023414460942149162,
-0.05758306756615639,
-0.06107722595334053,
0.03822967782616615,
0.00622214563190937,
0.005740722641348839,
-0.07901501655578613,
0.04433755576610565,
0.05480526015162468,
-0.07324925810098648,
-0.0019175084307789803,
0.01634172908961773,
-0.07634168118238449,
-0.015808546915650368,
-0.07772320508956909,
0.01829794980585575,
-0.06051284074783325,
0.0750921219587326,
-0.02889111079275608,
0.04450321942567825,
-0.10539428889751434,
-0.004008299671113491,
-0.04678637906908989,
-0.0561341755092144,
-0.06132722273468971,
0.03316272050142288,
0.038640595972537994,
-0.06261012703180313,
0.06302521377801895,
-0.0397411473095417,
-0.007167347241193056,
-0.05664422735571861,
-0.012428594753146172,
-0.04900259152054787,
-0.02869696170091629,
-0.060119710862636566,
0.011725112795829773,
0.03679588809609413,
0.07513073086738586,
0.002162742894142866,
-0.0054335989989340305,
-0.04218854382634163,
0.05519162118434906,
0.03584743291139603,
0.07266712188720703,
-0.03589542210102081,
0.03447958081960678,
-0.0029948288574814796,
-0.032796118408441544,
-0.02079116925597191,
0.05404400825500488,
-0.035715050995349884,
0.025480998679995537,
0.12127705663442612,
0.046629149466753006,
0.029061047360301018,
0.06440366059541702,
0.04194325953722,
-0.07477899640798569,
0.015099985525012016,
-0.0012284832773730159,
0.04057411849498749,
-0.012318484485149384,
-4.4157065559602415e-8,
-0.06646599620580673,
-0.01683155633509159,
-0.034908220171928406,
0.022343652322888374,
-0.07539607584476471,
-0.08288123458623886,
0.05016703903675079,
0.0929991826415062,
-0.03682025149464607,
-0.03696789592504501,
0.02690015360713005,
-0.06419200450181961,
-0.05843263864517212,
-0.06867394596338272,
0.030687376856803894,
0.07548514008522034,
0.050194475799798965,
-0.024296941235661507,
-0.0017321930499747396,
0.018862854689359665,
-0.06898535788059235,
0.056215692311525345,
0.03675903379917145,
-0.027826271951198578,
-0.00545813562348485,
0.010175307281315327,
-0.02336110547184944,
0.04988442361354828,
0.07629169523715973,
-0.10563335567712784,
0.05698581039905548,
0.036229755729436874,
-0.01753985323011875,
-0.030809717252850533,
-0.04230751097202301,
0.009708927012979984,
0.06441028416156769,
0.03720378875732422,
-0.06334581971168518,
-0.01356325950473547,
0.03211643919348717,
0.025226032361388206,
0.031480684876441956,
0.06846402585506439,
-0.0071126245893538,
0.060536302626132965,
0.005747718270868063,
-0.05356697365641594,
0.036114588379859924,
-0.10629065334796906,
-0.0031227641738951206,
-0.01869106851518154,
-0.0739533081650734,
0.10240227729082108,
-0.01033894345164299,
0.018039170652627945,
0.0686844140291214,
0.021237226203083992,
0.04321958124637604,
0.011543689295649529,
0.02830079384148121,
0.03319667652249336,
-0.04323357343673706,
0.01752898097038269
] |
Fast Feedforward 3D Gaussian Splatting Compression | https://openreview.net/forum?id=DCandSZ2F1 | [
"Yihang Chen",
"Qianyi Wu",
"Mengyao Li",
"Weiyao Lin",
"Mehrtash Harandi",
"Jianfei Cai"
] | Poster | With 3D Gaussian Splatting (3DGS) advancing real-time and high-fidelity rendering for novel view synthesis, storage requirements pose challenges for their widespread adoption. Although various compression techniques have been proposed, previous art suffers from a common limitation: for any existing 3DGS, per-scene optimization is needed to achieve compression, making the compression sluggish and slow. To address this issue, we introduce Fast Compression of 3D Gaussian Splatting (FCGS), an optimization-free model that can compress 3DGS representations rapidly in a single feed-forward pass, which significantly reduces compression time from minutes to seconds. To enhance compression efficiency, we propose a multi-path entropy module that assigns Gaussian attributes to different entropy constraint paths for balance between size and fidelity. We also carefully design both inter- and intra-Gaussian context models to remove redundancies among the unstructured Gaussian blobs. Overall, FCGS achieves a compression ratio of over 20X while maintaining fidelity, surpassing most per-scene SOTA optimization-based methods. Code: github.com/YihangChen-ee/FCGS. | 3DGS, compression, optimization-free | A 3DGS compression framework that compresses any existing 3DGS rapidly without fine-tuning. | 702 | 2410.08017 | [
-0.007369586732238531,
-0.06243063136935234,
-0.018216155469417572,
-0.005581308621913195,
0.04950786009430885,
-0.027009349316358566,
-0.02687334641814232,
-0.09634170681238174,
0.04381377995014191,
-0.09837744385004044,
-0.05334530398249626,
-0.029696406796574593,
0.048869531601667404,
0.051405344158411026,
-0.008596100844442844,
0.009110272862017155,
0.07790457457304001,
0.10299385339021683,
-0.09817646443843842,
0.06686802953481674,
0.06787249445915222,
-0.06289750337600708,
0.000013514958482119255,
-0.01567758247256279,
0.0754670649766922,
-0.039891038089990616,
0.05064299702644348,
-0.04223369061946869,
0.01875898614525795,
0.0006176747265271842,
0.024868987500667572,
-0.018384352326393127,
0.020940087735652924,
0.02250470593571663,
-0.014236893504858017,
0.04597209021449089,
-0.04103942960500717,
-0.01740836724638939,
-0.047575268894433975,
0.0073131113313138485,
-0.016341153532266617,
0.11872486770153046,
-0.025379134342074394,
0.06969177722930908,
0.03609076142311096,
0.056176211684942245,
-0.010179402306675911,
-0.08375154435634613,
-0.0024160295724868774,
-0.02265341393649578,
-0.050611790269613266,
0.009575929492712021,
-0.07080524414777756,
0.028881842270493507,
0.03353114426136017,
0.05292101949453354,
-0.03906533122062683,
-0.015376051887869835,
-0.01898658461868763,
0.028148382902145386,
-0.00855900440365076,
-0.0016217561205849051,
0.03735736757516861,
-0.02210049144923687,
0.016157615929841995,
0.00571026373654604,
0.0649522989988327,
-0.03211478143930435,
-0.04100419208407402,
0.06401451677083969,
-0.09957260638475418,
-0.00415822584182024,
-0.05152188986539841,
0.015436467714607716,
-0.030219977721571922,
0.05563627928495407,
0.04420306906104088,
-0.02745700627565384,
-0.03697909042239189,
-0.060941439121961594,
0.047052301466464996,
-0.03324945643544197,
0.031857091933488846,
0.061317846179008484,
0.01956879533827305,
-0.025479862466454506,
0.00762089341878891,
0.10177711397409439,
0.0003222628147341311,
0.02986067347228527,
-0.0035884773824363947,
0.0024680134374648333,
0.009057934395968914,
-0.05426103249192238,
-0.020645126700401306,
-0.034502170979976654,
-0.01148658525198698,
-0.056573834270238876,
0.05335884541273117,
-0.0237522404640913,
0.019724778831005096,
-0.06496013700962067,
0.07177647948265076,
0.0036828021984547377,
0.07890410721302032,
-0.039300549775362015,
0.09112785011529922,
0.07668183743953705,
-0.013330609537661076,
0.039484694600105286,
0.03948250412940979,
0.06670071929693222,
0.011223585344851017,
-0.06277873367071152,
-0.0015101716853678226,
0.0967884287238121,
-0.06529507040977478,
-0.02627701871097088,
0.009819209575653076,
0.08062417060136795,
-0.005278402008116245,
-0.022735409438610077,
-0.09841568768024445,
-0.007710190024226904,
-0.014970767311751842,
-0.0821484923362732,
-0.06927391141653061,
2.638215603480216e-33,
0.005384594667702913,
0.047521330416202545,
0.0015107258222997189,
0.029392531141638756,
0.02982363849878311,
0.09697430580854416,
-0.006596413906663656,
0.006759313866496086,
-0.049284543842077255,
0.038327816873788834,
-0.0633474662899971,
-0.025587692856788635,
-0.06309448182582855,
0.09045863151550293,
-0.0026659974828362465,
0.018850723281502724,
-0.06934383511543274,
-0.005043437704443932,
-0.007496868725866079,
0.04406770318746567,
0.024251002818346024,
-0.017706185579299927,
-0.016930077224969864,
-0.05662008002400398,
-0.04024099186062813,
-0.003566066501662135,
0.02521795593202114,
-0.07266289740800858,
-0.03072522021830082,
-0.02250365912914276,
-0.04539365693926811,
-0.0068669444881379604,
-0.02050723321735859,
-0.05183306708931923,
0.017880836501717567,
-0.07977882027626038,
-0.0276301521807909,
-0.039775095880031586,
-0.04170602560043335,
-0.002474089851602912,
-0.06231553852558136,
0.08858154714107513,
-0.012108066119253635,
-0.01491235475987196,
-0.1296461969614029,
0.08052167296409607,
0.09691788256168365,
0.08389537036418915,
-0.01149629894644022,
0.02312658354640007,
0.09458616375923157,
0.01827714778482914,
-0.020905431360006332,
0.018863065168261528,
-0.014872264117002487,
-0.06701870262622833,
0.06850256770849228,
-0.031439732760190964,
0.08121280372142792,
0.04884402081370354,
-0.015896860510110855,
0.015175693668425083,
0.001731258351355791,
0.049556344747543335,
-0.05459346994757652,
0.012584885582327843,
0.05810791254043579,
0.05649936571717262,
0.04310303553938866,
-0.0051541137509047985,
-0.07791029661893845,
-0.031282950192689896,
0.02557012252509594,
-0.036425597965717316,
-0.04986098036170006,
-0.02335205487906933,
-0.018697477877140045,
-0.07175258547067642,
0.004772378597408533,
0.033547788858413696,
-0.1264028251171112,
0.02396700717508793,
0.004251603037118912,
-0.014794188551604748,
-0.02121194265782833,
-0.03517477586865425,
-0.050596777349710464,
-0.0001224069419549778,
-0.08236904442310333,
-0.05172993615269661,
-0.01652117818593979,
-0.10796000808477402,
0.03573707118630409,
-0.09772329032421112,
-0.06053588166832924,
-2.2102484160595534e-33,
0.052567772567272186,
-0.022427145391702652,
-0.05521942675113678,
0.02429589070379734,
0.03939376771450043,
-0.05847698822617531,
0.017213603481650352,
0.009529642760753632,
0.07994896918535233,
-0.008882132358849049,
-0.039006441831588745,
0.03608003631234169,
0.0323391817510128,
0.0665687546133995,
-0.008126894943416119,
0.025757458060979843,
-0.0290276650339365,
-0.0150612723082304,
-0.07465468347072601,
-0.08851839601993561,
-0.008986297994852066,
0.025866249576210976,
-0.025812067091464996,
0.009183825924992561,
-0.04560806602239609,
0.13035403192043304,
-0.024335162714123726,
0.007787838578224182,
-0.005454637575894594,
0.02912154048681259,
-0.040016256272792816,
-0.0009761644760146737,
0.014804130420088768,
-0.06822992861270905,
-0.048313505947589874,
-0.002690064487978816,
0.05217771232128143,
0.04003048688173294,
-0.03846791386604309,
-0.020367247983813286,
0.007642970886081457,
0.040103767067193985,
0.016360606998205185,
0.08132178336381912,
0.027769947424530983,
0.06604138761758804,
-0.018546653911471367,
-0.06145656108856201,
0.05796068534255028,
0.14589133858680725,
-0.08835537731647491,
0.06748773157596588,
-0.05953391268849373,
0.010493701323866844,
0.024190513417124748,
-0.13890281319618225,
-0.20010440051555634,
-0.04791480675339699,
0.08610115945339203,
-0.006960026454180479,
0.00856164749711752,
-0.045228395611047745,
0.00223156763240695,
-0.06634194403886795,
-0.026058688759803772,
0.014590374194085598,
0.027916425839066505,
-0.03906874358654022,
0.009574406780302525,
0.08487008512020111,
-0.037695907056331635,
0.020162636414170265,
-0.014711789786815643,
0.08551843464374542,
-0.015595834702253342,
0.018266994506120682,
0.047955892980098724,
-0.005818156059831381,
0.08893801271915436,
0.05072791501879692,
0.04903705418109894,
0.041915517300367355,
0.02752825990319252,
0.12679749727249146,
0.07156706601381302,
0.045486487448215485,
-0.051118604838848114,
0.04045964404940605,
0.03775149956345558,
-0.005878286901861429,
0.0093543017283082,
0.06193624809384346,
0.06523914635181427,
0.06811200082302094,
-0.0019045916851609945,
-4.44978347502456e-8,
-0.019017761573195457,
0.019306661561131477,
0.015298692509531975,
-0.04228485748171806,
-0.02425679937005043,
-0.031255606561899185,
0.0023304775822907686,
0.065785713493824,
0.0319950245320797,
-0.10088630020618439,
0.11846059560775757,
-0.0633542463183403,
-0.04947506636381149,
-0.015610349364578724,
0.027314849197864532,
0.030501781031489372,
-0.0391199104487896,
-0.045951958745718,
0.009775967337191105,
-0.015127776190638542,
-0.06131340563297272,
0.034519702196121216,
-0.03959864005446434,
0.05185786634683609,
0.012561172246932983,
-0.00022578694915864617,
-0.006710920017212629,
0.047098156064748764,
0.12921203672885895,
-0.023597178980708122,
0.00925214122980833,
-0.005567335989326239,
0.021454796195030212,
0.0006845116149634123,
-0.11659598350524902,
-0.013741745613515377,
-0.0272003635764122,
0.025689747184515,
0.02975429780781269,
-0.03167897090315819,
0.03914999961853027,
-0.012525925412774086,
0.022779613733291626,
-0.016035208478569984,
-0.00014734109572600573,
0.019580528140068054,
0.0606578029692173,
-0.09763400256633759,
0.001300021423958242,
0.06629050523042679,
0.006492955610156059,
-0.02943328581750393,
-0.05905202031135559,
0.09541906416416168,
0.10747813433408737,
-0.0463104285299778,
0.00915540661662817,
0.030427543446421623,
0.12871040403842926,
0.016122428700327873,
-0.04991947486996651,
-0.005665207281708717,
-0.02037360891699791,
-0.012883339077234268
] |
Scalable Benchmarking and Robust Learning for Noise-Free Ego-Motion and 3D Reconstruction from Noisy Video | https://openreview.net/forum?id=Pz9zFea4MQ | [
"Xiaohao Xu",
"Tianyi Zhang",
"Shibo Zhao",
"Xiang Li",
"Sibo Wang",
"Yongqi Chen",
"Ye Li",
"Bhiksha Raj",
"Matthew Johnson-Roberson",
"Sebastian Scherer",
"Xiaonan Huang"
] | Poster | We aim to redefine robust ego-motion estimation and photorealistic 3D reconstruction by addressing a critical limitation: the reliance on noise-free data in existing models. While such sanitized conditions simplify evaluation, they fail to capture the unpredictable, noisy complexities of real-world environments. Dynamic motion, sensor imperfections, and synchronization perturbations lead to sharp performance declines when these models are deployed in practice, revealing an urgent need for frameworks that embrace and excel under real-world noise.
To bridge this gap, we tackle three core challenges: scalable data generation, comprehensive benchmarking, and model robustness enhancement. First, we introduce a scalable noisy data synthesis pipeline that generates diverse datasets simulating complex motion, sensor imperfections, and synchronization errors. Second, we leverage this pipeline to create Robust-Ego3D, a benchmark rigorously designed to expose noise-induced performance degradation, highlighting the limitations of current learning-based methods in ego-motion accuracy and 3D reconstruction quality. Third, we propose Correspondence-guided Gaussian Splatting (CorrGS), a novel method that progressively refines an internal clean 3D representation by aligning noisy observations with rendered RGB-D frames from clean 3D map, enhancing geometric alignment and appearance restoration through visual correspondence.
Extensive experiments on synthetic and real-world data demonstrate that CorrGS consistently outperforms prior state-of-the-art methods, particularly in scenarios involving rapid motion and dynamic illumination. We will release our code and benchmark to advance robust 3D vision, setting a new standard for ego-motion estimation and high-fidelity reconstruction in noisy environments. | Benchmarking, Robustness, Neural 3D Reconstruction, Ego-Motion Estimation, SLAM | We reframe ego-motion estimation and photorealistic 3D reconstruction under generalized noisy sensing conditions as model-data feedback loop framework. | 696 | 2501.14319 | [
-0.07235017418861389,
-0.10320672392845154,
0.03605259582400322,
0.022400619462132454,
0.05402318760752678,
-0.027181431651115417,
-0.037289880216121674,
-0.11950002610683441,
-0.011625883169472218,
-0.049581464380025864,
-0.0031078080646693707,
-0.0383247509598732,
-0.031039204448461533,
0.022898510098457336,
-0.07376313209533691,
-0.0013647557934746146,
0.09997360408306122,
0.1387934386730194,
-0.09421539306640625,
0.04268476739525795,
-0.034038253128528595,
-0.030543729662895203,
-0.0206814743578434,
0.06922519952058792,
0.10968300700187683,
0.018972482532262802,
0.0419309139251709,
-0.04600882902741432,
-0.07163511216640472,
-0.016697155311703682,
0.0350654236972332,
-0.11273761093616486,
-0.033614382147789,
-0.05891435220837593,
0.004327310714870691,
0.011241444386541843,
0.014076643623411655,
-0.02841864712536335,
0.02586016245186329,
-0.00914798490703106,
0.0205101128667593,
0.05203163996338844,
-0.007477110717445612,
-0.017168061807751656,
0.05492837727069855,
0.07299047708511353,
0.06257251650094986,
-0.09504726529121399,
0.039825838059186935,
0.0022224271669983864,
-0.0882255956530571,
-0.0014116540551185608,
-0.038037367165088654,
0.01592644490301609,
0.05350778251886368,
0.054100051522254944,
0.028279587626457214,
-0.07519719749689102,
0.036740198731422424,
-0.008196404203772545,
0.08085513114929199,
-0.039684560149908066,
0.021471036598086357,
0.022854071110486984,
-0.018917562440037727,
-0.0642317607998848,
0.03240647539496422,
-0.010894098319113255,
0.0293567031621933,
0.060168053954839706,
-0.09531378746032715,
0.04547719284892082,
0.013937356881797314,
0.027207542210817337,
-0.003148839343339205,
0.0461709089577198,
0.007093758787959814,
-0.03661070764064789,
-0.0032452659215778112,
-0.03625673055648804,
0.06618117541074753,
-0.034287773072719574,
-0.027109825983643532,
-0.0034373048692941666,
0.06182456389069557,
-0.007564036641269922,
0.01887628063559532,
0.13382035493850708,
-0.003911575768142939,
-0.0003052347165066749,
-0.1055721864104271,
0.0008812296437099576,
-0.10621548444032669,
-0.04477154091000557,
0.037925418466329575,
-0.015564457513391972,
0.026988741010427475,
-0.059869974851608276,
0.05622527748346329,
0.02897326461970806,
0.016598738729953766,
0.033705003559589386,
-0.012346507981419563,
0.02886923775076866,
0.05717566981911659,
-0.004549442790448666,
0.059975847601890564,
0.06663940101861954,
-0.02285061776638031,
0.009788802824914455,
0.03463927283883095,
-0.02876235544681549,
0.05897556245326996,
-0.03726036474108696,
0.01670685037970543,
0.016263574361801147,
-0.04515187442302704,
0.020950987935066223,
0.005448297597467899,
0.052942562848329544,
0.032542385160923004,
-0.04828638955950737,
0.0027455443050712347,
0.05617915093898773,
0.031136341392993927,
-0.026144951581954956,
-0.0182400681078434,
-4.43436096146759e-34,
0.048310667276382446,
0.08589228987693787,
0.06551022082567215,
0.0013074628077447414,
0.08577640354633331,
0.021548593416810036,
-0.02177683636546135,
-0.05830799788236618,
-0.04720184579491615,
0.008864878676831722,
0.01635071262717247,
-0.07185837626457214,
-0.031181560829281807,
0.052590202540159225,
-0.01723242551088333,
0.025538407266139984,
-0.051698703318834305,
-0.028743257746100426,
0.022208306938409805,
0.11258015036582947,
0.08089709281921387,
-0.07559782266616821,
-0.03767261654138565,
0.012310820631682873,
-0.04448695480823517,
0.038461025804281235,
0.044775381684303284,
0.11714766174554825,
0.0048582409508526325,
-0.01667110249400139,
-0.045247677713632584,
-0.023124195635318756,
0.004006931558251381,
0.04267461225390434,
0.049803394824266434,
-0.04889584332704544,
-0.016487881541252136,
-0.07197529077529907,
-0.04981384426355362,
-0.023110071197152138,
0.01703726127743721,
0.10368119925260544,
-0.04573893919587135,
0.008029190823435783,
-0.061828725039958954,
0.011291188187897205,
0.07180490344762802,
0.09408050030469894,
-0.05012303963303566,
-0.015079534612596035,
0.0718778669834137,
-0.005395606160163879,
-0.029580876231193542,
-0.0658816546201706,
-0.10962916910648346,
-0.044830385595560074,
0.09558403491973877,
-0.05386798083782196,
0.01890980452299118,
0.045761097222566605,
-0.050767142325639725,
0.017462490126490593,
-0.01543509028851986,
-0.04061567038297653,
-0.011223972775042057,
0.03863736242055893,
0.02609526365995407,
0.10456478595733643,
0.03763660788536072,
-0.021659022197127342,
0.016574015840888023,
-0.037787582725286484,
-0.03154266998171806,
0.007434636354446411,
0.07095678150653839,
0.048569537699222565,
0.01117168739438057,
-0.005419503431767225,
0.028642544522881508,
0.010406364686787128,
-0.108616903424263,
0.004424575250595808,
-0.0024910406209528446,
0.02100965566933155,
-0.06948557496070862,
-0.04331832751631737,
-0.06626737862825394,
-0.005764438305050135,
-0.055129729211330414,
-0.006938050501048565,
-0.04546356573700905,
-0.08476134389638901,
-0.0033675089944154024,
-0.0114098796620965,
-0.06567761301994324,
5.888889661348761e-34,
0.0856504887342453,
0.016641896218061447,
-0.021114787086844444,
0.09242583066225052,
-0.005523369647562504,
-0.03730808570981026,
-0.016927722841501236,
0.01702532358467579,
0.05787085369229317,
-0.07337535917758942,
-0.03738866001367569,
-0.057958781719207764,
0.04066416993737221,
0.04627854749560356,
0.004827269818633795,
-0.0070966687053442,
0.014059259556233883,
-0.05489009991288185,
-0.05397491529583931,
-0.01580946333706379,
0.034442558884620667,
0.007315031718462706,
0.01664605736732483,
0.03615734726190567,
-0.023395463824272156,
0.12093141674995422,
-0.0674857646226883,
-0.0031028997618705034,
0.040194205939769745,
-0.021290618926286697,
0.01039133407175541,
-0.008632196113467216,
0.024906741455197334,
-0.07264887541532516,
-0.08685746043920517,
0.08420995622873306,
0.014438789337873459,
-0.03080759570002556,
0.00009611930727260187,
0.014841579832136631,
0.03572733327746391,
0.005579323507845402,
-0.07801522314548492,
0.03682532534003258,
0.0558931864798069,
0.032383933663368225,
-0.01430431380867958,
0.00938291847705841,
-0.02127031236886978,
0.06883957982063293,
-0.10123533755540848,
0.12905800342559814,
-0.06699196249246597,
0.00011524791625561193,
0.03548891842365265,
-0.0618763193488121,
-0.11239767074584961,
-0.02123303897678852,
0.12702548503875732,
-0.0042871152982115746,
-0.04009202495217323,
-0.005553601309657097,
-0.057519566267728806,
-0.001543295569717884,
-0.023191463202238083,
0.061058804392814636,
-0.005196101497858763,
-0.026755372062325478,
0.02201290987432003,
0.0687701553106308,
0.037578318268060684,
-0.01567750796675682,
-0.0201242808252573,
-0.014883599244058132,
0.0019884733483195305,
-0.030314777046442032,
0.02597067505121231,
-0.01057523861527443,
0.053765084594488144,
0.047784529626369476,
0.07661481946706772,
-0.06701405346393585,
0.11712731420993805,
0.11240383982658386,
0.017622165381908417,
0.0701291561126709,
0.006250509526580572,
0.029971623793244362,
-0.004181370139122009,
0.04453689977526665,
-0.006547574885189533,
0.043666139245033264,
-0.05284326896071434,
0.0018764560809358954,
-0.043676503002643585,
-3.628879596817569e-8,
-0.1045118197798729,
0.03130210191011429,
0.019940055906772614,
-0.039387527853250504,
-0.04982516169548035,
-0.035846684128046036,
0.015973396599292755,
0.05434663966298103,
0.08731426298618317,
-0.08765346556901932,
0.05385105684399605,
-0.10826669633388519,
-0.021572740748524666,
-0.016293037682771683,
-0.04611702263355255,
0.049844346940517426,
-0.02555638551712036,
0.00101229059509933,
-0.014318058267235756,
0.05094987899065018,
-0.12509147822856903,
0.06550392508506775,
-0.032894283533096313,
-0.029083145782351494,
0.06512977182865143,
0.0007496680482290685,
0.004545008763670921,
0.05409381538629532,
0.08096183836460114,
-0.026013173162937164,
-0.019188113510608673,
-0.03040289878845215,
0.10913225263357162,
-0.059148672968149185,
-0.07060227543115616,
-0.005104823503643274,
-0.002277816878631711,
0.012189590372145176,
0.03123978152871132,
-0.10673999041318893,
-0.04339174926280975,
0.08799482882022858,
-0.013539664447307587,
-0.00632724491879344,
-0.021424537524580956,
0.03281863406300545,
0.0025444095954298973,
-0.08206530660390854,
-0.017017411068081856,
-0.0023299993481487036,
0.06000996381044388,
-0.00849198643118143,
-0.025744538754224777,
0.035860300064086914,
0.08738385885953903,
0.04105713218450546,
0.021235166117548943,
-0.059100937098264694,
0.03412949666380882,
0.06956371665000916,
-0.007623214740306139,
-0.06851904839277267,
-0.01798473857343197,
0.0295159500092268
] |
LaMP: Language-Motion Pretraining for Motion Generation, Retrieval, and Captioning | https://openreview.net/forum?id=LYawG8YkPa | [
"Zhe Li",
"Weihao Yuan",
"Yisheng HE",
"Lingteng Qiu",
"Shenhao Zhu",
"Xiaodong Gu",
"Weichao Shen",
"Yuan Dong",
"Zilong Dong",
"Laurence Tianruo Yang"
] | Poster | Language plays a vital role in the realm of human motion. Existing methods have largely depended on CLIP text embeddings for motion generation, yet they fall short in effectively aligning language and motion due to CLIP’s pretraining on static image-text pairs. This work introduces LaMP, a novel Language-Motion Pretraining model, which transitions from a language-vision to a more suitable language-motion latent space. It addresses key limitations by generating motion-informative text embeddings, significantly enhancing the relevance and semantics of generated motion sequences. With LaMP, we advance three key tasks: text-to-motion generation, motion-text retrieval, and motion captioning through aligned language-motion representation learning. For generation, LaMP instead of CLIP provides the text condition, and an autoregressive masked prediction is designed to achieve mask modeling without rank collapse in transformers. For retrieval, motion features from LaMP’s motion transformer interact with query tokens to retrieve text features from the text transformer, and vice versa. For captioning, we finetune a large language model with the language-informative motion features to develop a strong motion captioning model. In addition, we introduce the LaMP-BertScore metric to assess the alignment of generated motions with textual descriptions. Extensive experimental results on multiple datasets demonstrate substantial improvements over previous methods across all three tasks. Project page: https://aigc3d.github.io/LaMP | motion generation; motion-language pretraining; multimodal | We propose LaMP, a novel language-motion pretraining model bridging language and motion gaps, enhancing motion sequence relevance and semantics in text-to-motion generation, retrieval, and captioning. | 693 | 2410.07093 | [
-0.06707915663719177,
-0.11637741327285767,
0.027681082487106323,
-0.019064804539084435,
0.04764906316995621,
0.089045450091362,
-0.01649869605898857,
-0.014597507193684578,
0.06595580279827118,
-0.037968575954437256,
0.028832819312810898,
-0.05780840665102005,
0.006321218330413103,
0.053281135857105255,
-0.015283911488950253,
-0.008993140421807766,
0.015053837560117245,
0.0774483010172844,
-0.03868648782372475,
-0.05493113771080971,
0.10467717796564102,
-0.020627528429031372,
0.07022134959697723,
-0.0007362373871728778,
0.09081583470106125,
0.015656601637601852,
-0.0021658753976225853,
-0.02737196534872055,
0.04327692091464996,
-0.013421020470559597,
0.005498959217220545,
0.02448693849146366,
0.01224333792924881,
0.06816806644201279,
-0.10144949704408646,
0.10796856135129929,
-0.05149561166763306,
-0.041087523102760315,
-0.019359765574336052,
-0.04446248710155487,
-0.025147458538413048,
-0.039207667112350464,
0.03753780946135521,
0.016940230503678322,
0.08629324287176132,
0.03561360016465187,
-0.01164550706744194,
-0.06582293659448624,
-0.011329623870551586,
0.003538542427122593,
-0.07705137878656387,
-0.008941402658820152,
0.05707336217164993,
0.052552320063114166,
-0.021356360986828804,
-0.013694249093532562,
0.0450507253408432,
-0.03650830686092377,
0.07203047722578049,
-0.058301642537117004,
0.027928810566663742,
-0.03597869351506233,
0.0881890133023262,
-0.007427127566188574,
-0.04137694463133812,
-0.05414781719446182,
0.029603656381368637,
-0.020582180470228195,
-0.006951027549803257,
0.009455726481974125,
-0.11626828461885452,
-0.018487613648176193,
0.009511670097708702,
0.0941532775759697,
-0.0749443769454956,
0.07786750048398972,
0.07047408074140549,
-0.09250248968601227,
-0.02448238991200924,
-0.07380218058824539,
0.041415438055992126,
-0.06637931615114212,
0.053286828100681305,
0.0782460868358612,
0.06167837232351303,
-0.002686323132365942,
0.03261726722121239,
0.05163953825831413,
-0.02049930952489376,
0.00040329317562282085,
-0.09873683005571365,
-0.060990333557128906,
0.0993131548166275,
-0.03047202154994011,
0.027929464355111122,
0.034690625965595245,
-0.03889956325292587,
-0.06422264128923416,
0.021050482988357544,
0.07420984655618668,
0.04221037030220032,
0.0558820404112339,
-0.01732989400625229,
0.01722036302089691,
-0.01927553489804268,
-0.029284296557307243,
0.04789694771170616,
0.043960411101579666,
0.0008853671606630087,
-0.06470835208892822,
0.06503105163574219,
-0.02044833078980446,
-0.016973072662949562,
0.019840223714709282,
0.0319591648876667,
0.004780198913067579,
0.01119141560047865,
-0.04720156639814377,
-0.01787360943853855,
0.03802795335650444,
0.03422205150127411,
0.08606462180614471,
-0.061833105981349945,
-0.0023091889452189207,
0.03554891422390938,
-0.07429859042167664,
-0.013836558908224106,
-1.464107233965945e-33,
0.06259392201900482,
0.03011935018002987,
0.054436735808849335,
0.020998578518629074,
0.033189065754413605,
-0.007487378548830748,
0.0031728658359497786,
0.07886762171983719,
-0.024007946252822876,
-0.05299784988164902,
0.0284944586455822,
-0.025995727628469467,
-0.02062712423503399,
0.06761793792247772,
-0.08470074832439423,
-0.024048717692494392,
0.0006175953894853592,
-0.025019856169819832,
0.019747963175177574,
0.05643994361162186,
-0.013786601833999157,
0.004540415480732918,
-0.0056932661682367325,
-0.010990936309099197,
0.009229764342308044,
0.021784404292702675,
0.040217749774456024,
-0.12496744841337204,
-0.025818748399615288,
-0.020107729360461235,
-0.1435893028974533,
-0.030769377946853638,
0.034540511667728424,
-0.04250624030828476,
0.05019301921129227,
-0.0500318668782711,
-0.018592629581689835,
-0.07412457466125488,
0.0025901382323354483,
-0.012261255644261837,
-0.021026888862252235,
0.00596403144299984,
-0.07387851178646088,
-0.08129402250051498,
-0.06606477499008179,
-0.05786930397152901,
0.03006669320166111,
0.0948048084974289,
-0.0930960401892662,
0.03252853825688362,
0.06241828575730324,
0.03449555113911629,
-0.0764390230178833,
-0.06159874424338341,
0.002314997371286154,
-0.03775028884410858,
0.04820898920297623,
-0.03232036158442497,
0.026156125590205193,
0.01587056927382946,
-0.03904155641794205,
0.02435280568897724,
0.03071131370961666,
-0.03558472543954849,
0.060458749532699585,
0.002407602733001113,
0.06560835987329483,
0.0660911574959755,
0.05101725459098816,
0.01367856189608574,
-0.05853073671460152,
-0.029225386679172516,
-0.06215745583176613,
-0.005186989437788725,
0.04308265820145607,
0.03987966850399971,
0.015262829139828682,
-0.10714009404182434,
0.0025263784918934107,
0.0589364692568779,
-0.07020749151706696,
-0.06909862160682678,
0.06784375756978989,
-0.07069551199674606,
-0.03476036712527275,
-0.05456091836094856,
0.062052588909864426,
-0.08532820641994476,
-0.008822181262075901,
-0.053334277123212814,
-0.008084583096206188,
-0.013797575607895851,
-0.010296742431819439,
0.05037450045347214,
0.04212406277656555,
1.0715922214821132e-33,
0.03664467856287956,
0.05992741882801056,
-0.09417645633220673,
0.02478381246328354,
-0.022459132596850395,
-0.008629056625068188,
0.06022869423031807,
0.005028951913118362,
-0.01999606192111969,
-0.03954442963004112,
-0.013002142310142517,
-0.04577767103910446,
-0.003863532794639468,
-0.062202103435993195,
0.04778238758444786,
0.021897438913583755,
0.049067478626966476,
-0.019546041265130043,
-0.03312890976667404,
0.0749795213341713,
-0.006280178669840097,
0.09505759179592133,
-0.068001389503479,
0.01669827662408352,
-0.05714825913310051,
0.07710402458906174,
0.05869918689131737,
0.02578483335673809,
-0.06799899786710739,
0.018032491207122803,
0.015979671850800514,
-0.04513448104262352,
0.010679838247597218,
-0.030335400253534317,
-0.09600331634283066,
0.025684228166937828,
0.0257481150329113,
-0.027098435908555984,
-0.041716787964105606,
0.05458386242389679,
0.13164760172367096,
-0.04384744167327881,
0.014582241885364056,
0.02883601374924183,
-0.11751925200223923,
-0.07071850448846817,
-0.0811488926410675,
0.03328617662191391,
0.016646219417452812,
0.05273907631635666,
-0.04524283856153488,
0.004451208282262087,
-0.10806428641080856,
-0.06860574334859848,
-0.029538821429014206,
-0.03399011120200157,
-0.09033183753490448,
-0.014204634353518486,
0.04366070032119751,
-0.029040174558758736,
0.00793757289648056,
0.0046269833110272884,
0.022828225046396255,
-0.028866011649370193,
0.0008694172138348222,
-0.026293739676475525,
-0.024813704192638397,
0.005378279834985733,
-0.0184064581990242,
-0.02467302419245243,
0.10049065202474594,
-0.016929471865296364,
0.06320764124393463,
0.036396101117134094,
0.03842233121395111,
-0.019425349310040474,
0.0773862823843956,
0.0668116882443428,
0.03857524320483208,
-0.10294689238071442,
0.017981622368097305,
-0.004502255469560623,
0.02208763360977173,
0.04400889202952385,
0.004524821415543556,
0.08801066875457764,
-0.05609698221087456,
0.05498313903808594,
0.013960933312773705,
0.06436928361654282,
-0.08857770264148712,
0.06831375509500504,
-0.04029523581266403,
0.034855280071496964,
-0.028067633509635925,
-3.1568514913260515e-8,
-0.13065820932388306,
0.08299664407968521,
-0.034740153700113297,
0.04876382276415825,
-0.03882639482617378,
-0.05584193393588066,
0.019015619531273842,
0.061399903148412704,
0.00423118332400918,
-0.06981908529996872,
0.09869126975536346,
-0.010430612601339817,
0.04995550215244293,
-0.028652358800172806,
-0.06021254137158394,
0.023799743503332138,
-0.010954081080853939,
0.035580918192863464,
-0.017128918319940567,
0.0289299413561821,
-0.021953541785478592,
0.012127029709517956,
-0.0063643911853432655,
-0.0282881036400795,
-0.005518315825611353,
0.03338506445288658,
-0.052172187715768814,
0.06838804483413696,
0.10958024114370346,
-0.02881518192589283,
0.024325963109731674,
0.019685635343194008,
0.05194931477308273,
-0.05257260426878929,
-0.008235056884586811,
0.05386219546198845,
-0.022223422303795815,
-0.08230512589216232,
0.011238589882850647,
-0.012638069689273834,
0.059476710855960846,
0.11229770630598068,
0.01461477018892765,
0.018790636211633682,
0.0028452863916754723,
0.009903945028781891,
0.036429427564144135,
-0.11659718304872513,
0.05491644889116287,
-0.02524644322693348,
0.0016746936598792672,
-0.09722809493541718,
0.02902381494641304,
0.09008534997701645,
0.08511617034673691,
0.06282531470060349,
0.025031205266714096,
0.02276765927672386,
0.03545060753822327,
0.04471178725361824,
-0.047094520181417465,
-0.004838505294173956,
-0.0046599810011684895,
0.011047272011637688
] |
Generating Physical Dynamics under Priors | https://openreview.net/forum?id=eNjXcP6C0H | [
"Zihan Zhou",
"Xiaoxue Wang",
"Tianshu Yu"
] | Poster | Generating physically feasible dynamics in a data-driven context is challenging, especially when adhering to physical priors expressed in specific equations or formulas. Existing methodologies often overlook the integration of ''physical priors'', resulting in violation of basic physical laws and suboptimal performance. In this paper, we introduce a novel framework that seamlessly incorporates physical priors into diffusion-based generative models to address this limitation. Our approach leverages two categories of priors: 1) distributional priors, such as roto-translational invariance, and 2) physical feasibility priors, including energy and momentum conservation laws and PDE constraints. By embedding these priors into the generative process, our method can efficiently generate physically realistic dynamics, encompassing trajectories and flows. Empirical evaluations demonstrate that our method produces high-quality dynamics across a diverse array of physical phenomena with remarkable robustness, underscoring its potential to advance data-driven studies in AI4Physics. Our contributions signify a substantial advancement in the field of generative modeling, offering a robust solution to generate accurate and physically consistent dynamics. | diffusion models, generative models, physical dynamics, priors | null | 688 | 2409.00730 | [
-0.07873546332120895,
-0.08011305332183838,
0.11071530729532242,
0.06909836083650589,
-0.016911882907152176,
-0.03868841007351875,
-0.10311660915613174,
-0.05254377797245979,
0.00947458017617464,
0.004046619404107332,
-0.019464822486042976,
-0.1026802584528923,
-0.031102238222956657,
0.02473524957895279,
-0.012217702344059944,
0.026102520525455475,
0.05086877942085266,
0.11327406018972397,
-0.09161978960037231,
-0.017910383641719818,
-0.010900411754846573,
0.07778209447860718,
-0.025595178827643394,
0.014932654798030853,
0.08249756693840027,
0.01692890375852585,
0.06754503399133682,
-0.002162839286029339,
0.007693973369896412,
-0.004779767710715532,
0.08485022932291031,
-0.006928991060703993,
-0.13566161692142487,
0.04213934764266014,
-0.012815000489354134,
0.07821337133646011,
-0.010713906027376652,
0.02840995416045189,
-0.030037924647331238,
-0.03405948728322983,
0.038715630769729614,
0.03914197161793709,
0.04017072543501854,
0.03619609773159027,
0.08795325458049774,
-0.022667931392788887,
0.024009469896554947,
-0.06878707557916641,
0.048673130571842194,
0.024087974801659584,
-0.015534250065684319,
0.01624831184744835,
0.009349904954433441,
0.02369365282356739,
0.029156027361750603,
0.017792576923966408,
0.09252084791660309,
-0.07924910634756088,
0.04203213378787041,
-0.05467618256807327,
0.0012762011028826237,
-0.02343759313225746,
-0.002603674540296197,
0.0015605248045176268,
-0.017527172341942787,
-0.04072181135416031,
0.042595673352479935,
0.1142238974571228,
0.037603288888931274,
0.05797262862324715,
-0.007991730235517025,
0.09795010089874268,
-0.10148017853498459,
0.010390788316726685,
0.060531169176101685,
-0.01823721081018448,
-0.002257012762129307,
0.001355443731881678,
-0.0033104526810348034,
-0.060522787272930145,
0.06820196658372879,
0.04247599467635155,
0.014832175336778164,
-0.027860725298523903,
-0.021936381235718727,
0.0838966816663742,
0.03825055807828903,
0.07719143480062485,
0.06254994124174118,
0.010969923809170723,
-0.08041927218437195,
0.02037987858057022,
-0.05981729179620743,
-0.029120709747076035,
0.014209403656423092,
0.0017085190629586577,
0.029362551867961884,
-0.1274726539850235,
0.07236070930957794,
0.02584894187748432,
-0.0023754630237817764,
0.03929174691438675,
-0.011910329572856426,
0.06688796728849411,
0.056605517864227295,
-0.07016061246395111,
0.03739520162343979,
0.05138329789042473,
-0.0548124760389328,
0.06682104617357254,
0.11592012643814087,
-0.0020623730961233377,
0.026479022577404976,
0.010836286470293999,
-0.05476115643978119,
-0.05927208811044693,
0.011062368750572205,
0.003627285361289978,
-0.008141949772834778,
-0.01434997282922268,
-0.01252537127584219,
-0.0027423417195677757,
-0.006851674057543278,
0.07034289836883545,
0.04615368694067001,
-0.09809526056051254,
-0.025803809985518456,
-9.04518382880276e-34,
-0.0035985857248306274,
-0.023783186450600624,
0.039238497614860535,
0.11719870567321777,
0.07131420075893402,
0.0006927357171662152,
-0.03114164248108864,
-0.04987318813800812,
-0.03839472681283951,
-0.00896212738007307,
-0.09140198677778244,
0.029757441952824593,
-0.09959006309509277,
0.027503851801156998,
0.026907367631793022,
-0.05880902335047722,
-0.025727178901433945,
0.04120171070098877,
-0.0039662024937570095,
0.041106533259153366,
0.012251663953065872,
-0.04727604612708092,
-0.06917937844991684,
-0.03261953219771385,
-0.06409068405628204,
0.08150339126586914,
0.022419577464461327,
0.01984243281185627,
0.034902285784482956,
-0.009737784042954445,
-0.11961081624031067,
-0.014116193167865276,
-0.05032523348927498,
0.0007546929409727454,
0.04490753263235092,
0.02384091168642044,
-0.006085751112550497,
-0.056491099298000336,
-0.07364827394485474,
0.012250533327460289,
-0.016443641856312752,
0.012601315043866634,
-0.05103812366724014,
0.026353521272540092,
-0.110074982047081,
-0.06429527699947357,
0.059496764093637466,
0.019398676231503487,
-0.0049734218046069145,
-0.04171786457300186,
-0.02534264512360096,
0.01699512079358101,
0.07981915026903152,
-0.056366823613643646,
-0.008834799751639366,
-0.07145166397094727,
0.01905824989080429,
-0.0961359441280365,
0.020089317113161087,
-0.01565210521221161,
0.014309454709291458,
0.07718882709741592,
-0.016525378450751305,
0.03626085817813873,
0.009391325525939465,
0.0022184113040566444,
-0.03903152048587799,
0.06897959113121033,
0.038752324879169464,
-0.012497016228735447,
0.0043510994873940945,
-0.012906152755022049,
-0.03152405470609665,
-0.045699868351221085,
0.10169779509305954,
0.03857656940817833,
0.0492832325398922,
-0.05322421342134476,
-0.007605941966176033,
-0.01669800840318203,
-0.07303682714700699,
-0.042802419513463974,
-0.021259676665067673,
0.02923625521361828,
0.009458575397729874,
0.01908641867339611,
-0.0314226970076561,
-0.034240927547216415,
-0.10582787543535233,
-0.06977558135986328,
-0.08300326019525528,
-0.06598404794931412,
0.010485328733921051,
-0.015892760828137398,
0.005632018204778433,
-2.6149564273682817e-33,
-0.04803527519106865,
0.04905111342668533,
0.01030740700662136,
0.0535578578710556,
-0.006440369877964258,
-0.044126663357019424,
-0.05654580146074295,
0.02542598731815815,
-0.021434908732771873,
-0.16262882947921753,
-0.04150301590561867,
-0.018686141818761826,
-0.03491416200995445,
-0.015104275196790695,
0.08322198688983917,
-0.010045449249446392,
-0.001984104048460722,
-0.038110408931970596,
0.0017622323939576745,
-0.00499257817864418,
-0.033374737948179245,
0.04558492451906204,
-0.05444098636507988,
-0.004673229530453682,
-0.021919922903180122,
0.0709678903222084,
-0.008778419345617294,
0.014182688668370247,
-0.03133787959814072,
0.023412607610225677,
-0.047897495329380035,
0.009236766956746578,
0.03998761251568794,
-0.0029442571103572845,
-0.1422325223684311,
0.11111344397068024,
0.13331946730613708,
0.035792551934719086,
-0.016850197687745094,
0.02604534849524498,
0.013314545154571533,
-0.05154167860746384,
-0.006564600393176079,
0.045189063996076584,
0.0019188578007742763,
-0.01565633900463581,
0.0409836545586586,
0.021695425733923912,
0.030920496210455894,
0.02262992039322853,
-0.013797467574477196,
0.0468171127140522,
-0.05484137311577797,
0.0016789709916338325,
-0.04582545533776283,
-0.06493621319532394,
-0.007438204251229763,
-0.07040434330701828,
0.045518483966588974,
-0.05648123845458031,
-0.07269256561994553,
-0.06739576160907745,
0.0657520666718483,
0.00847417488694191,
-0.07470308989286423,
-0.007492642849683762,
-0.06157434731721878,
-0.015710705891251564,
-0.03145721182227135,
-0.028218260034918785,
-0.017574094235897064,
-0.038304805755615234,
-0.00418659346178174,
0.02011510170996189,
0.10884401202201843,
-0.09467024356126785,
0.08388582617044449,
-0.006798085290938616,
0.03567099943757057,
-0.007197350729256868,
0.04442017152905464,
0.03573306277394295,
0.038527972996234894,
-0.011754455976188183,
0.04442932829260826,
-0.05814022198319435,
-0.06074085459113121,
-0.0768621563911438,
-0.011616863310337067,
0.07116173207759857,
0.04257882386445999,
0.02911803126335144,
0.005133912432938814,
0.023773858323693275,
-0.0858202651143074,
-4.181984536444361e-8,
0.014378347434103489,
0.036496203392744064,
0.08826308697462082,
-0.0232987180352211,
0.022982528433203697,
0.054572392255067825,
-0.004445184022188187,
-0.02567490004003048,
0.04389204829931259,
-0.11105073988437653,
0.03890523314476013,
-0.074200838804245,
0.010089805349707603,
-0.06126869097352028,
-0.003365903627127409,
0.10109193623065948,
-0.020879952237010002,
-0.037760935723781586,
-0.06724519282579422,
0.00762106291949749,
-0.020010849460959435,
0.0565665028989315,
-0.07044581323862076,
-0.02053019404411316,
0.08646673709154129,
0.008073121309280396,
-0.016504548490047455,
-0.03870436176657677,
0.060343801975250244,
0.01108134351670742,
-0.0517226979136467,
-0.01962997391819954,
0.021085599437355995,
0.02197800762951374,
-0.07536096125841141,
0.018790170550346375,
0.019066162407398224,
0.010982317849993706,
-0.041356444358825684,
-0.1043422669172287,
-0.04462527856230736,
0.09024505317211151,
-0.06352841109037399,
0.03465618938207626,
-0.006499267648905516,
0.08937455713748932,
-0.03979665786027908,
-0.004592798184603453,
0.06580260396003723,
0.135735422372818,
-0.018004266545176506,
0.008915245532989502,
0.047012366354465485,
0.037493519484996796,
0.006206704769283533,
0.06152929365634918,
-0.028642762452363968,
-0.019639013335108757,
0.0588439479470253,
-0.0540420264005661,
0.01003442332148552,
0.018817424774169922,
0.004792553838342428,
0.03146197646856308
] |
Denoising with a Joint-Embedding Predictive Architecture | https://openreview.net/forum?id=d4njmzM7jf | [
"Dengsheng Chen",
"Jie Hu",
"Xiaoming Wei",
"Enhua Wu"
] | Poster | Joint-embedding predictive architectures (JEPAs) have shown substantial promise in self-supervised representation learning, yet their application in generative modeling remains underexplored. Conversely, diffusion models have demonstrated significant efficacy in modeling arbitrary probability distributions. In this paper, we introduce Denoising with a Joint-Embedding Predictive Architecture (D-JEPA), pioneering the integration of JEPA within generative modeling. By recognizing JEPA as a form of masked image modeling, we reinterpret it as a generalized next-token prediction strategy, facilitating data generation in an auto-regressive manner. Furthermore, we incorporate diffusion loss to model the per-token probability distribution, enabling data generation in a continuous space. We also adapt flow matching loss as an alternative to diffusion loss, thereby enhancing the flexibility of D-JEPA. Empirically, with increased GFLOPs, D-JEPA consistently achieves lower FID scores with fewer training epochs, indicating its good scalability. Our base, large, and huge models outperform all previous generative models across all scales on ImageNet conditional generation benchmarks. Beyond image generation, D-JEPA is well-suited for other continuous data modeling, including video and audio. | AIGC, JEPA, Diffusion Model, Flow Matching, Image Synthetic | In this paper, we introduce Denoising with a Joint-Embedding Predictive Architecture (D-JEPA), pioneering the application of JEPA in generative modeling. | 682 | 2410.03755 | [
-0.06379424035549164,
-0.10189222544431686,
0.09441854059696198,
0.0446857213973999,
0.06503668427467346,
-0.020854046568274498,
-0.04652094468474388,
-0.037105709314346313,
0.06445565819740295,
-0.05699383094906807,
-0.05571393668651581,
-0.06376995146274567,
0.028167307376861572,
0.07154025137424469,
-0.043003667145967484,
-0.03632969409227371,
0.05587531253695488,
0.1010982096195221,
-0.10400775820016861,
-0.08041016012430191,
-0.00645019905641675,
0.03990009054541588,
0.019079232588410378,
-0.03022577613592148,
0.11957408487796783,
0.008865856565535069,
0.04202092066407204,
-0.080005943775177,
0.00025377102429047227,
-0.042081985622644424,
0.07445671409368515,
-0.06316554546356201,
-0.06949329376220703,
0.018464354798197746,
-0.023945633322000504,
0.006460786797106266,
-0.04216406121850014,
0.050204548984766006,
0.05340839549899101,
-0.048702847212553024,
0.03553535044193268,
0.013264589942991734,
0.04562763497233391,
0.03286096081137657,
0.04429559037089348,
0.03149278834462166,
-0.006589571479707956,
-0.0833500549197197,
0.0018346289871260524,
0.026084043085575104,
0.004634705372154713,
-0.041271667927503586,
-0.06614086776971817,
0.06480645388364792,
-0.03571678325533867,
0.011580361053347588,
0.05609201639890671,
0.010216254740953445,
-0.022132283076643944,
0.04503173753619194,
-0.038369063287973404,
-0.10909847915172577,
0.006428752560168505,
0.0067797270603477955,
-0.022607211023569107,
-0.02849443256855011,
0.0238689836114645,
0.10857493430376053,
-0.01874089241027832,
0.027680452913045883,
-0.007112491875886917,
0.033950336277484894,
-0.07702109217643738,
0.01981159672141075,
0.03111395053565502,
0.05360327661037445,
0.057335998862981796,
-0.028956806287169456,
0.07697637379169464,
-0.09180209040641785,
0.057954832911491394,
0.046999864280223846,
0.10336383432149887,
-0.022507207468152046,
0.09962274134159088,
0.001925553078763187,
0.017144082114100456,
0.06730733066797256,
-0.04050871729850769,
-0.03446240350604057,
-0.094096839427948,
-0.017050841823220253,
-0.02911021001636982,
0.0045188115909695625,
0.06262455880641937,
0.04649752378463745,
-0.009093177504837513,
-0.10209181904792786,
-0.058337677270174026,
0.09566057473421097,
-0.042094334959983826,
0.06260222941637039,
0.04188770428299904,
-0.019901664927601814,
0.06396107375621796,
-0.062396787106990814,
0.060653895139694214,
0.0433979295194149,
0.011922579258680344,
-0.06735630333423615,
0.01196998916566372,
0.035953227430582047,
0.021158583462238312,
-0.03210029378533363,
0.03105698898434639,
-0.020446917042136192,
0.0011826087720692158,
0.013682343065738678,
0.035248998552560806,
0.023335594683885574,
-0.017587140202522278,
-0.010280461981892586,
-0.02134915068745613,
0.032607123255729675,
0.020180508494377136,
-0.03622850775718689,
-0.04524712637066841,
3.02056534489432e-33,
0.001103910501115024,
-0.024495743215084076,
0.016834814101457596,
0.008474239148199558,
0.05242901295423508,
-0.025341415777802467,
-0.049658775329589844,
0.009348317049443722,
-0.02406049147248268,
-0.07168909162282944,
-0.04224134236574173,
-0.040376272052526474,
-0.11220689117908478,
0.12316355854272842,
0.02019323967397213,
0.021406985819339752,
-0.04710298031568527,
0.03602619469165802,
0.00807790644466877,
-0.03503913804888725,
0.0042316727340221405,
0.006828734651207924,
-0.05380350723862648,
-0.013663681223988533,
-0.006602875422686338,
0.008887127041816711,
0.07890302687883377,
-0.05807159096002579,
0.007400805130600929,
0.008086161687970161,
-0.14448454976081848,
0.046526405960321426,
0.021789195016026497,
-0.0023694175761193037,
0.032334670424461365,
-0.03338892385363579,
-0.03989719599485397,
-0.017740027979016304,
-0.03405710309743881,
-0.012484759092330933,
-0.023711876943707466,
0.05228213220834732,
-0.05811915546655655,
-0.026930643245577812,
-0.0743536427617073,
-0.11531195789575577,
0.0203254334628582,
0.019162241369485855,
-0.05713503807783127,
-0.04323730617761612,
0.09312324970960617,
-0.01623842492699623,
-0.06139038875699043,
-0.05219481140375137,
-0.059867240488529205,
-0.012639659456908703,
0.04011247679591179,
-0.024005159735679626,
0.02093658037483692,
-0.002452101092785597,
0.028505196794867516,
0.05215398967266083,
-0.01052902452647686,
0.06901515275239944,
0.0048034838400781155,
0.011313200928270817,
0.04242544248700142,
0.06545062363147736,
0.007764937821775675,
-0.006943987682461739,
0.0035336227156221867,
0.04508485272526741,
-0.02262718230485916,
-0.10887106508016586,
0.04120418801903725,
0.016947878524661064,
0.06925974041223526,
-0.0705966204404831,
-0.014245202764868736,
0.04332897439599037,
-0.039824146777391434,
-0.032291002571582794,
-0.08023864030838013,
-0.06282147765159607,
0.03338820859789848,
0.015430626459419727,
0.023503810167312622,
-0.08413523435592651,
-0.051687005907297134,
-0.029972093179821968,
-0.014377630315721035,
-0.027916960418224335,
0.008180887438356876,
0.056822389364242554,
0.007336915470659733,
-2.920355186170689e-33,
-0.04857749119400978,
0.16968299448490143,
-0.04415682703256607,
-0.005970353726297617,
-0.016002321615815163,
-0.07934712618589401,
0.028546471148729324,
0.03635362535715103,
0.05293232575058937,
-0.07942204177379608,
-0.05010819435119629,
-0.024425018578767776,
0.018869182094931602,
-0.011175142601132393,
-0.05554438382387161,
-0.027623672038316727,
-0.00932571105659008,
-0.029042234644293785,
-0.06774149090051651,
-0.01163925789296627,
-0.0011476597283035517,
0.10349512845277786,
-0.020616715773940086,
0.11118149757385254,
-0.031388357281684875,
0.07902240008115768,
-0.013845984824001789,
0.08281297236680984,
-0.006798847112804651,
-0.051432088017463684,
0.004793605767190456,
-0.003004767931997776,
0.0494210422039032,
-0.05217723920941353,
-0.08813206106424332,
0.05992061644792557,
0.022484539076685905,
0.009130707010626793,
-0.06388004124164581,
0.053185898810625076,
0.053895946592092514,
-0.0404767282307148,
-0.0816754475235939,
0.04859362170100212,
-0.011649752035737038,
-0.05381853133440018,
-0.038173552602529526,
0.003059549257159233,
0.10596726089715958,
0.02845185250043869,
-0.04490234702825546,
0.10345280170440674,
-0.1414368897676468,
0.04814179614186287,
-0.055896732956171036,
-0.08563028275966644,
-0.010308918543159962,
-0.008815902285277843,
0.045035604387521744,
0.03868458792567253,
-0.021694045513868332,
-0.07499212771654129,
0.05851348489522934,
-0.05775060877203941,
-0.03379508852958679,
-0.08115711063146591,
-0.08335839211940765,
0.05135585740208626,
0.00741576636210084,
0.12038003653287888,
0.00958301406353712,
0.020847637206315994,
-0.026345642283558846,
0.0268479622900486,
0.02477935515344143,
-0.08905099332332611,
0.07371649891138077,
-0.009689879603683949,
0.02570599690079689,
0.01455073431134224,
0.019506679847836494,
0.025358764454722404,
-0.02162751741707325,
0.011772754602134228,
0.10611706227064133,
0.0009708619909361005,
0.0125465402379632,
-0.006124417297542095,
0.04794032871723175,
0.05308903753757477,
-0.030192798003554344,
0.0010740255238488317,
-0.014427151530981064,
0.03531922027468681,
-0.0478530116379261,
-4.3125726989501345e-8,
-0.028864340856671333,
0.07511956244707108,
0.03165208548307419,
0.0329539030790329,
0.014523541554808617,
-0.03166709840297699,
0.037602297961711884,
0.07519431412220001,
0.01922069676220417,
-0.004917795769870281,
0.0656091645359993,
0.000640505226328969,
-0.07539861649274826,
-0.0819355845451355,
0.02400960959494114,
0.06494118273258209,
0.04015835374593735,
0.010745293460786343,
0.011050958186388016,
-0.028240330517292023,
-0.029752092435956,
-0.019991017878055573,
-0.013049053028225899,
-0.05961493402719498,
0.018418392166495323,
-0.05825081467628479,
0.04433742165565491,
0.021730473265051842,
0.04004335403442383,
0.03663589060306549,
-0.03289969265460968,
0.03250591456890106,
0.05469261854887009,
-0.06697151064872742,
-0.05844377726316452,
0.09976568073034286,
0.10159081965684891,
-0.023474987596273422,
-0.028715653344988823,
-0.033515773713588715,
-0.014552227221429348,
0.0505201481282711,
-0.026844151318073273,
-0.00998061615973711,
0.006795869208872318,
0.016300247982144356,
0.09720786660909653,
-0.026164362207055092,
0.018939383327960968,
0.03290572762489319,
0.03179870545864105,
-0.08349993824958801,
0.044026173651218414,
0.08849409967660904,
0.0875283032655716,
-0.023500630632042885,
0.01877896673977375,
-0.0007214805227704346,
0.06318838894367218,
0.09646439552307129,
-0.011054244823753834,
0.0028683620039373636,
0.005651494022458792,
0.025037793442606926
] |
Hyper-Connections | https://openreview.net/forum?id=9FqARW7dwB | [
"Defa Zhu",
"Hongzhi Huang",
"Zihao Huang",
"Yutao Zeng",
"Yunyao Mao",
"Banggu Wu",
"Qiyang Min",
"Xun Zhou"
] | Poster | We present hyper-connections, a simple yet effective method that can serve as an alternative to residual connections. This approach specifically addresses common drawbacks observed in residual connection variants, such as the seesaw effect between gradient vanishing and representation collapse. Theoretically, hyper-connections allow the network to adjust the strength of connections between features at different depths and dynamically rearrange layers. We conduct experiments focusing on the pre-training of large language models, including dense and sparse models, where hyper-connections show significant performance improvements over residual connections. Additional experiments conducted on vision tasks also demonstrate similar improvements. We anticipate that this method will be broadly applicable and beneficial across a wide range of AI problems. | Network Architecture, Residual Connections, LLMs, Pre-training | null | 681 | 2409.19606 | [
0.0011639099102467299,
-0.08998223394155502,
0.019620690494775772,
0.06812223047018051,
-0.005361063871532679,
0.08650674670934677,
-0.023407284170389175,
-0.07647483795881271,
-0.017172543331980705,
-0.0699080377817154,
-0.038071051239967346,
0.012064647860825062,
-0.02268456295132637,
0.02282404713332653,
-0.040337830781936646,
0.054480936378240585,
0.12931355834007263,
0.048894159495830536,
-0.07935748994350433,
-0.008492699824273586,
-0.011001450940966606,
-0.04667007178068161,
-0.008567006327211857,
-0.0021182317286729813,
0.0901004821062088,
0.025929730385541916,
-0.030638545751571655,
-0.10863181203603745,
0.01907859370112419,
-0.0032480726949870586,
0.012968378141522408,
-0.028760667890310287,
-0.09693188220262527,
0.02988428622484207,
-0.02150052599608898,
0.08377294987440109,
-0.05194041505455971,
0.05813445523381233,
-0.03974005952477455,
-0.02175767719745636,
0.036988791078329086,
-0.02038007415831089,
0.007326969411224127,
0.06047927588224411,
0.07402268797159195,
0.021679764613509178,
-0.02921324595808983,
-0.032065391540527344,
0.04638242721557617,
-0.009571119211614132,
-0.09688983857631683,
-0.017164774239063263,
0.014639005996286869,
0.06365983188152313,
0.001639806549064815,
0.03581373766064644,
0.0359540618956089,
0.04655707627534866,
-0.029591921716928482,
0.09135988354682922,
0.08386952430009842,
-0.0753471851348877,
-0.013643534854054451,
-0.05329523980617523,
-0.016746649518609047,
0.010241813957691193,
0.05786814168095589,
0.009070655331015587,
-0.028151970356702805,
0.05747625231742859,
0.016088878735899925,
0.13016782701015472,
-0.05725042149424553,
0.00033720897044986486,
0.035989660769701004,
0.053380087018013,
0.04599748179316521,
-0.06494695693254471,
0.07382920384407043,
-0.07640455663204193,
0.08314521610736847,
0.05211479216814041,
0.03199666365981102,
-0.022832456976175308,
0.086881622672081,
0.05887956544756889,
0.0038146234583109617,
-0.07836954295635223,
-0.021280765533447266,
-0.034321386367082596,
-0.003934558480978012,
-0.09006371349096298,
0.03613681718707085,
-0.006794567685574293,
0.010965269058942795,
0.014729456044733524,
0.0018268510466441512,
-0.0892169326543808,
-0.03561249002814293,
0.05561070144176483,
0.019683953374624252,
-0.02682495303452015,
-0.0025638111401349306,
-0.0008538339170627296,
0.05059873312711716,
0.03952983021736145,
0.035327792167663574,
0.03912012651562691,
-0.03778509050607681,
-0.04198675602674484,
0.05671097710728645,
-0.04902762919664383,
0.026851937174797058,
0.018864070996642113,
-0.0043967352248728275,
-0.0050820596516132355,
0.0321236215531826,
-0.017343085259199142,
-0.004049843642860651,
0.11395017057657242,
-0.019695766270160675,
0.02954401820898056,
0.030780542641878128,
0.01389512699097395,
0.0498979352414608,
0.003241373458877206,
-0.0517270490527153,
6.553424168517447e-34,
0.02293103002011776,
0.05060992389917374,
0.054361991584300995,
0.006278854329138994,
0.07554302364587784,
-0.05832948535680771,
0.01513889990746975,
-0.0247334074229002,
-0.017047826200723648,
0.007113977335393429,
-0.09429134428501129,
-0.03204503655433655,
-0.08953050523996353,
0.09797213971614838,
-0.06509493291378021,
-0.02031809464097023,
0.025023682042956352,
0.06719204783439636,
-0.025392191484570503,
-0.021818747743964195,
-0.10293116420507431,
-0.07078162580728531,
0.0027489522472023964,
-0.06462790817022324,
-0.026597850024700165,
-0.023382918909192085,
0.04931703954935074,
-0.06274547427892685,
-0.02863151580095291,
0.021692709997296333,
-0.1105152815580368,
-0.005365699529647827,
0.022656967863440514,
0.05184859782457352,
-0.03377905115485191,
0.01344288419932127,
-0.014648274518549442,
0.005069544073194265,
0.02430756390094757,
0.032226257026195526,
-0.002809293568134308,
0.05254340171813965,
-0.015638580545783043,
-0.023437973111867905,
-0.062280915677547455,
-0.03933161497116089,
0.03944490849971771,
0.0995664969086647,
-0.1184016764163971,
-0.030507979914546013,
0.0019395608687773347,
0.049013689160346985,
-0.07784819602966309,
-0.02724790945649147,
-0.04556697979569435,
-0.042958036065101624,
-0.00013013783609494567,
0.08836788684129715,
0.020158318802714348,
0.0005756308091804385,
0.06372425705194473,
0.028040466830134392,
-0.09974443912506104,
0.11951637268066406,
0.05465862900018692,
0.04893718287348747,
0.013624679297208786,
0.036247190088033676,
0.013808797113597393,
-0.014333880506455898,
-0.09067433327436447,
0.00593856954947114,
-0.00480625731870532,
-0.0624929741024971,
-0.026155846193432808,
0.02129361219704151,
0.013180766254663467,
-0.12910552322864532,
0.03065522201359272,
0.07391658425331116,
-0.005223860964179039,
-0.010795810259878635,
0.041391678154468536,
-0.015849078074097633,
-0.05397981405258179,
0.008631017990410328,
0.026255102828145027,
-0.11906957626342773,
0.036216918379068375,
-0.057935792952775955,
-0.05485126003623009,
-0.042148929089307785,
0.03342725709080696,
-0.067588210105896,
0.043729688972234726,
-6.345657378719531e-34,
-0.009908877313137054,
0.06938798725605011,
-0.024230429902672768,
0.005655419081449509,
-0.0011677151778712869,
-0.011571488343179226,
0.043953586369752884,
0.0031199047807604074,
-0.07088489085435867,
-0.025103764608502388,
0.08748999238014221,
-0.013138808310031891,
-0.025767449289560318,
-0.026171069592237473,
0.05087883770465851,
0.005914253182709217,
-0.009628258645534515,
-0.00866164080798626,
0.04797748103737831,
0.050510622560977936,
0.03168431296944618,
0.011924169026315212,
-0.05090028792619705,
0.026727303862571716,
-0.007650628220289946,
-0.04359976574778557,
-0.006147627253085375,
0.029513005167245865,
-0.029535407200455666,
-0.0024883884470909834,
-0.01371421106159687,
0.0007378386217169464,
-0.06523984670639038,
-0.05411910638213158,
-0.06643251329660416,
0.15120944380760193,
0.021247895434498787,
-0.011469878256320953,
-0.00661076745018363,
0.059782400727272034,
0.03358956426382065,
-0.06779912859201431,
-0.01664584130048752,
-0.0011201204033568501,
-0.022741423919796944,
-0.01916966214776039,
-0.11826840788125992,
-0.07417380064725876,
-0.058932963758707047,
0.07991998642683029,
-0.07381066679954529,
0.049676425755023956,
-0.03199845924973488,
0.0296331699937582,
-0.1605498343706131,
-0.006883916910737753,
-0.031309083104133606,
0.021884087473154068,
0.052915558218955994,
-0.01937011256814003,
-0.04750527814030647,
-0.054070889949798584,
-0.01935383677482605,
-0.036302417516708374,
0.004600308369845152,
-0.05826201289892197,
-0.018473152071237564,
-0.03713437169790268,
0.037341002374887466,
-0.0317394845187664,
0.08696651458740234,
0.0011521445121616125,
-0.027845848351716995,
0.05626704916357994,
-0.026590874418616295,
0.03347316384315491,
-0.026201900094747543,
-0.04340474680066109,
0.007971350103616714,
0.019418856129050255,
0.024323677644133568,
0.06181361898779869,
0.03411313146352768,
0.048549652099609375,
0.13129156827926636,
0.14398342370986938,
-0.00418102415278554,
0.06733044236898422,
0.005692641716450453,
0.03735610470175743,
-0.03953701630234718,
-0.05063065513968468,
0.00997376348823309,
-0.03359127417206764,
-0.11065637320280075,
-4.13627425643881e-8,
-0.08542489260435104,
0.03205148130655289,
0.016361109912395477,
-0.010172860696911812,
0.05493471398949623,
-0.038697052747011185,
-0.026469457894563675,
0.12281908839941025,
-0.05555548518896103,
-0.001670729834586382,
-0.012479184195399284,
0.01522334385663271,
-0.006906633265316486,
-0.00761702423915267,
0.038441941142082214,
0.04849676415324211,
-0.006384093780070543,
0.04792780056595802,
-0.02911251410841942,
-0.011344033293426037,
0.06316463649272919,
0.036370594054460526,
0.03563980385661125,
0.07423721253871918,
0.03611607477068901,
-0.04500359669327736,
-0.035808365792036057,
0.07712141424417496,
0.011272636242210865,
0.016745595261454582,
0.00954504031687975,
0.03641452267765999,
-0.01385840680450201,
0.002804163843393326,
0.015226342715322971,
0.12023331969976425,
0.024534625932574272,
0.005803662817925215,
-0.11278234422206879,
0.015775075182318687,
0.00410136254504323,
0.035215865820646286,
0.02906923182308674,
0.005090342368930578,
0.09541809558868408,
0.06677643209695816,
-0.008148649707436562,
-0.12786856293678284,
-0.023399941623210907,
-0.05659673735499382,
-0.012461286969482899,
0.032527245581150055,
-0.020477844402194023,
0.042136821895837784,
0.04663063958287239,
-0.005327511113137007,
0.026766592636704445,
0.045618996024131775,
-0.05724850296974182,
0.07536157965660095,
-0.018370339646935463,
0.1056116595864296,
-0.026889577507972717,
-0.0409567691385746
] |
Efficient Masked AutoEncoder for Video Object Counting and A Large-Scale Benchmark | https://openreview.net/forum?id=sY3anJ8C68 | [
"Bing Cao",
"Quanhao Lu",
"Jiekang Feng",
"Qilong Wang",
"Pengfei Zhu",
"Qinghua Hu"
] | Poster | The dynamic imbalance of the fore-background is a major challenge in video object counting, which is usually caused by the sparsity of target objects. This remains understudied in existing works and often leads to severe under-/over-prediction errors. To tackle this issue in video object counting, we propose a density-embedded Efficient Masked Autoencoder Counting (E-MAC) framework in this paper. To empower the model’s representation ability on density regression, we develop a new Density-Embedded Masked mOdeling (DEMO) method, which first takes the density map as an auxiliary modality to perform multimodal self-representation learning for image and density map. Although DEMO contributes to effective cross-modal regression guidance, it also brings in redundant background information, making it difficult to focus on the foreground regions. To handle this dilemma, we propose an efficient spatial adaptive masking derived from density maps to boost efficiency. Meanwhile, we employ an optical flow-based temporal collaborative fusion strategy to effectively capture the dynamic variations across frames, aligning features to derive multi-frame density residuals. The counting
accuracy of the current frame is boosted by harnessing the information from adjacent frames. In addition, considering that most existing datasets are limited to human-centric scenarios, we propose a large video bird counting dataset, DroneBird, in natural scenarios for migratory bird protection. Extensive experiments on three crowd datasets and our DroneBird validate our superiority against
the counterparts. The code and dataset are available. | Video object counting, masked autoencoder, multimodal self-representation learning | Density-Embedded Masked AutoEncoder for Video Object Counting and A Large-Scale Benchmark | 678 | 2411.13056 | [
0.028591910377144814,
-0.141569584608078,
0.007149768061935902,
0.08947161585092545,
0.09804388135671616,
0.05360131338238716,
0.020215095952153206,
-0.1079954206943512,
0.051325585693120956,
-0.08449790626764297,
-0.04014824703335762,
-0.10684080421924591,
-0.015898313373327255,
0.10368679463863373,
-0.027515633031725883,
-0.04927317053079605,
0.046570584177970886,
0.06480082124471664,
-0.04189910367131233,
-0.07860203087329865,
0.029756644740700722,
-0.07231760770082474,
0.013056651689112186,
0.006075142417103052,
0.08186937123537064,
0.03404752537608147,
-0.011693197302520275,
-0.07198254019021988,
0.015068788081407547,
-0.03546496108174324,
0.06383015215396881,
-0.05211327597498894,
0.030967317521572113,
0.03226215019822121,
-0.012960868887603283,
-0.07325950264930725,
-0.03863031044602394,
0.05795109644532204,
-0.07884705811738968,
-0.03065924160182476,
-0.043831564486026764,
0.016596026718616486,
0.06988022476434708,
0.007327806670218706,
-0.0033263538498431444,
0.06169314682483673,
0.008344568312168121,
-0.05511721596121788,
0.025121089071035385,
-0.022352168336510658,
0.056368786841630936,
0.020060740411281586,
-0.04350845888257027,
0.026351680979132652,
0.023760471493005753,
0.026235492900013924,
0.03268514573574066,
-0.08081275224685669,
0.022125020623207092,
0.031040553003549576,
0.029911521822214127,
-0.0550210066139698,
0.0739441066980362,
-0.005126406438648701,
-0.0038519345689564943,
-0.017347266897559166,
0.03515973687171936,
0.0577603317797184,
-0.014726422727108002,
0.041499629616737366,
-0.014979740604758263,
0.015108891762793064,
0.04598904401063919,
0.004449859727174044,
0.06323710829019547,
0.004485486540943384,
-0.013500970788300037,
-0.07244919240474701,
0.03512699902057648,
-0.08217675983905792,
0.06695397943258286,
0.02218102663755417,
0.00031921392655931413,
-0.010053880512714386,
0.06962820142507553,
0.019806956872344017,
0.018368834629654884,
0.062177594751119614,
-0.1119348555803299,
-0.0014797495678067207,
-0.04369809851050377,
-0.017015855759382248,
-0.0061126649379730225,
0.019098404794931412,
0.03753958269953728,
-0.04250493273139,
-0.007011191453784704,
-0.06549767404794693,
0.0075711412355303764,
0.024992765858769417,
-0.00647498806938529,
-0.028660625219345093,
-0.04673866182565689,
0.029185736551880836,
0.032502491027116776,
0.007374411448836327,
0.06493518501520157,
0.07473302632570267,
0.05369365215301514,
0.022181307896971703,
0.06825128197669983,
-0.0011472617043182254,
0.038641780614852905,
0.03188566491007805,
-0.01599682681262493,
0.0631774440407753,
0.04001074284315109,
0.05109194293618202,
0.05139733850955963,
0.030252110213041306,
-0.037621770054101944,
-0.04598254710435867,
-0.04847577586770058,
-0.017524726688861847,
0.06318933516740799,
-0.0036164370831102133,
-0.022275183349847794,
-1.3429259805171562e-33,
0.0303924772888422,
0.00021426065359264612,
0.019979307428002357,
0.0028822121676057577,
0.03965179622173309,
-0.05872564762830734,
0.03672993183135986,
0.06477832049131393,
-0.012601963244378567,
0.07336165010929108,
-0.017669185996055603,
-0.03559103608131409,
-0.043477971106767654,
0.10777387022972107,
-0.024490613490343094,
-0.012071081437170506,
-0.04067987576127052,
0.034029848873615265,
-0.01789865456521511,
-0.016332823783159256,
-0.015853408724069595,
-0.052672695368528366,
-0.07599279284477234,
0.0015144156059250236,
-0.00844731554389,
0.047906894236803055,
0.03733757883310318,
-0.019206741824746132,
0.02356036938726902,
0.035360198467969894,
-0.10583563148975372,
-0.007512771524488926,
0.02322262153029442,
0.051214996725320816,
0.045444127172231674,
-0.0822153091430664,
-0.031184636056423187,
0.0014041924150660634,
-0.01290727499872446,
-0.024047354236245155,
0.01030734833329916,
0.010128523223102093,
-0.099666528403759,
-0.0577857531607151,
-0.07370468229055405,
-0.053581077605485916,
0.0863298699259758,
0.05080580338835716,
-0.1382557600736618,
0.04529186710715294,
0.13614650070667267,
0.030413735657930374,
-0.03853340819478035,
-0.07395420223474503,
-0.07489736378192902,
-0.039908088743686676,
0.03572099655866623,
-0.001894469722174108,
0.09359918534755707,
-0.011001110076904297,
0.029225585982203484,
0.059799645096063614,
-0.07394047826528549,
-0.004365180619060993,
0.07477612048387527,
0.00884966365993023,
0.05587233230471611,
0.049498748034238815,
-0.0034836186096072197,
0.01906415820121765,
-0.006957998499274254,
0.07567113637924194,
-0.03773197531700134,
-0.03387865051627159,
0.027881190180778503,
0.06880763173103333,
0.027009017765522003,
-0.007702897302806377,
-0.061597615480422974,
0.07028927654027939,
-0.019368799403309822,
0.042700573801994324,
0.024805938825011253,
-0.033805325627326965,
-0.03067690320312977,
0.02229687198996544,
0.04418998956680298,
0.0061670332215726376,
0.014962857589125633,
-0.01772734709084034,
0.03330986201763153,
-0.12042272835969925,
0.013892523013055325,
-0.052537523210048676,
-0.03206713870167732,
1.7284114449475746e-33,
0.024317773059010506,
0.055442653596401215,
-0.035392533987760544,
0.01806926541030407,
0.04191756248474121,
-0.09641215950250626,
-0.009013992734253407,
0.015496712177991867,
0.05462103709578514,
-0.11128029227256775,
-0.04823531582951546,
-0.012282822281122208,
0.023291325196623802,
0.008739754557609558,
0.027660571038722992,
-0.04866016283631325,
-0.10536560416221619,
0.021647773683071136,
-0.12209991365671158,
0.005800615064799786,
-0.013209803961217403,
0.041341785341501236,
0.006362686399370432,
0.036123182624578476,
-0.1459556370973587,
0.062024492770433426,
-0.016160057857632637,
0.03852255269885063,
0.05712615326046944,
0.016112586483359337,
0.07768440246582031,
0.00856531411409378,
0.03503510728478432,
-0.05458031967282295,
-0.06963270157575607,
0.03239206597208977,
0.014449603855609894,
-0.025220725685358047,
-0.06409326195716858,
0.043842997401952744,
-0.005831606220453978,
-0.04555269703269005,
-0.0652751550078392,
-0.014973397366702557,
-0.026328787207603455,
-0.012320147827267647,
0.012483694590628147,
0.0033712643198668957,
0.1005094051361084,
0.03623325005173683,
-0.08462007343769073,
0.10015310347080231,
-0.10693467408418655,
0.034996699541807175,
-0.02607441321015358,
-0.03294355794787407,
-0.06415034830570221,
0.009649908170104027,
0.11652111262083054,
0.02868730202317238,
-0.0746891126036644,
-0.05685416981577873,
0.0193694606423378,
-0.02933785878121853,
0.006946586072444916,
-0.0283671785145998,
-0.08039657771587372,
-0.03328481689095497,
-0.0024846484884619713,
0.061699047684669495,
0.05173609405755997,
-0.0469064861536026,
-0.013964023441076279,
0.09863421320915222,
-0.009308106265962124,
-0.007201003842055798,
0.06073017790913582,
-0.018856961280107498,
0.032030437141656876,
0.03410542756319046,
-0.06688885390758514,
-0.09671369194984436,
0.02882457710802555,
0.07519780099391937,
0.05286858603358269,
0.10378776490688324,
0.009232687763869762,
-0.050666678696870804,
-0.01094128843396902,
0.009283085353672504,
-0.06407037377357483,
0.023793315514922142,
-0.03590403124690056,
-0.007588841021060944,
-0.0032751949038356543,
-3.888380462058194e-8,
-0.034976255148649216,
0.012305820360779762,
-0.027487309649586678,
-0.010328933596611023,
-0.031217774376273155,
-0.03300212696194649,
0.032653238624334335,
0.05749204009771347,
0.03498772159218788,
-0.04714391753077507,
0.08627491444349289,
-0.07290171086788177,
0.03684023767709732,
-0.020101342350244522,
-0.025390414521098137,
0.014638867229223251,
0.03787831962108612,
0.053352195769548416,
0.0012528825318440795,
0.05299326032400131,
-0.03494850918650627,
0.026019547134637833,
0.03781529888510704,
-0.1009979099035263,
-0.013819740153849125,
-0.03923826292157173,
0.05757031962275505,
0.10420624911785126,
0.07307924330234528,
-0.012144040316343307,
0.033069461584091187,
0.04800866171717644,
-0.007530580274760723,
-0.027155639603734016,
-0.03992685303092003,
0.05693550780415535,
0.02845667488873005,
0.01312896516174078,
-0.06310568749904633,
-0.011293093673884869,
0.014598858542740345,
-0.013958374969661236,
-0.03274601325392723,
0.018076477572321892,
-0.02363048493862152,
0.09786849468946457,
0.03880148380994797,
-0.12084664404392242,
-0.09222396463155746,
-0.044396400451660156,
0.06871271878480911,
-0.04269987717270851,
0.00018603091302793473,
0.12437224388122559,
0.09296774864196777,
-0.010344274342060089,
0.04260530322790146,
-0.03227732703089714,
0.06195726990699768,
0.09374568611383438,
-0.05485263839364052,
0.0006195296300575137,
0.0005705627263523638,
0.011165283620357513
] |
TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting | https://openreview.net/forum?id=rDe9yQQYKt | [
"FENG SHIBO",
"Wanjin Feng",
"Xingyu Gao",
"Peilin Zhao",
"Zhiqi Shen"
] | Poster | Spiking Neural Networks (SNNs) offer a promising, biologically inspired approach for processing spatiotemporal data, particularly for time series forecasting.
However, conventional neuron models like the Leaky Integrate-and-Fire (LIF) struggle to capture long-term dependencies and effectively process multi-scale temporal dynamics.
To overcome these limitations, we introduce the Temporal Segment Leaky Integrate-and-Fire (TS-LIF) model, featuring a novel dual-compartment architecture.
The dendritic and somatic compartments specialize in capturing distinct frequency components, providing functional heterogeneity that enhances the neuron's ability to process both low- and high-frequency information.
Furthermore, the newly introduced direct somatic current injection reduces information loss during intra-neuronal transmission, while dendritic spike generation improves multi-scale information extraction.
We provide a theoretical stability analysis of the TS-LIF model and explain how each compartment contributes to distinct frequency response characteristics.
Experimental results show that TS-LIF outperforms traditional SNNs in time series forecasting, demonstrating better accuracy and robustness, even with missing data.
TS-LIF advances the application of SNNs in time-series forecasting, providing a biologically inspired approach that captures complex temporal dynamics and offers potential for practical implementation in diverse forecasting scenarios. | spiking neural network, time series forecasting, Application | We proposed a Temporal Segment Spiking Neuron Network (TS-LIF) for multivariate time series forecasting, supported by stability analysis and frequency response analysis to demonstrate its effectiveness and efficiency. | 677 | null | [
-0.05310268700122833,
-0.12011241167783737,
0.08392852544784546,
0.049160510301589966,
-0.011935388669371605,
0.048379216343164444,
-0.009791537187993526,
-0.028189344331622124,
0.05359594523906708,
-0.02873116359114647,
-0.06454034894704819,
-0.026989484205842018,
-0.03727279603481293,
0.08522959053516388,
-0.00489472970366478,
-0.048019930720329285,
-0.018404526636004448,
0.009895633906126022,
-0.0198329109698534,
-0.06904789060354233,
0.09711505472660065,
0.040727727115154266,
-0.061896175146102905,
0.06398871541023254,
-0.0066615925170481205,
0.008030577562749386,
-0.0049719675444066525,
-0.04748962074518204,
-0.013459408655762672,
-0.027966346591711044,
0.05946691706776619,
-0.006907389033585787,
-0.0961809754371643,
0.012413698248565197,
-0.0705651268362999,
-0.029997030273079872,
-0.10032156109809875,
0.039101433008909225,
0.039899539202451706,
0.0070191603153944016,
0.083904929459095,
0.008622960187494755,
0.009417015127837658,
0.027520306408405304,
0.006073744501918554,
0.03881536424160004,
0.022821983322501183,
-0.0892668217420578,
-0.042216308414936066,
0.016182836145162582,
0.04264722391963005,
0.040434401482343674,
0.08092007040977478,
0.08744946122169495,
0.014781215228140354,
0.07948675006628036,
-0.025488048791885376,
0.040623798966407776,
-0.012686544097959995,
0.03172982856631279,
-0.07632412016391754,
-0.030378982424736023,
0.02002141624689102,
-0.06726757436990738,
0.005384291987866163,
0.03591936454176903,
0.029154445976018906,
0.035847581923007965,
0.028874512761831284,
-0.00878230668604374,
0.0018149444367736578,
0.07366668432950974,
-0.06782681494951248,
-0.08845876902341843,
0.023883219808340073,
0.004459237214177847,
0.04514557123184204,
0.029370661824941635,
0.03691558167338371,
-0.04305015876889229,
0.06414633244276047,
0.050741616636514664,
0.013508173637092113,
-0.029285185039043427,
0.1389947086572647,
0.052842069417238235,
0.03538715839385986,
0.028687139973044395,
-0.05463733151555061,
-0.022340141236782074,
-0.008652414195239544,
-0.03082488290965557,
0.053038619458675385,
-0.0005829592701047659,
-0.0035915642511099577,
-0.03670912981033325,
0.039177801460027695,
0.021394630894064903,
-0.007462298963218927,
-0.0024318236391991377,
0.0031071710400283337,
0.024660954251885414,
-0.007861047051846981,
0.015238607302308083,
0.08356563001871109,
-0.06432930380105972,
0.05896089971065521,
0.0017768506659194827,
0.01290292851626873,
0.016485219821333885,
0.050509337335824966,
0.07491210103034973,
0.07040052115917206,
0.04735845327377319,
0.04537399113178253,
-0.008055640384554863,
-0.04795901104807854,
0.049095749855041504,
0.06309214234352112,
0.0652170479297638,
0.005604194942861795,
0.01680452562868595,
-0.0643518716096878,
0.037784475833177567,
-0.0019120393553748727,
-0.0017272704280912876,
-0.06540808826684952,
1.0239330887249698e-33,
-0.09415967017412186,
-0.0020033768378198147,
-0.015275297686457634,
0.0011123056756332517,
-0.011654756963253021,
-0.09826566278934479,
-0.02134721167385578,
-0.014887749217450619,
0.056379832327365875,
0.0633205994963646,
-0.11295516043901443,
0.06765732914209366,
-0.07028687000274658,
0.12799394130706787,
0.03188927099108696,
-0.01853429153561592,
-0.03391953185200691,
0.0027910801582038403,
0.04333061724901199,
-0.08676032721996307,
0.03849615901708603,
-0.002018839353695512,
-0.006170451175421476,
-0.01426607370376587,
0.0015872393269091845,
0.0025126058608293533,
0.05751423537731171,
-0.029969312250614166,
-0.03168559819459915,
0.017521990463137627,
-0.08918099850416183,
0.04367528855800629,
-0.03287515416741371,
-0.05995096266269684,
0.09457521140575409,
-0.06552063673734665,
0.03415074199438095,
-0.021644657477736473,
-0.0043619731441140175,
-0.05575583130121231,
0.010790644213557243,
0.02286870777606964,
-0.07535616308450699,
-0.06642772257328033,
-0.07979127019643784,
-0.07548393309116364,
0.029840264469385147,
-0.00012010962382191792,
0.015317803248763084,
-0.11363890022039413,
0.009559266269207,
-0.028223223984241486,
0.04333467409014702,
-0.037841878831386566,
0.02782447449862957,
0.02668537013232708,
0.04008370265364647,
-0.006788891274482012,
0.06338856369256973,
0.1064850464463234,
-0.03920978307723999,
0.0069822147488594055,
0.030630989000201225,
0.030231667682528496,
0.06007848680019379,
0.050304438918828964,
-0.033958084881305695,
0.02831222116947174,
0.08136232197284698,
-0.08912380784749985,
0.06677819788455963,
0.04848584905266762,
0.03090359829366207,
-0.04386528953909874,
-0.04403868317604065,
0.03228508681058884,
0.05020787566900253,
0.008300219662487507,
-0.07383058220148087,
0.034339964389801025,
0.014608790166676044,
-0.14147472381591797,
-0.06648663431406021,
0.07918302714824677,
0.029942313209176064,
-0.04310818016529083,
0.09873591363430023,
0.018266797065734863,
-0.09515368193387985,
-0.04701772332191467,
-0.0760631114244461,
-0.04029976576566696,
0.09006116539239883,
-0.054502490907907486,
-0.03958556428551674,
-2.000226146475997e-33,
-0.039141811430454254,
0.014301293529570103,
-0.04084152728319168,
-0.03510715067386627,
0.022727733477950096,
-0.0041004354134202,
-0.005278764758259058,
0.04276330769062042,
0.013522175140678883,
0.04788218066096306,
-0.03157215192914009,
0.009366150945425034,
-0.021244050934910774,
-0.020963624119758606,
0.06738937646150589,
-0.04515884071588516,
-0.05148765444755554,
0.022636815905570984,
0.03856485337018967,
-0.05115561932325363,
0.020497048273682594,
0.021973205730319023,
-0.13093291223049164,
0.027398783713579178,
-0.030187804251909256,
0.03857886791229248,
-0.056849271059036255,
0.0912870541214943,
-0.04762331023812294,
0.03331257030367851,
-0.09512938559055328,
0.002316896105185151,
0.015559983439743519,
-0.022264115512371063,
0.0032733692787587643,
0.05929664522409439,
0.035977188497781754,
-0.05827470123767853,
0.00374883646145463,
-0.05500483140349388,
0.02420252189040184,
0.05889733135700226,
0.014064494520425797,
0.03440907225012779,
0.04310702905058861,
-0.05555634945631027,
-0.11207303404808044,
0.04691120237112045,
0.026170114055275917,
0.11799504607915878,
-0.011019480414688587,
-0.004518761765211821,
-0.09791548550128937,
-0.12237373739480972,
-0.09001060575246811,
-0.06193959340453148,
0.017300372943282127,
-0.05818759277462959,
0.0012985088396817446,
-0.003162678796797991,
-0.017969535663723946,
-0.0631527304649353,
-0.019341612234711647,
-0.08888968825340271,
-0.02441014163196087,
-0.03440496325492859,
0.010656887665390968,
-0.05712438374757767,
0.08537434786558151,
-0.04605793580412865,
0.08066926896572113,
0.027021726593375206,
0.10547351837158203,
-0.0013078039046376944,
-0.035771384835243225,
-0.042534396052360535,
-0.056849848479032516,
-0.00940285436809063,
0.015591052360832691,
0.013102111406624317,
-0.06893313676118851,
-0.01162280235439539,
-0.012817609123885632,
0.053837887942790985,
0.032441359013319016,
0.0751197338104248,
0.052527766674757004,
0.02627423405647278,
-0.010931040160357952,
-0.08903716504573822,
-0.007261872757226229,
-0.018801141530275345,
-0.011904576793313026,
0.0923362746834755,
-0.023538103327155113,
-4.381276141884882e-8,
0.05714261904358864,
0.03798462823033333,
0.019466601312160492,
0.058743640780448914,
-0.0017612105002626777,
-0.06849324703216553,
0.0342622734606266,
-0.040002886205911636,
-0.03348647430539131,
-0.06406045705080032,
0.11804011464118958,
-0.06055176630616188,
-0.02768440544605255,
0.010932308621704578,
0.03296910971403122,
0.0439131036400795,
0.017076615244150162,
0.02774522267282009,
-0.010497740469872952,
-0.09153400361537933,
0.02549975924193859,
0.019770296290516853,
0.009324117563664913,
-0.011903303675353527,
0.0340740904211998,
0.0046987030655145645,
-0.0793493241071701,
0.0909346416592598,
0.06909656524658203,
-0.05732956528663635,
0.07420661300420761,
-0.012453469447791576,
0.019792398437857628,
0.04781630262732506,
-0.04134127497673035,
0.01707354187965393,
0.007949302904307842,
0.031730443239212036,
0.010582382790744305,
0.02480032667517662,
-0.06340794265270233,
0.013611829839646816,
0.04273808002471924,
0.01764713227748871,
-0.1301814317703247,
-0.050939034670591354,
0.05451911687850952,
-0.03293589502573013,
0.05182613059878349,
0.06511208415031433,
-0.014679056592285633,
0.018398305401206017,
-0.04148140922188759,
0.03230147436261177,
0.06904691457748413,
0.030427830293774605,
-0.01882307231426239,
-0.04459017887711525,
0.03443688154220581,
-0.007557089906185865,
0.011138206347823143,
0.030165912583470345,
-0.11106576025485992,
0.0011558027472347021
] |
ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time | https://openreview.net/forum?id=QoDDNkx4fP | [
"Yi Ding",
"Bolian Li",
"Ruqi Zhang"
] | Poster | Vision Language Models (VLMs) have become essential backbones for multi-modal intelligence, yet significant safety challenges limit their real-world application. While textual inputs can often be effectively safeguarded, adversarial visual inputs can often easily bypass VLM defense mechanisms. Existing defense methods are either resource-intensive, requiring substantial data and compute, or fail to simultaneously ensure safety and usefulness in responses. To address these limitations, we propose a novel two-phase inference-time alignment framework, **E**valuating **T**hen **A**ligning (ETA): i) Evaluating input visual contents and output responses to establish a robust safety awareness in multimodal settings, and ii) Aligning unsafe behaviors at both shallow and deep levels by conditioning the VLMs' generative distribution with an interference prefix and performing sentence-level best-of-$N$ to search the most harmless and helpful generation paths. Extensive experiments show that ETA outperforms baseline methods in terms of harmlessness, helpfulness, and efficiency, reducing the unsafe rate by 87.5\% in cross-modality attacks and achieving 96.6\% win-ties in GPT-4 helpfulness evaluation. | VLMs, safety alignment, inference time | We propose a robust inference time alignment framework, ETA, to safeguard VLMs and enhance both safety and usefulness. | 675 | 2410.06625 | [
-0.03246081620454788,
-0.10139118134975433,
0.006559071131050587,
-0.005459737963974476,
0.12894895672798157,
0.06994129717350006,
-0.03476741537451744,
-0.016094224527478218,
0.05630769208073616,
-0.03909265249967575,
0.002409935463219881,
-0.11630281060934067,
0.022140176966786385,
0.017537379637360573,
-0.059132643043994904,
-0.012093991041183472,
0.05686550587415695,
0.0900174155831337,
-0.12867911159992218,
-0.01680142618715763,
0.09569346159696579,
0.06162165477871895,
0.02145109325647354,
0.005796221550554037,
-0.007317286916077137,
0.04199477285146713,
-0.018639517948031425,
-0.05329037457704544,
0.029623782262206078,
-0.10795338451862335,
-0.027392452582716942,
-0.0009464510367251933,
-0.030232800170779228,
0.04798899218440056,
-0.07701393216848373,
0.024502983316779137,
-0.030489951372146606,
-0.03164420649409294,
0.04911300539970398,
-0.08762145787477493,
-0.053499262779951096,
-0.04489464312791824,
-0.008271447382867336,
-0.06133934110403061,
0.04387460649013519,
-0.004610455594956875,
0.037667158991098404,
-0.05472864955663681,
-0.012868293561041355,
-0.022085560485720634,
-0.0790596678853035,
-0.012882279232144356,
0.030867060646414757,
0.019109081476926804,
-0.0274119321256876,
0.06374601274728775,
0.06415453553199768,
0.05378730967640877,
0.011404487304389477,
0.025595562532544136,
-0.03668198734521866,
-0.049162641167640686,
-0.032604627311229706,
0.0152932433411479,
-0.029242100194096565,
-0.015606132335960865,
0.015535691753029823,
-0.016044054180383682,
-0.03478094935417175,
0.018328340724110603,
-0.01877475157380104,
0.0821562185883522,
-0.019395295530557632,
-0.00995875708758831,
0.03734583780169487,
0.14001917839050293,
0.0668579712510109,
-0.09413184970617294,
0.07392982393503189,
-0.13184379041194916,
0.09505314379930496,
0.01229317020624876,
0.07594603300094604,
0.03002765029668808,
0.130327507853508,
0.0265712458640337,
-0.043695542961359024,
0.0644446462392807,
0.09072436392307281,
-0.011725102551281452,
-0.09057268500328064,
-0.10337822884321213,
0.083138607442379,
0.05852247029542923,
0.1060752123594284,
0.015028043650090694,
-0.004637337289750576,
-0.059165891259908676,
-0.020998449996113777,
0.04025636985898018,
-0.013216043822467327,
0.05731988698244095,
0.010992749594151974,
-0.06932887434959412,
0.061141833662986755,
-0.03380507975816727,
0.09365113824605942,
0.018722789362072945,
0.0063176024705171585,
-0.055723804980516434,
0.06866402924060822,
0.022760573774576187,
0.049584247171878815,
0.005588021129369736,
0.012964311987161636,
-0.024242296814918518,
-0.045385271310806274,
-0.02609531581401825,
0.028186209499835968,
-0.019109586253762245,
0.021627960726618767,
0.017012901604175568,
0.020925048738718033,
0.025306379422545433,
0.060221005231142044,
-0.0676022320985794,
-0.003128054551780224,
1.9742889095428513e-34,
0.052853476256132126,
0.07444914430379868,
0.003264915430918336,
0.02010858617722988,
0.08229563385248184,
-0.008705331943929195,
-0.08945237845182419,
0.006250417325645685,
-0.03006131201982498,
0.022591302171349525,
-0.04350282624363899,
-0.060941729694604874,
-0.0782439261674881,
0.08521432429552078,
0.03167865052819252,
0.04063988849520683,
-0.008776246570050716,
0.06010442227125168,
-0.09235156327486038,
0.0387163907289505,
0.0427701361477375,
-0.04721280187368393,
-0.04905512556433678,
-0.03371445834636688,
0.008616753853857517,
0.01356419362127781,
0.06780257076025009,
-0.0687192901968956,
-0.05145615339279175,
0.031430937349796295,
-0.14956773817539215,
0.009377611801028252,
-0.007204460445791483,
0.02825966291129589,
0.1200438067317009,
-0.01677757315337658,
-0.02968010865151882,
-0.03768930211663246,
-0.03054679185152054,
0.007218565791845322,
-0.04579870030283928,
0.07332902401685715,
0.014845379628241062,
-0.017404604703187943,
-0.005548759829252958,
-0.03463652729988098,
-0.03120141662657261,
0.05549376830458641,
-0.10304902493953705,
-0.05218999460339546,
0.007779907435178757,
0.04414403438568115,
0.038668934255838394,
-0.0775313451886177,
-0.08454674482345581,
-0.0034898463636636734,
0.011514446698129177,
0.015325608663260937,
0.01568218134343624,
0.023456482216715813,
-0.027474701404571533,
0.026633957400918007,
-0.04406796023249626,
0.03438442572951317,
-0.012586535885930061,
-0.024151068180799484,
-0.035982295870780945,
-0.012922246940433979,
0.05454395338892937,
0.023096412420272827,
-0.024716518819332123,
0.001979454653337598,
-0.03294311463832855,
-0.05786001682281494,
-0.025290608406066895,
-0.01404703501611948,
0.09297975897789001,
-0.0316525362432003,
0.08398400992155075,
0.03883446380496025,
0.0234838854521513,
-0.030726874247193336,
0.033747024834156036,
-0.023454975336790085,
-0.046552158892154694,
-0.05504419654607773,
0.04582887142896652,
-0.030101828277111053,
-0.042884040623903275,
-0.013786899857223034,
0.006440210621803999,
-0.025899121537804604,
0.02899930439889431,
-0.044703032821416855,
-0.026232169941067696,
-7.122212364924254e-34,
0.04066990688443184,
0.03771422803401947,
-0.01870516687631607,
0.0512799508869648,
-0.071471206843853,
-0.04745334014296532,
0.03612164780497551,
0.04713395610451698,
0.03487365320324898,
-0.07855121046304703,
0.003683623392134905,
-0.015407491475343704,
-0.025005755946040154,
-0.031141839921474457,
0.08550478518009186,
-0.03724489361047745,
0.021206725388765335,
0.0549795888364315,
0.010976250283420086,
0.006319065578281879,
0.0666278824210167,
0.02960154414176941,
-0.1231217086315155,
-0.008659807965159416,
-0.008320444263517857,
0.059272654354572296,
-0.06507015228271484,
0.01691584102809429,
-0.04064517840743065,
0.05550095811486244,
-0.03771628811955452,
0.04007386416196823,
0.017990300431847572,
-0.0366988480091095,
-0.02481548860669136,
0.05686033144593239,
0.06924423575401306,
-0.06690201908349991,
0.0077517363242805,
0.0581485815346241,
0.05711403116583824,
0.01190602034330368,
-0.08884606510400772,
-0.04651753604412079,
-0.04734519496560097,
-0.017052466049790382,
-0.03143228590488434,
-0.019475344568490982,
-0.006517097819596529,
0.0003996338928118348,
0.004907610826194286,
0.0028809288050979376,
-0.09160814434289932,
0.00278048450127244,
-0.04329218342900276,
-0.09901399910449982,
-0.01701653003692627,
-0.01317284069955349,
0.043058715760707855,
0.011115513741970062,
-0.03342505544424057,
0.02900431863963604,
-0.03502817824482918,
-0.03626847267150879,
0.0029454289469867945,
0.042709916830062866,
-0.04469349980354309,
-0.01751350797712803,
0.07291269302368164,
-0.10926633328199387,
0.11271658539772034,
-0.03568722680211067,
-0.014537526294589043,
0.04509449750185013,
-0.04072411358356476,
-0.09305283427238464,
0.03877544775605202,
-0.027101892977952957,
0.019353443756699562,
-0.003649630583822727,
0.06913361698389053,
-0.07591181248426437,
0.04628681018948555,
0.08945691585540771,
0.0069114151410758495,
0.06636582314968109,
0.06183893606066704,
0.04063818231225014,
0.013497893698513508,
0.02593785710632801,
-0.0960654467344284,
0.03315433859825134,
-0.029475480318069458,
0.10897893458604813,
-0.022813424468040466,
-4.527040431412388e-8,
-0.034185755997896194,
0.06101192906498909,
-0.006650768686085939,
0.05962834134697914,
-0.04221866652369499,
-0.043424855917692184,
-0.06254658848047256,
0.03440646827220917,
0.03516889736056328,
-0.03637256845831871,
0.0351083017885685,
0.003791545983403921,
-0.019536403939127922,
-0.042656198143959045,
-0.0033352049067616463,
0.06223351135849953,
-0.024223603308200836,
0.026790333911776543,
-0.076768659055233,
0.02378765121102333,
-0.007447497919201851,
0.05042311176657677,
-0.016745969653129578,
-0.0431971475481987,
-0.013797834515571594,
-0.037556957453489304,
-0.04896856099367142,
0.03606882318854332,
0.09400849044322968,
-0.0001055519242072478,
-0.0140862250700593,
0.047710876911878586,
-0.019232217222452164,
0.009945733472704887,
-0.06139965355396271,
0.1033790186047554,
-0.03077756054699421,
-0.005887079052627087,
0.03481169790029526,
0.03862548619508743,
0.010859981179237366,
0.05063754320144653,
-0.01919681951403618,
0.03708282858133316,
0.05073690786957741,
-0.01971675641834736,
0.050109852105379105,
-0.15856051445007324,
0.013196881860494614,
-0.014672941528260708,
0.014669583179056644,
0.02639174833893776,
0.05636807531118393,
0.12140414118766785,
-0.022129887714982033,
0.0018339934758841991,
0.04723779112100601,
-0.06384032964706421,
0.02410687692463398,
0.06564997136592865,
0.05465064197778702,
0.014403415843844414,
0.016414791345596313,
-0.05874941125512123
] |
Ranking-aware adapter for text-driven image ordering with CLIP | https://openreview.net/forum?id=KbCh7zbw2K | [
"Wei-Hsiang Yu",
"Yen-Yu Lin",
"Ming-Hsuan Yang",
"Yi-Hsuan Tsai"
] | Poster | Recent advances in vision-language models (VLMs) have made significant progress in downstream tasks that require quantitative concepts such as facial age estimation and image quality assessment, enabling VLMs to explore applications like image ranking and retrieval. However, existing studies typically focus on the reasoning based on a single image and heavily depend on text prompting, limiting their ability to learn comprehensive understanding from multiple images. To address this, we propose an effective yet efficient approach that reframes the CLIP model into a learning-to-rank task and introduces a lightweight adapter to augment CLIP for text-guided image ranking. Specifically, our approach incorporates learnable prompts to adapt to new instructions for ranking purposes and an auxiliary branch with ranking-aware attention, leveraging text-conditioned visual differences for additional supervision in image ranking. Our ranking-aware adapter consistently outperforms fine-tuned CLIPs on various tasks and achieves competitive results compared to state-of-the-art models designed for specific tasks like facial age estimation and image quality assessment. Overall, our approach primarily focuses on ranking images with a single instruction, which provides a natural and generalized way of learning from visual differences across images, bypassing the need for extensive text prompts tailored to individual tasks. | Vision Language Models, CLIP, Learning-to-Rank | A lightweight CLIP-based ranking adapter with novel relational ranking-aware attention to for text-driven image ordering. | 669 | 2412.06760 | [
-0.0038450821302831173,
-0.06216484680771828,
0.05374035984277725,
0.048921067267656326,
0.07180237770080566,
0.018331436440348625,
-0.032132577151060104,
0.025296056643128395,
-0.002449012128636241,
-0.000025940082196029834,
0.0002808083372656256,
-0.018583396449685097,
0.03496173396706581,
0.09639418125152588,
-0.01652209460735321,
0.01632019318640232,
0.01808108203113079,
0.05231513828039169,
-0.06052674353122711,
-0.04428501054644585,
0.016215220093727112,
-0.02722923830151558,
0.11331075429916382,
-0.03330202400684357,
0.03857193887233734,
-0.0021594539284706116,
-0.061606068164110184,
-0.011388740502297878,
0.018645016476511955,
-0.03872609883546829,
-0.013368777930736542,
-0.04493439570069313,
0.04638227820396423,
0.04102008789777756,
-0.07306532561779022,
0.053548768162727356,
-0.05089165270328522,
0.0394844226539135,
-0.02601017989218235,
-0.03041565790772438,
-0.03868195042014122,
-0.014164901338517666,
0.016860561445355415,
-0.014313438907265663,
0.03779369592666626,
0.0031631062738597393,
-0.015021147206425667,
-0.07340002804994583,
0.009580866433680058,
0.027200378477573395,
-0.07223061472177505,
-0.07104964554309845,
-0.0272502563893795,
0.04624566808342934,
-0.08422435820102692,
0.09100665152072906,
0.02189476601779461,
-0.03841756656765938,
0.00025129117420874536,
0.012536739930510521,
-0.05183020979166031,
-0.07231688499450684,
-0.0068232533521950245,
-0.006261763628572226,
-0.010591013357043266,
-0.05582113936543465,
0.06780213117599487,
-0.05163193494081497,
-0.09762799739837646,
-0.013431371189653873,
-0.027694279327988625,
0.03648136556148529,
0.007135561667382717,
0.06554783135652542,
-0.01998932659626007,
0.04878034070134163,
-0.0024140910245478153,
-0.08326598256826401,
0.010340402834117413,
-0.04337327554821968,
0.000044804099161410704,
-0.03475332260131836,
0.05544237419962883,
0.029737742617726326,
0.12880082428455353,
-0.047919467091560364,
0.05145932734012604,
-0.017494678497314453,
-0.03836710378527641,
0.008243815042078495,
-0.027742065489292145,
-0.044589851051568985,
0.05883532389998436,
0.03043469786643982,
0.07897090911865234,
0.06060663238167763,
-0.0217705350369215,
-0.09626057744026184,
-0.05577531084418297,
0.067142054438591,
-0.025358179584145546,
0.03536820039153099,
-0.05404316261410713,
-0.026918534189462662,
0.055562686175107956,
0.025149399414658546,
0.03897193446755409,
0.008834253065288067,
0.021912410855293274,
-0.13035409152507782,
0.08533324301242828,
0.010734920389950275,
0.01518684346228838,
-0.034488677978515625,
0.01248305942863226,
-0.011108015663921833,
0.03342604637145996,
-0.009089270606637001,
0.07188679277896881,
-0.008339285850524902,
0.041876379400491714,
0.020830262452363968,
-0.05119108408689499,
-0.03159581497311592,
0.0356505811214447,
0.012888546101748943,
0.0683356374502182,
1.267462549272227e-33,
0.10281108319759369,
0.06387564539909363,
0.10427354276180267,
0.0772654339671135,
0.08183220028877258,
0.003719699801877141,
-0.04825397953391075,
0.03066703863441944,
-0.13777077198028564,
-0.05522579699754715,
0.008219331502914429,
-0.023082561790943146,
-0.03479340672492981,
0.09670578688383102,
0.014838973060250282,
-0.02441488578915596,
-0.03909450024366379,
0.06985758244991302,
-0.01613030768930912,
-0.015573415905237198,
-0.008779134601354599,
-0.0449938140809536,
0.003956560976803303,
-0.04249190166592598,
0.03338325396180153,
-0.03859454393386841,
0.05458207428455353,
-0.03830794617533684,
-0.015796059742569923,
-0.035181183367967606,
-0.09016528725624084,
-0.04419950023293495,
0.04572617635130882,
0.011614167131483555,
0.02950426936149597,
-0.06935150921344757,
0.006864381954073906,
-0.014292045496404171,
0.06976520270109177,
0.010007896460592747,
0.03968284651637077,
0.010224412195384502,
-0.026512954384088516,
-0.014444290660321712,
-0.09159297496080399,
-0.026223165914416313,
-0.05218242481350899,
0.1166350245475769,
-0.04954662546515465,
0.013524445705115795,
0.0510721392929554,
0.007189363706856966,
-0.021638808771967888,
-0.020520247519016266,
-0.04556433483958244,
-0.016500720754265785,
0.04976312816143036,
0.026085617020726204,
-0.014426250010728836,
-0.051824599504470825,
0.04092087596654892,
0.06765131652355194,
-0.05178384110331535,
0.07647428661584854,
0.032235436141490936,
-0.009797880426049232,
0.04118414595723152,
0.02408882975578308,
0.026344267651438713,
0.007290787063539028,
-0.06723569333553314,
0.07167106866836548,
-0.036260318011045456,
-0.05027986690402031,
0.034930188208818436,
0.05770900472998619,
0.044163305312395096,
-0.10062186419963837,
0.09716156125068665,
-0.0035662236623466015,
0.0333150178194046,
-0.038625821471214294,
0.023182375356554985,
-0.09212955087423325,
-0.00927999522536993,
-0.030811460688710213,
0.04785523936152458,
-0.09479311108589172,
0.04414299502968788,
-0.06380226463079453,
-0.018899057060480118,
-0.027469871565699577,
-0.018512869253754616,
0.0237539354711771,
0.047515977174043655,
1.1047552130062139e-33,
0.03180544078350067,
0.03418032452464104,
0.022479575127363205,
0.0278018768876791,
0.03799120709300041,
-0.006006691139191389,
0.023480458185076714,
0.0154216717928648,
-0.043962374329566956,
-0.1118781566619873,
-0.04514370486140251,
0.0085719283670187,
-0.015583188273012638,
-0.061382509768009186,
0.03338869288563728,
-0.0002863932750187814,
-0.02705535851418972,
-0.014018356800079346,
-0.11256876587867737,
0.008592966943979263,
-0.028139840811491013,
0.0884900689125061,
-0.021665826439857483,
0.020691918209195137,
-0.12548020482063293,
0.07524976134300232,
0.03895151615142822,
0.062148306518793106,
0.0067675295285880566,
-0.005914923269301653,
0.006487080827355385,
-0.01826155185699463,
0.04010730981826782,
-0.015538213774561882,
-0.06604775041341782,
0.09023649990558624,
0.033188559114933014,
-0.09657929092645645,
-0.07485108077526093,
0.16717462241649628,
0.056689586490392685,
-0.049978598952293396,
-0.08809938281774521,
-0.042231813073158264,
-0.04303169623017311,
-0.04700430855154991,
-0.029019564390182495,
-0.038826536387205124,
-0.06830473244190216,
0.064273402094841,
-0.06861597299575806,
0.05534316599369049,
-0.14084608852863312,
0.05527755245566368,
-0.07278065383434296,
-0.055922310799360275,
-0.10433845967054367,
0.06172975152730942,
0.0683705285191536,
-0.0018375266809016466,
0.03128889948129654,
0.0018062845338135958,
0.021390967071056366,
-0.04926573857665062,
-0.025263147428631783,
0.008165162988007069,
-0.04875270277261734,
-0.025720153003931046,
0.006269232369959354,
0.03395666927099228,
0.032659534364938736,
-0.08821118623018265,
0.0521073155105114,
-0.0449969582259655,
0.03713900223374367,
0.020679332315921783,
0.025102391839027405,
0.05999759957194328,
-0.04695058614015579,
-0.06517330557107925,
0.05503775551915169,
-0.01384829729795456,
-0.007120329886674881,
0.07573357224464417,
0.07535997033119202,
0.11453306674957275,
0.0736289769411087,
0.05515367165207863,
0.05312556400895119,
0.013569519855082035,
-0.08957985788583755,
-0.030124755576252937,
0.011674126610159874,
0.05422383174300194,
-0.057206325232982635,
-3.49743771721478e-8,
-0.05841943621635437,
0.01065086480230093,
-0.07236287742853165,
0.09090743958950043,
-0.008777388371527195,
-0.012646595947444439,
-0.03752457723021507,
0.05554305389523506,
0.011938685551285744,
-0.01864064671099186,
0.0750100165605545,
0.034367457032203674,
0.016324350610375404,
-0.05943389609456062,
-0.015337045304477215,
0.07606831938028336,
-0.011923315934836864,
0.09294641762971878,
0.007747751194983721,
-0.0596545971930027,
-0.011461608111858368,
0.05508866906166077,
0.07234427332878113,
-0.019389402121305466,
-0.030955389142036438,
-0.04393656551837921,
-0.016888396814465523,
0.10615797340869904,
0.05113384872674942,
-0.022314205765724182,
0.06452412903308868,
0.049121029675006866,
0.049722831696271896,
-0.022381776943802834,
0.0668184757232666,
0.02637714333832264,
-0.009384269826114178,
-0.04750056937336922,
-0.01730278693139553,
-0.05571060627698898,
0.035000018775463104,
0.06884680688381195,
-0.02828970178961754,
0.07987692952156067,
0.017252538353204727,
0.011020535603165627,
0.02116641029715538,
-0.09235458821058273,
0.04844311624765396,
0.0003122880298178643,
-0.008820516988635063,
-0.07745236903429031,
0.016227878630161285,
0.06415802240371704,
0.03149590268731117,
-0.005171247757971287,
0.022807583212852478,
0.0374283529818058,
0.03338383138179779,
0.09509874135255814,
0.03184780851006508,
-0.009962203912436962,
-0.003934964537620544,
0.001353501807898283
] |
ReMatching Dynamic Reconstruction Flow | https://openreview.net/forum?id=bwhI6bCGY1 | [
"Sara Oblak",
"Despoina Paschalidou",
"Sanja Fidler",
"Matan Atzmon"
] | Poster | Reconstructing a dynamic scene from image inputs is a fundamental computer
vision task with many downstream applications. Despite recent advancements, existing approaches still struggle to achieve high-quality reconstructions from unseen viewpoints and timestamps. This work introduces the ReMatching framework, designed to improve reconstruction quality by incorporating deformation priors into dynamic reconstruction models. Our approach advocates for velocity-field based priors, for which we suggest a matching procedure that can seamlessly supplement existing dynamic reconstruction pipelines. The framework is highly adaptable and can be applied to various dynamic representations. Moreover, it supports integrating multiple types of model priors and enables combining simpler ones to create more complex classes. Our evaluations on popular benchmarks involving both synthetic and real-world dynamic scenes demonstrate that augmenting current state-of-the-art methods with our approach leads to a clear improvement in reconstruction accuracy. | Dynamic Reconstruction, Flow Modeling, Gaussian Splatting, Novel view synthesis | We introduce the ReMatching framework—a novel method for designing and incorporating deformation priors into dynamic reconstruction models, ensuring fidelity to input data while adhering to the specified priors. | 655 | 2411.00705 | [
-0.053159281611442566,
-0.011629488319158554,
0.09768979996442795,
-0.0012762617552652955,
0.0359010174870491,
-0.0016722811851650476,
-0.08134572952985764,
-0.056060340255498886,
0.07071084529161453,
-0.11030115932226181,
-0.10058718174695969,
-0.04356316104531288,
-0.07325691729784012,
0.0570094957947731,
-0.07207683473825455,
-0.043759170919656754,
0.06672628223896027,
0.14605991542339325,
-0.07760096341371536,
0.009452194906771183,
-0.08319506794214249,
0.010905447416007519,
0.025141796097159386,
0.02066543884575367,
0.08776352554559708,
0.04535163566470146,
0.016935521736741066,
0.03193370997905731,
0.007367108482867479,
-0.03661356866359711,
-0.024266749620437622,
-0.01436684001237154,
-0.09381653368473053,
-0.011339526623487473,
-0.039224397391080856,
0.00398836238309741,
0.03542720153927803,
0.028305526822805405,
-0.05077056214213371,
0.02638595923781395,
-0.024004768580198288,
0.061742182821035385,
0.013418982736766338,
-0.011264595203101635,
0.07921194285154343,
0.06654636561870575,
0.06738661974668503,
-0.05491865798830986,
0.05675796419382095,
0.013211934827268124,
-0.01364736445248127,
0.04311102256178856,
0.0010381739120930433,
0.04994037002325058,
-0.019178034737706184,
0.12276115268468857,
0.07134237140417099,
-0.10019341111183167,
0.03206950053572655,
-0.0069398158229887486,
-0.009544290602207184,
-0.008637920953333378,
0.06220345199108124,
0.06341364979743958,
-0.03835142403841019,
-0.019221046939492226,
-0.015835914760828018,
0.01114946324378252,
0.07764404267072678,
0.020722009241580963,
-0.0654827430844307,
0.05072631314396858,
0.02924252860248089,
-0.0804101824760437,
-0.02017272263765335,
0.012366038747131824,
0.05069303140044212,
-0.0037008312065154314,
0.03089204803109169,
-0.05534656345844269,
0.05917426943778992,
-0.033248674124479294,
-0.04986363649368286,
-0.008026382885873318,
-0.02578052692115307,
0.051923591643571854,
-0.05534930154681206,
0.06752793490886688,
0.05026213452219963,
0.06690610945224762,
-0.09519153088331223,
-0.05885878577828407,
-0.020462729036808014,
-0.038142137229442596,
0.04716774448752403,
-0.0373491570353508,
0.025211771950125694,
-0.07526833564043045,
0.10723438113927841,
0.01604543626308441,
-0.04334050044417381,
-0.027767516672611237,
0.02512444369494915,
-0.010229065082967281,
0.07987751066684723,
-0.01692245341837406,
0.04371834546327591,
0.10885005444288254,
0.0016573817702010274,
-0.0009429471683688462,
0.09603554010391235,
-0.06878262758255005,
0.06263332068920135,
0.016823947429656982,
0.03541354089975357,
-0.005085265263915062,
-0.02348160929977894,
-0.02565954625606537,
-0.052338261157274246,
0.013501058332622051,
0.032813869416713715,
-0.023928288370370865,
-0.042329419404268265,
-0.004883007146418095,
0.03459496423602104,
-0.03758492320775986,
0.021422937512397766,
7.177427079456804e-34,
0.008057228289544582,
0.0745568573474884,
0.09016436338424683,
0.07215216010808945,
0.043010808527469635,
0.02041843719780445,
0.028827697038650513,
-0.031947869807481766,
-0.03422300145030022,
-0.00766051234677434,
-0.09640749543905258,
-0.050911568105220795,
-0.024696387350559235,
0.04423500597476959,
-0.055170778185129166,
-0.06941311806440353,
-0.05462666600942612,
0.05475083366036415,
-0.04056492820382118,
0.14164844155311584,
-0.019969850778579712,
0.026520628482103348,
-0.050049182027578354,
-0.08447173237800598,
-0.021097922697663307,
0.04021349176764488,
0.03678424656391144,
0.07607414573431015,
0.011234337463974953,
-0.027208317071199417,
-0.029228709638118744,
-0.04296500235795975,
-0.01948486641049385,
0.013035545125603676,
0.04130033403635025,
-0.0062865219078958035,
0.03966346010565758,
-0.028222015127539635,
-0.017595523968338966,
-0.025183025747537613,
0.004292769823223352,
0.06436373293399811,
-0.08789562433958054,
-0.02979048527777195,
-0.06483758240938187,
-0.061136484146118164,
0.06544391810894012,
0.07757328450679779,
-0.06776843219995499,
0.021003223955631256,
0.08159130811691284,
0.023746922612190247,
-0.01584322564303875,
-0.056331533938646317,
-0.08874267339706421,
-0.052310120314359665,
-0.0036495958920568228,
-0.020701009780168533,
0.0005227252258919179,
0.10321978479623795,
0.07583700865507126,
0.07291226089000702,
-0.08312653750181198,
-0.014347657561302185,
-0.0359162874519825,
0.008669238537549973,
0.04188194125890732,
0.0024095014669001102,
-0.01718365028500557,
0.015166208148002625,
0.014043375849723816,
-0.03875992074608803,
-0.03204231709241867,
-0.04640814661979675,
0.10625196993350983,
0.040240149945020676,
0.019840914756059647,
0.07247735559940338,
0.0010217915987595916,
0.00783575139939785,
-0.13257825374603271,
0.0038144243881106377,
0.034493882209062576,
0.019009584560990334,
0.04426819458603859,
0.05321779474616051,
0.05705200508236885,
-0.028049547225236893,
-0.020572790876030922,
-0.06811109930276871,
-0.030034860596060753,
-0.07536844909191132,
-0.0366891510784626,
-0.023712441325187683,
0.055773232132196426,
-1.2766410479357374e-33,
0.048273976892232895,
0.04350210353732109,
-0.005794913973659277,
0.044586312025785446,
-0.04433487728238106,
-0.052799440920352936,
-0.08384615927934647,
0.05837669596076012,
0.0166917834430933,
-0.10532952845096588,
-0.013758453540503979,
0.01023050770163536,
-0.03129054233431816,
-0.026954928413033485,
0.008141751401126385,
-0.019672760739922523,
-0.05457298085093498,
-0.10932853072881699,
-0.04245652258396149,
-0.03241606056690216,
0.0770392194390297,
0.015058222226798534,
-0.04973771423101425,
-0.039640676230192184,
-0.11985094845294952,
0.1106477901339531,
-0.015225164592266083,
0.012446293607354164,
0.009770006872713566,
-0.04936477169394493,
-0.013273336924612522,
-0.06921568512916565,
-0.019589120522141457,
0.004884904250502586,
-0.09275418519973755,
0.07165779173374176,
0.020860454067587852,
-0.003777762409299612,
-0.007424991577863693,
0.028854994103312492,
-0.018856646493077278,
-0.013728621415793896,
-0.07793830335140228,
0.05765380710363388,
-0.023712702095508575,
-0.03129635378718376,
0.019055308774113655,
0.0554145947098732,
0.0568738728761673,
0.07367043942213058,
-0.06832214444875717,
0.06573785096406937,
-0.09941640496253967,
0.042843740433454514,
-0.03014499694108963,
-0.011521556414663792,
-0.014067419804632664,
-0.05129396170377731,
0.015316646546125412,
0.011724484153091908,
-0.025846347212791443,
-0.029769040644168854,
-0.03495462238788605,
0.0559123270213604,
0.02238234505057335,
0.012820028699934483,
-0.04440062493085861,
-0.07481749355792999,
-0.07025063782930374,
0.05153357982635498,
0.024293329566717148,
0.004763023462146521,
-0.04150017723441124,
0.011006515473127365,
0.06422166526317596,
-0.07492823153734207,
0.021402278915047646,
0.06163319572806358,
0.07937577366828918,
0.056830015033483505,
0.0054563432931900024,
-0.03937791660428047,
0.11402718722820282,
0.07982906699180603,
0.08192319422960281,
0.08364845812320709,
-0.06396627426147461,
-0.06393164396286011,
0.026346899569034576,
0.0615152046084404,
-0.02728530392050743,
-0.02445734292268753,
-0.015508664771914482,
-0.0033005806617438793,
-0.022245939821004868,
-4.148503762735345e-8,
-0.08608698099851608,
0.07779080420732498,
0.005011122673749924,
0.0010263107251375914,
-0.05895564705133438,
-0.047171659767627716,
0.0014715008437633514,
0.08944404125213623,
0.0166750717908144,
-0.09130217134952545,
0.05013350024819374,
-0.0616886243224144,
0.035715993493795395,
-0.024307047948241234,
-0.06383458524942398,
0.08080947399139404,
-0.027571184560656548,
0.0022097043693065643,
-0.06398037075996399,
0.06799356639385223,
-0.03557126596570015,
-0.0005320655182003975,
-0.03268308565020561,
0.04050326347351074,
0.04415693134069443,
-0.000915298645850271,
-0.03342186287045479,
-0.009541474282741547,
0.0832628533244133,
-0.0017924063140526414,
-0.0341864638030529,
0.03237702324986458,
0.008376158773899078,
-0.05031639337539673,
-0.06098419055342674,
-0.02731660008430481,
0.01056150533258915,
0.022361060604453087,
-0.0299670547246933,
-0.03548058494925499,
-0.0684589222073555,
0.00519432220607996,
0.021440742537379265,
0.009628848172724247,
0.07633057236671448,
0.060340043157339096,
0.007048307918012142,
-0.028609268367290497,
-0.02327614836394787,
0.009156009182333946,
0.028506433591246605,
0.033555544912815094,
0.013791614212095737,
0.034474071115255356,
0.07002954930067062,
-0.002694532973691821,
0.04869862273335457,
-0.03418542072176933,
0.04366390034556389,
0.0522817000746727,
-0.03912029787898064,
0.018412167206406593,
0.03328511863946915,
-0.014722365885972977
] |
Analyzing and Boosting the Power of Fine-Grained Visual Recognition for Multi-modal Large Language Models | https://openreview.net/forum?id=p3NKpom1VL | [
"Hulingxiao He",
"Geng Li",
"Zijun Geng",
"Jinglin Xu",
"Yuxin Peng"
] | Poster | Multi-modal large language models (MLLMs) have shown remarkable abilities in various visual understanding tasks. However, MLLMs still struggle with fine-grained visual recognition (FGVR), which aims to identify subordinate-level categories from images. This can negatively impact more advanced capabilities of MLLMs, such as object-centric visual question answering and reasoning. In our study, we revisit three quintessential capabilities of MLLMs for FGVR, including object information extraction, category knowledge reserve, object-category alignment, and position of the root cause as a misalignment problem. To address this issue, we present Finedefics, an MLLM that enhances the model's FGVR capability by incorporating informative attribute descriptions of objects into the training phase. We employ contrastive learning on object-attribute pairs and attribute-category pairs simultaneously and use examples from similar but incorrect categories as hard negatives, naturally bringing representations of visual objects and category names closer. Extensive evaluations across multiple popular FGVR datasets demonstrate that Finedefics outperforms existing MLLMs of comparable parameter sizes, showcasing its remarkable efficacy. The code is available at [https://github.com/PKU-ICST-MIPL/Finedefics_ICLR2025](https://github.com/PKU-ICST-MIPL/Finedefics_ICLR2025). | Multimodal Large Language Models, Fine-Grained Visual Recognition | null | 649 | 2501.15140 | [
0.00042909514741040766,
-0.1015680655837059,
0.016481023281812668,
0.006924021523445845,
0.0811762809753418,
-0.02075223997235298,
-0.055728327482938766,
0.01589934527873993,
0.021877527236938477,
-0.028381917625665665,
-0.027055539190769196,
-0.10223276168107986,
0.005230983719229698,
0.08779015392065048,
-0.017362607643008232,
-0.012587280012667179,
0.0679994598031044,
0.0838075801730156,
-0.14664563536643982,
0.008281157352030277,
0.04496056213974953,
0.014471777714788914,
0.061057593673467636,
-0.00902697816491127,
0.0049224537797272205,
0.018151715397834778,
-0.014037669636309147,
-0.030252326279878616,
0.018733970820903778,
-0.08565180003643036,
-0.0009897475829347968,
0.02571844682097435,
-0.01559113897383213,
0.10458974540233612,
-0.00965082086622715,
0.11093372851610184,
-0.026241960003972054,
0.011109137907624245,
0.004308633506298065,
-0.06079639494419098,
-0.08633644133806229,
-0.010319496504962444,
0.03973593935370445,
-0.03828931599855423,
0.03902514651417732,
-0.005404769442975521,
0.019081249833106995,
0.0013235030928626657,
0.07162239402532578,
-0.004791781771928072,
-0.08508598804473877,
0.004970868583768606,
-0.0063263652846217155,
0.07426164299249649,
-0.06873868405818939,
0.03364761173725128,
0.053213879466056824,
-0.0374586284160614,
-0.016992582008242607,
0.035955820232629776,
-0.03396459296345711,
0.019191835075616837,
0.014064663089811802,
0.02222900278866291,
-0.04592568427324295,
-0.007834467105567455,
-0.009615973569452763,
-0.01466823648661375,
-0.04635625705122948,
-0.005134149454534054,
-0.01770656183362007,
0.009916920214891434,
0.005185174290090799,
0.024689901620149612,
0.03307192772626877,
0.06939485669136047,
0.013837279751896858,
-0.009809314273297787,
0.04543435201048851,
-0.11278930306434631,
0.03661927208304405,
0.046709492802619934,
0.048882585018873215,
-0.0016048813704401255,
0.13270901143550873,
0.006699268240481615,
-0.07806894928216934,
0.05216219276189804,
-0.021615536883473396,
-0.020759033039212227,
-0.04659513011574745,
-0.06990649551153183,
0.09137517958879471,
0.025893565267324448,
0.11421288549900055,
0.038924165070056915,
0.03190826624631882,
-0.08870959281921387,
-0.05921541526913643,
0.07514690607786179,
-0.02361130341887474,
0.07532736659049988,
0.0037869333755224943,
-0.044799163937568665,
0.04473014920949936,
-0.02276032418012619,
0.02961510233581066,
0.0005445491406135261,
0.04712538793683052,
-0.1422375738620758,
0.06622041761875153,
0.06711242347955704,
-0.011162412352859974,
0.029201282188296318,
-0.01322789117693901,
-0.04641622304916382,
-0.01607724092900753,
-0.0074541401118040085,
0.027424179017543793,
-0.08232312649488449,
0.008477919735014439,
-0.007728687487542629,
-0.020129859447479248,
-0.006198743358254433,
0.0881483182311058,
-0.04003103822469711,
-0.04057125747203827,
1.7852893921510418e-33,
0.08744274079799652,
0.03180130198597908,
0.07830806076526642,
0.03818264603614807,
0.05256810784339905,
-0.03531681373715401,
-0.10311080515384674,
0.008958393707871437,
-0.020722582936286926,
0.10180839151144028,
-0.017180832102894783,
0.0170277152210474,
-0.10378903895616531,
0.062402162700891495,
0.042023371905088425,
-0.02860063500702381,
-0.0495721809566021,
-0.03291281312704086,
-0.10633169859647751,
0.008431728929281235,
-0.04057294502854347,
-0.014264844357967377,
0.03199860826134682,
-0.017711171880364418,
0.0104701928794384,
0.02648565173149109,
0.08894676715135574,
-0.07459602504968643,
-0.07655702531337738,
0.01188734546303749,
-0.08912841975688934,
-0.013785082846879959,
0.040333833545446396,
-0.004924928769469261,
0.007635960355401039,
-0.07135665416717529,
-0.027770737186074257,
-0.07491941750049591,
0.021991606801748276,
0.023183777928352356,
0.010154200717806816,
0.010977564379572868,
-0.09546729177236557,
-0.0034795638639479876,
-0.0263112410902977,
0.01722223125398159,
0.011913665570318699,
0.016211913898587227,
-0.014528829604387283,
0.015300646424293518,
0.06270453333854675,
0.013893049210309982,
-0.023893026635050774,
-0.08041892945766449,
-0.02061004936695099,
0.01630493998527527,
-0.022029701620340347,
0.013372056186199188,
0.03495590016245842,
-0.047577984631061554,
-0.02955211140215397,
0.007943407632410526,
0.015603061765432358,
0.08319688588380814,
0.028622088953852654,
0.0002480351540725678,
-0.012520957738161087,
0.10091236978769302,
0.012897560372948647,
0.07255430519580841,
-0.05317908897995949,
0.04054269939661026,
-0.009465697221457958,
-0.07279612123966217,
0.03387772664427757,
0.028932884335517883,
0.04061603918671608,
-0.08951582759618759,
0.018125096336007118,
0.05110975354909897,
0.03880305588245392,
-0.008353346958756447,
0.04873644933104515,
-0.06971905380487442,
-0.08616232126951218,
-0.01602810248732567,
0.07927178591489792,
-0.0768951028585434,
0.027230611070990562,
-0.050173092633485794,
-0.03657853975892067,
0.02094622515141964,
-0.031438812613487244,
-0.020762909203767776,
0.01022742222994566,
-1.7505649216857668e-33,
0.017495067790150642,
-0.05063984915614128,
-0.05405675992369652,
0.013146999292075634,
-0.0030246865935623646,
-0.06383633613586426,
0.08271154761314392,
0.028958972543478012,
-0.022642487660050392,
-0.04456588253378868,
0.004704286344349384,
0.0031538461335003376,
-0.06987112015485764,
-0.06255830079317093,
0.037068214267492294,
-0.006206008605659008,
0.008960237726569176,
0.06634575128555298,
0.00676317885518074,
0.04251944273710251,
0.03176756203174591,
0.10147034376859665,
-0.086081363260746,
-0.012947030365467072,
-0.05354919657111168,
0.049799706786870956,
-0.08765602856874466,
0.02160092070698738,
-0.07763177901506424,
-0.02085156738758087,
0.027860304340720177,
0.0196482315659523,
0.022293204441666603,
-0.05145483836531639,
-0.07520674169063568,
0.02137530781328678,
0.04488563910126686,
-0.11925028264522552,
-0.0022154301404953003,
0.10611732304096222,
0.0660383477807045,
-0.03018481470644474,
-0.10036661475896835,
-0.05058268830180168,
-0.03827793896198273,
-0.08309847116470337,
-0.03255241736769676,
-0.058950964361429214,
0.03311394155025482,
0.04395461082458496,
-0.03227109834551811,
-0.022791806608438492,
-0.07722470909357071,
0.013891431502997875,
0.032814670354127884,
-0.03277705982327461,
-0.07379094511270523,
-0.002159496769309044,
0.054270196706056595,
0.060567229986190796,
-0.06142864748835564,
0.016155865043401718,
0.03759676590561867,
-0.0394270084798336,
0.022257784381508827,
-0.02942357398569584,
-0.0179513581097126,
-0.06884117424488068,
0.09136473387479782,
-0.02988341450691223,
0.025335246697068214,
-0.04410368204116821,
-0.020787931978702545,
0.030319804325699806,
0.045431461185216904,
-0.0029770194087177515,
0.03035688027739525,
0.012493136338889599,
-0.010581888258457184,
-0.07519702613353729,
0.03861549124121666,
-0.10410969704389572,
0.029627501964569092,
0.120751291513443,
0.07669602334499359,
0.0481218658387661,
0.0215712059289217,
0.0725020170211792,
0.09785813093185425,
0.010972133837640285,
-0.009688414633274078,
-0.011838823556900024,
0.06069101020693779,
0.08207816630601883,
0.02423461526632309,
-4.142867737755296e-8,
-0.10269271582365036,
0.09383571147918701,
-0.07190509140491486,
-0.0017985845915973186,
0.009192976169288158,
-0.010436734184622765,
-0.0919516533613205,
0.07627379149198532,
-0.012296206317842007,
-0.019969690591096878,
0.033089812844991684,
0.029948968440294266,
-0.0700882077217102,
-0.059606555849313736,
0.0009867389453575015,
0.09007314592599869,
-0.04575324431061745,
0.027209637686610222,
0.03188084438443184,
0.10718058794736862,
0.08560901880264282,
-0.024220291525125504,
0.03985656797885895,
0.04526720941066742,
-0.05411043018102646,
-0.05735135078430176,
-0.10082963109016418,
0.07738473266363144,
0.04703065752983093,
-0.018852226436138153,
-0.037897396832704544,
0.04441731423139572,
0.01836545206606388,
-0.06137160584330559,
0.02626163512468338,
0.03858349472284317,
-0.03458065539598465,
-0.013820318505167961,
0.023583104833960533,
0.004331057425588369,
0.07420898973941803,
0.024309882894158363,
0.0015686206752434373,
0.03682051599025726,
0.027139659970998764,
0.03506668284535408,
-0.014344949275255203,
-0.02303207479417324,
-0.012240493670105934,
-0.0018195246811956167,
-0.03551328182220459,
0.026685502380132675,
-0.009018800221383572,
0.11999586224555969,
-0.008487694896757603,
-0.06701774150133133,
0.044178400188684464,
0.0033948784694075584,
0.051743634045124054,
-0.004514235537499189,
0.040852490812540054,
0.021769287064671516,
-0.04877270385622978,
-0.029043084010481834
] |
γ−MoD: Exploring Mixture-of-Depth Adaptation for Multimodal Large Language Models | https://openreview.net/forum?id=q44uq3tc2D | [
"Yaxin Luo",
"Gen Luo",
"Jiayi Ji",
"Yiyi Zhou",
"Xiaoshuai Sun",
"Zhiqiang Shen",
"Rongrong Ji"
] | Poster | Despite the significant progress in multimodal large language models (MLLMs), their high computational cost remains a barrier to real-world deployment. Inspired by the mixture of depths (MoDs) in natural language processing, we aim to address this limitation from the perspective of ``activated tokens''. Our key insight is that if most tokens are redundant for the layer computation, then can be skipped directly via the MoD layer. However, directly converting the dense layers of MLLMs to MoD layers leads to substantial performance degradation. To address this issue, we propose an innovative MoD adaptation strategy for existing MLLMs called $\gamma$-MoD. In $\gamma$-MoD, a novel metric is proposed to guide the deployment of MoDs in the MLLM, namely rank of attention maps (ARank). Through ARank, we can effectively identify which layer is redundant and should be replaced with the MoD layer. Based on ARank, we further propose two novel designs to maximize the computational sparsity of MLLM while maintaining its performance, namely shared vision-language router and masked routing learning. With these designs, more than 90% dense layers of the MLLM can be effectively converted to the MoD ones. To validate our method, we apply it to three popular MLLMs, and conduct extensive experiments on 9 benchmark datasets. Experimental results not only validate the significant efficiency benefit of $\gamma$-MoD to existing MLLMs but also confirm its generalization ability on various MLLMs. For example, with a minor performance drop, i.e., -1.5%, $\gamma$-MoD can reduce the training and inference time of LLaVA-HR by 31.0% and 53.2%, respectively. | mixture of depths; multimodal large language models | null | 648 | 2410.13859 | [
-0.005608569364994764,
-0.17663651704788208,
0.04829385131597519,
0.04086213931441307,
0.14557954668998718,
0.016955377534031868,
-0.012590876780450344,
0.022703733295202255,
0.04026137664914131,
-0.04535982385277748,
0.02714267373085022,
-0.04619358107447624,
0.04661022126674652,
0.010840234346687794,
0.020055202767252922,
0.0061170849949121475,
0.0748492032289505,
0.13815487921237946,
-0.10508427023887634,
-0.061650969088077545,
0.028719624504446983,
0.016092026606202126,
0.04987926781177521,
-0.036643315106630325,
0.03386687487363815,
0.036769479513168335,
-0.07288644462823868,
-0.05123753845691681,
0.00975178275257349,
-0.05107523500919342,
0.02070774883031845,
0.005679589230567217,
-0.03292611241340637,
0.07516855746507645,
-0.06850727647542953,
0.061041075736284256,
-0.052522677928209305,
0.020235974341630936,
-0.002317368518561125,
-0.0964195653796196,
-0.07812360674142838,
0.0311721283942461,
-0.048827607184648514,
0.02677123434841633,
0.022249707952141762,
0.00846839975565672,
-0.007182403467595577,
-0.07090456038713455,
0.02232947200536728,
-0.004858795553445816,
-0.03139382600784302,
-0.031432848423719406,
-0.020757343620061874,
0.08820725232362747,
-0.039059631526470184,
-0.0019685898441821337,
-0.034474004060029984,
-0.04167851060628891,
0.0034688354935497046,
0.006863443646579981,
-0.08403544872999191,
-0.07137570530176163,
0.01522529311478138,
-0.07800625264644623,
0.020159156993031502,
-0.055900201201438904,
0.01621619053184986,
-0.05115954577922821,
-0.021318668499588966,
0.05553114041686058,
-0.04495042562484741,
0.07491325587034225,
-0.01084831915795803,
0.03915323689579964,
0.011023173108696938,
0.054481737315654755,
0.04709624499082565,
-0.01677355170249939,
0.0784819945693016,
-0.10497350990772247,
0.07215984165668488,
0.021792162209749222,
0.06446345150470734,
-0.027020249515771866,
0.19814220070838928,
-0.042386364191770554,
-0.0005348869599401951,
0.01600443571805954,
-0.024778839200735092,
0.016632938757538795,
-0.05450346693396568,
-0.058274611830711365,
0.07320700585842133,
-0.002531249076128006,
-0.02885664626955986,
-0.04117853567004204,
-0.035352155566215515,
-0.05700656771659851,
-0.0097741624340415,
0.09678038954734802,
-0.015441960655152798,
0.022190673276782036,
-0.060362037271261215,
-0.03326857089996338,
0.023405499756336212,
-0.004329826217144728,
0.03449869528412819,
0.07052841782569885,
0.02419584058225155,
-0.04831244796514511,
0.02816738933324814,
0.031645480543375015,
0.035815898329019547,
-0.003951867576688528,
0.0070288521237671375,
-0.04313907027244568,
0.03439363092184067,
0.011745354160666466,
0.03588658943772316,
0.06924288719892502,
-0.0333855077624321,
0.00001044454620569013,
0.024265741929411888,
-0.07215556502342224,
0.008364873006939888,
-0.03675348311662674,
-0.0037673639599233866,
6.7586021156786695e-34,
0.06914481520652771,
0.10593413561582565,
-0.045644428580999374,
0.0637025311589241,
0.050543468445539474,
-0.01675504446029663,
0.03004380874335766,
0.002958140801638365,
-0.013934895396232605,
-0.04161816090345383,
-0.0596923790872097,
-0.02539817988872528,
-0.1130460873246193,
0.09689351171255112,
0.014615559950470924,
-0.0005993882659822702,
0.003156566759571433,
0.016083385795354843,
-0.05007762834429741,
-0.09361111372709274,
-0.02250177413225174,
-0.023735327646136284,
0.043744638562202454,
0.005842727608978748,
0.07210493832826614,
0.02042386122047901,
0.04695606231689453,
-0.11622003465890884,
-0.047425877302885056,
0.04568028822541237,
-0.0742608979344368,
0.004028111696243286,
0.0033291801810264587,
0.05748532712459564,
0.024770285934209824,
-0.015742044895887375,
0.030248194932937622,
-0.03486202657222748,
0.03321636840701103,
-0.0353752076625824,
-0.06867100298404694,
0.026459340006113052,
-0.040416136384010315,
-0.024811211973428726,
-0.10264523327350616,
-0.021088136360049248,
0.028907014057040215,
0.08191429823637009,
-0.04182862862944603,
-0.04440242424607277,
0.0748034194111824,
-0.03518681973218918,
-0.019228747114539146,
-0.029766051098704338,
0.007335892878472805,
-0.10625553876161575,
0.035328976809978485,
0.03095184452831745,
0.08704220503568649,
0.020435407757759094,
0.00431623961776495,
0.030421903356909752,
-0.0031064129434525967,
0.14493660628795624,
0.0861412063241005,
0.007426097523421049,
-0.02302052080631256,
0.04364581033587456,
0.021352481096982956,
0.003065169556066394,
-0.07067491114139557,
0.05959896370768547,
0.041036225855350494,
0.03008907288312912,
-0.07713182270526886,
-0.016267046332359314,
0.010675337165594101,
-0.10712045431137085,
0.0671507865190506,
0.061539966613054276,
0.004098912701010704,
0.02421114593744278,
0.007756818551570177,
-0.09301009774208069,
-0.027219364419579506,
-0.028736405074596405,
0.06939088553190231,
-0.10602754354476929,
0.000418758048908785,
-0.04641009494662285,
0.010412517935037613,
-0.032879430800676346,
0.0006153363501653075,
-0.06052819639444351,
-0.03718350827693939,
-7.210559948938165e-34,
-0.008584471419453621,
0.040380824357271194,
-0.01737084984779358,
0.05447021871805191,
-0.059500906616449356,
-0.05970800668001175,
0.06152929365634918,
0.022282926365733147,
-0.06049363315105438,
-0.059081051498651505,
-0.05682379752397537,
-0.06844928115606308,
0.06090211123228073,
0.01013594213873148,
0.0752856656908989,
-0.015515228733420372,
-0.02904829941689968,
0.04256938397884369,
-0.0329899825155735,
0.05050547048449516,
-0.02831997349858284,
0.061779070645570755,
-0.049772024154663086,
0.026137985289096832,
-0.047092337161302567,
0.034667398780584335,
-0.09797345101833344,
0.06701891869306564,
0.011189394630491734,
-0.06429222226142883,
-0.04576170817017555,
0.04147626459598541,
-0.009830542840063572,
-0.1028074398636818,
-0.029633646830916405,
0.08771254122257233,
0.03227873519062996,
0.0062729003839194775,
-0.03715736046433449,
0.10923201590776443,
0.05205138027667999,
-0.035269975662231445,
-0.061228472739458084,
-0.03918791189789772,
-0.003590112552046776,
0.013199149630963802,
-0.048913151025772095,
-0.09157156199216843,
-0.008133070543408394,
0.024537917226552963,
0.03570849448442459,
0.029016191139817238,
-0.05002624914050102,
0.030741361901164055,
-0.02318533882498741,
-0.048232302069664,
-0.060979392379522324,
-0.05801279842853546,
0.036818087100982666,
-0.032785482704639435,
-0.0050573814660310745,
-0.0446084588766098,
0.010607851669192314,
-0.018946928903460503,
0.027911052107810974,
-0.030582712963223457,
0.024218689650297165,
-0.040353912860155106,
0.03567139059305191,
-0.055535830557346344,
0.04656827822327614,
-0.09348618984222412,
0.03910362347960472,
0.035040199756622314,
0.01280028186738491,
0.0356634296476841,
0.0660473182797432,
0.024347463622689247,
-0.017256546765565872,
-0.04834175854921341,
0.01794450357556343,
-0.00027297742781229317,
0.009735074825584888,
0.028521860018372536,
0.08106569200754166,
0.10431979596614838,
0.07204430550336838,
0.06998956948518753,
0.050986748188734055,
-0.011493115685880184,
-0.013431929983198643,
0.03827071189880371,
0.039689019322395325,
0.04649509862065315,
0.05448410287499428,
-4.3910400648883297e-8,
-0.07136273384094238,
0.06220300495624542,
-0.05781490355730057,
-0.013991388492286205,
0.0689229667186737,
-0.0352124348282814,
-0.04060051590204239,
0.08356548100709915,
0.03471978008747101,
0.009569731540977955,
0.03386617824435234,
0.021823914721608162,
-0.007717064116150141,
-0.046814415603876114,
0.006627538241446018,
0.04732104390859604,
-0.03447655588388443,
0.04359054937958717,
0.003857152070850134,
-0.005687262397259474,
-0.041354771703481674,
0.1007542759180069,
0.06781003624200821,
-0.023811016231775284,
0.013216339983046055,
-0.07915709167718887,
-0.0077129858545959,
0.01366649940609932,
0.017970621585845947,
-0.05419599637389183,
-0.031470317393541336,
0.008346060290932655,
0.008571241050958633,
-0.06858532875776291,
-0.040253881365060806,
0.08810335397720337,
0.037563711404800415,
-0.019550729542970657,
-0.04344315081834793,
0.019440045580267906,
0.07040034979581833,
0.012191825546324253,
-0.048622746020555496,
0.050471626222133636,
0.05906441807746887,
0.013928866945207119,
-0.060735028237104416,
-0.10290781408548355,
-0.014639992266893387,
-0.06818418949842453,
0.024361293762922287,
-0.00040653912583366036,
0.012913708575069904,
0.0454019270837307,
0.0658324658870697,
0.0153138292953372,
0.03030187077820301,
0.027863726019859314,
0.0514669269323349,
0.04385873302817345,
0.01821475848555565,
-0.011756878346204758,
-0.07447675615549088,
-0.05432145297527313
] |
Circuit Representation Learning with Masked Gate Modeling and Verilog-AIG Alignment | https://openreview.net/forum?id=US9k5TXVLZ | [
"Haoyuan WU",
"Haisheng Zheng",
"Yuan Pu",
"Bei Yu"
] | Poster | Understanding the structure and function of circuits is crucial for electronic design automation (EDA).
Circuits can be formulated as And-Inverter graphs (AIGs), enabling efficient implementation of representation learning through graph neural networks (GNNs).
Masked modeling paradigms have been proven effective in graph representation learning.
However, masking augmentation to original circuits will destroy their logical equivalence, which is unsuitable for circuit representation learning.
Moreover, existing masked modeling paradigms often prioritize structural information at the expense of abstract information such as circuit function.
To address these limitations, we introduce MGVGA, a novel constrained masked modeling paradigm incorporating masked gate modeling (MGM) and Verilog-AIG alignment (VGA).
Specifically, MGM preserves logical equivalence by masking gates in the latent space rather than in the original circuits, subsequently reconstructing the attributes of these masked gates.
Meanwhile, large language models (LLMs) have demonstrated an excellent understanding of the Verilog code functionality.
Building upon this capability, VGA performs masking operations on original circuits and reconstructs masked gates under the constraints of equivalent Verilog codes, enabling GNNs to learn circuit functions from LLMs.
We evaluate MGVGA on various logic synthesis tasks for EDA and show the superior performance of MGVGA compared to previous state-of-the-art methods.
Our code is available at https://github.com/wuhy68/MGVGA. | circuit representation learning, masked graph modeling, large language models, multimodal alignment | null | 642 | 2502.12732 | [
-0.06636154651641846,
0.02382780611515045,
0.015168843790888786,
0.01695626974105835,
-0.017443615943193436,
-0.04019096493721008,
-0.062138721346855164,
-0.027111927047371864,
-0.0010728678898885846,
-0.020549770444631577,
0.03027239255607128,
-0.06372720003128052,
0.003522739978507161,
0.010220393538475037,
0.026601260527968407,
0.020024271681904793,
0.0016255102818831801,
0.07375750690698624,
-0.04406216740608215,
-0.12367314845323563,
0.08183165639638901,
-0.018631739541888237,
-0.017135992646217346,
-0.012238451279699802,
0.028918931260704994,
0.0618642196059227,
0.02776576764881611,
0.010187757201492786,
0.019212843850255013,
-0.07730865478515625,
0.013583753257989883,
-0.031735435128211975,
0.02347344532608986,
0.035437632352113724,
-0.017870847135782242,
0.08045335859060287,
-0.07445983588695526,
0.017983904108405113,
0.04520268738269806,
-0.11704857647418976,
-0.08360353857278824,
0.04417289048433304,
0.05288830026984215,
0.013699878938496113,
0.08028016239404678,
-0.02727665938436985,
-0.04637599363923073,
-0.014895634725689888,
0.00617144163697958,
-0.03351167216897011,
0.0034373439848423004,
-0.024471618235111237,
-0.016246382147073746,
-0.019448157399892807,
0.04192311689257622,
0.01957727037370205,
0.004824087489396334,
-0.03100687637925148,
-0.009751091711223125,
0.058240506798028946,
-0.009812366217374802,
-0.044088803231716156,
0.007210735231637955,
0.040234558284282684,
-0.08239121735095978,
0.04104883223772049,
0.028881072998046875,
0.0033235412556678057,
-0.017711125314235687,
0.02278018742799759,
-0.021068410947918892,
-0.005191014613956213,
-0.07939782738685608,
0.02995963767170906,
0.0071323830634355545,
0.06613130867481232,
0.07296017557382584,
0.005700551439076662,
0.043668583035469055,
-0.0564718134701252,
-0.010257997550070286,
0.02157842554152012,
-0.011984563432633877,
0.04556344822049141,
0.06051584333181381,
-0.027481991797685623,
-0.06540045142173767,
0.011111347004771233,
0.04865909740328789,
-0.007414259482175112,
-0.03230762109160423,
-0.06549651175737381,
0.010527608916163445,
0.04227061569690704,
0.04863671585917473,
-0.0001328588550677523,
0.06545066833496094,
-0.035194605588912964,
-0.05066513270139694,
0.0658501535654068,
0.0012523820623755455,
0.07700730860233307,
0.0014844664838165045,
-0.07479611784219742,
0.006626913323998451,
0.022588934749364853,
0.010299383662641048,
0.046918533742427826,
-0.02791946567595005,
-0.16581596434116364,
0.03839389607310295,
0.1227649375796318,
0.03838353231549263,
0.08665567636489868,
0.018942303955554962,
-0.03339546546339989,
0.01757250726222992,
-0.056754276156425476,
0.07227472960948944,
0.05887081101536751,
-0.09376450628042221,
0.03756073862314224,
-0.06339001655578613,
0.05083427205681801,
-0.009407870471477509,
-0.08005707710981369,
-0.11034192889928818,
2.5548041773847974e-33,
0.008432513102889061,
0.05783513933420181,
-0.027682457119226456,
-0.003185151843354106,
0.11585249751806259,
0.08249390870332718,
0.01888890005648136,
0.07315539568662643,
-0.004879195708781481,
0.07949869334697723,
-0.05051662027835846,
-0.09670466184616089,
-0.04271157830953598,
0.05089595541357994,
-0.01654057390987873,
0.005257946904748678,
-0.060782916843891144,
-0.024050848558545113,
-0.05836624279618263,
-0.10059403628110886,
0.012150618247687817,
-0.03372276946902275,
0.024593601003289223,
0.0275583416223526,
0.036075085401535034,
0.045312900096178055,
-0.0027403798885643482,
-0.023863887414336205,
-0.06007455289363861,
-0.0066669113002717495,
-0.04978051036596298,
0.0173309575766325,
0.028857775032520294,
0.006805648561567068,
-0.01014051865786314,
0.007578979711979628,
0.0704837441444397,
-0.09932298213243484,
0.08226766437292099,
-0.04100830480456352,
-0.038793791085481644,
0.05010638386011124,
-0.03909287601709366,
-0.03361046686768532,
-0.059222448617219925,
0.02506883442401886,
0.015433704480528831,
-0.028082096949219704,
0.026801427826285362,
-0.05188224837183952,
-0.010553741827607155,
-0.014040467329323292,
-0.00025496509624645114,
-0.07114206999540329,
0.02483706921339035,
-0.02587093412876129,
0.050398048013448715,
0.08467832207679749,
0.07679083943367004,
0.12898385524749756,
-0.055896274745464325,
0.06169939786195755,
-0.0391240194439888,
0.052100010216236115,
0.0012127319350838661,
0.06175021454691887,
-0.016038527712225914,
-0.02231738343834877,
-0.019123923033475876,
-0.029234504327178,
-0.04051778092980385,
0.030360957607626915,
-0.11196695268154144,
-0.021774807944893837,
-0.008379333652555943,
-0.03569275513291359,
-0.016210442408919334,
-0.07472099363803864,
-0.031273216009140015,
-0.003994329832494259,
-0.02839529700577259,
0.03239034488797188,
-0.04744990915060043,
0.01807507872581482,
-0.028726886957883835,
-0.09743400663137436,
0.08886611461639404,
-0.011102111078798771,
0.05754523724317551,
-0.07209376245737076,
-0.006491075269877911,
-0.06649329513311386,
0.06446576863527298,
0.020298663526773453,
0.04021930322051048,
-4.01048734128817e-33,
-0.07611023634672165,
0.09585759043693542,
0.04315253719687462,
0.013848822563886642,
-0.05065303295850754,
-0.0982263907790184,
0.01794702745974064,
-0.029786765575408936,
-0.010978774167597294,
0.021319346502423286,
-0.009513074532151222,
0.0020045535638928413,
-0.032184455543756485,
-0.054322369396686554,
0.0034585529938340187,
0.01052739005535841,
-0.0697544664144516,
0.01701970212161541,
0.028824038803577423,
0.02333260141313076,
-0.014286668039858341,
0.08946400880813599,
-0.0900249257683754,
-0.014817561022937298,
0.04208755120635033,
0.0444486029446125,
-0.024090228602290154,
0.08256950974464417,
0.08517113327980042,
0.03954584151506424,
0.01823985017836094,
0.04069045931100845,
-0.0028451168909668922,
0.0213016327470541,
-0.0029113658238202333,
0.03156086057424545,
0.04251743480563164,
-0.003244049148634076,
0.05632850527763367,
-0.04398918151855469,
0.00989652331918478,
-0.008793430402874947,
-0.13022662699222565,
0.04254833981394768,
-0.06560744345188141,
0.07279028743505478,
0.03751925006508827,
0.005584636237472296,
-0.04459806904196739,
-0.014739228412508965,
-0.07672850787639618,
0.0045021953992545605,
0.015935100615024567,
-0.057391926646232605,
-0.05332444608211517,
-0.04338366165757179,
0.004059957340359688,
0.026029009371995926,
0.11238699406385422,
-0.05570243299007416,
0.010729656554758549,
-0.09810832887887955,
0.12969286739826202,
-0.10043047368526459,
-0.04685966670513153,
-0.025500865653157234,
0.022467175498604774,
0.059136442840099335,
0.004284524358808994,
0.023661920800805092,
0.03908536583185196,
-0.014948667027056217,
-0.043291542679071426,
-0.006384044885635376,
0.008327069692313671,
-0.044669926166534424,
-0.02377740666270256,
-0.027693813666701317,
-0.038541700690984726,
-0.023284243419766426,
0.001177536672912538,
0.007793570403009653,
-0.054111048579216,
0.06467638909816742,
-0.019585080444812775,
0.035989340394735336,
0.02055690623819828,
0.09013545513153076,
0.046438779681921005,
0.030465353280305862,
-0.09679725021123886,
0.0828515887260437,
-0.02974219061434269,
0.06742832064628601,
-0.07579223066568375,
-3.719067365182127e-8,
-0.07230965793132782,
-0.03838006779551506,
0.030432768166065216,
-0.0018818738171830773,
0.062412407249212265,
-0.05496576800942421,
0.07850176841020584,
0.02907218225300312,
-0.02477499097585678,
-0.05082156881690025,
0.08290454000234604,
-0.004964346997439861,
0.00045542881707660854,
-0.07415735721588135,
-0.018884340301156044,
0.07881458103656769,
0.006528194528073072,
0.01785280928015709,
0.05284636467695236,
0.09085410088300705,
-0.0551723837852478,
-0.01642971858382225,
-0.018500937148928642,
0.05122007057070732,
0.07386518269777298,
-0.14947786927223206,
-0.07520182430744171,
0.08241746574640274,
0.06768001616001129,
0.05439221113920212,
-0.013072202913463116,
0.03055683523416519,
0.16754694283008575,
0.014770773239433765,
0.025365857407450676,
0.07651162892580032,
0.03845951706171036,
-0.047292307019233704,
0.0014373711310327053,
-0.028800202533602715,
-0.036966241896152496,
-0.02821587398648262,
-0.020075755193829536,
0.0035003747325390577,
-0.07474987208843231,
-0.032306693494319916,
0.05684994161128998,
-0.07646601647138596,
0.006635780446231365,
-0.012279345653951168,
-0.016762666404247284,
-0.03271941468119621,
-0.03930133208632469,
-0.027668669819831848,
-0.01339930109679699,
0.022986218333244324,
-0.03543408587574959,
-0.02699848636984825,
0.08209461718797684,
0.05422452837228775,
-0.020715391263365746,
0.018262919038534164,
0.017460353672504425,
-0.06948645412921906
] |
AuroraCap: Efficient, Performant Video Detailed Captioning and a New Benchmark | https://openreview.net/forum?id=tTDUrseRRU | [
"Wenhao Chai",
"Enxin Song",
"Yilun Du",
"Chenlin Meng",
"Vashisht Madhavan",
"Omer Bar-Tal",
"Jenq-Neng Hwang",
"Saining Xie",
"Christopher D Manning"
] | Poster | Video detailed captioning is a key task which aims to generate comprehensive and coherent textual descriptions of video content, benefiting both video understanding and generation. In this paper, we propose AuroraCap, a video captioner based on a large multimodal model. We follow the simplest architecture design without additional parameters for temporal modeling. To address the overhead caused by lengthy video sequences, we implement the token merging strategy, reducing the number of input visual tokens. Surprisingly, we found that this strategy results in little performance loss. AuroraCap shows superior performance on various video and image captioning benchmarks, for example, obtaining a CIDEr of 88.9 on Flickr30k, beating GPT-4V (55.3) and Gemini-1.5 Pro (82.2). However, existing video caption benchmarks only include simple descriptions, consisting of a few dozen words, which limits research in this field. Therefore, we develop VDC, a video detailed captioning benchmark with over one thousand carefully annotated structured captions. In addition, we propose a new LLM-assisted metric VDCscore for bettering evaluation, which adopts a divide-and-conquer strategy to transform long caption evaluation into multiple short question-answer pairs. With the help of human Elo ranking, our experiments show that this benchmark better correlates with human judgments of video detailed captioning quality. | Video Captioning, Benchmark, Multimodel Large Language Model | null | 640 | 2410.03051 | [
0.023117098957300186,
-0.09928473830223083,
-0.0034932782873511314,
-0.02912650629878044,
0.13529428839683533,
-0.010945923626422882,
-0.07060341536998749,
0.020279908552765846,
0.08397755771875381,
-0.0511394739151001,
-0.0482732355594635,
-0.08937329798936844,
0.012824621982872486,
0.08516696840524673,
-0.0074248965829610825,
-0.04959700629115105,
0.06918271631002426,
0.0675247386097908,
-0.06571602821350098,
-0.03413662314414978,
0.012185052037239075,
-0.02611631155014038,
0.07057004421949387,
-0.033788908272981644,
0.04776269197463989,
0.03790681064128876,
-0.026666710153222084,
-0.02480304427444935,
0.008407718501985073,
-0.05117974802851677,
-0.0073519377037882805,
-0.03530019149184227,
-0.03520628809928894,
0.06854203343391418,
-0.08495206385850906,
0.06067509576678276,
-0.043689679354429245,
-0.010058579035103321,
0.0017111703054979444,
-0.09312508255243301,
0.01214538887143135,
0.016821835190057755,
0.031070485711097717,
0.011439800262451172,
0.07472123950719833,
0.061596356332302094,
0.01820272020995617,
-0.09821932017803192,
0.06754112243652344,
0.03894323483109474,
-0.04416436329483986,
-0.019648224115371704,
0.02131309174001217,
0.017550038173794746,
-0.05802959203720093,
0.05128880590200424,
0.045256342738866806,
-0.10307135432958603,
0.05097167193889618,
-0.04425879195332527,
-0.0209854356944561,
-0.07015982270240784,
0.06660536676645279,
0.015557602047920227,
0.0399601012468338,
-0.10292799770832062,
0.1295047402381897,
-0.0007066386751830578,
0.014953779056668282,
-0.026778092607855797,
-0.08175424486398697,
0.044437799602746964,
-0.013245378620922565,
0.09227049350738525,
-0.10990601032972336,
0.07547833770513535,
0.05063923820853233,
-0.05239765718579292,
0.03651925176382065,
-0.0862581804394722,
0.05536985024809837,
-0.00801792647689581,
0.05445230379700661,
0.022660862654447556,
0.037822213023900986,
-0.03153868019580841,
0.031301241368055344,
0.006713992450386286,
-0.04107239469885826,
0.06841377913951874,
-0.09354556351900101,
-0.018442323431372643,
0.09940504282712936,
-0.03521306812763214,
-0.002703194972127676,
0.038557447493076324,
0.011832142248749733,
-0.0811380073428154,
-0.01955278031527996,
0.054511554539203644,
0.03651675581932068,
0.04943619668483734,
-0.023707231506705284,
-0.06144874542951584,
0.05043580010533333,
-0.010102702304720879,
0.07241206616163254,
0.020819980651140213,
-0.018206747248768806,
0.020648878067731857,
0.06018683314323425,
0.026202164590358734,
-0.04572317376732826,
-0.011061448603868484,
0.07367904484272003,
0.06333436816930771,
-0.03229036182165146,
-0.02817150019109249,
0.01737021841108799,
0.0588609017431736,
0.0006199581548571587,
-0.0019146277336403728,
-0.011282441206276417,
-0.0004797495494130999,
-0.03151087462902069,
-0.04411100596189499,
0.05668136477470398,
2.6957051750880956e-34,
0.034175578504800797,
0.020954517647624016,
0.037348512560129166,
0.037091098725795746,
0.07350950688123703,
0.0002947537577711046,
0.004807671997696161,
0.10378070175647736,
-0.12176991999149323,
-0.06260158121585846,
0.01683185063302517,
-0.08100496232509613,
-0.014953142032027245,
0.07693475484848022,
0.009863482788205147,
-0.024894073605537415,
-0.0784817636013031,
0.04487437382340431,
-0.0367533303797245,
0.06759491562843323,
-0.02709924429655075,
0.009684913791716099,
0.04322867840528488,
0.046485867351293564,
0.053365614265203476,
-0.036497052758932114,
0.04819982126355171,
-0.08706830441951752,
-0.004058487247675657,
-0.018485605716705322,
-0.09602291136980057,
-0.053460367023944855,
0.05584169924259186,
0.001844819518737495,
0.10478190332651138,
-0.07133804261684418,
-0.06159540265798569,
-0.027722522616386414,
-0.01706031709909439,
-0.07044914364814758,
-0.045579250901937485,
-0.0102439159527421,
-0.10190223902463913,
-0.08612442016601562,
-0.06887612491846085,
-0.0033506620675325394,
-0.018359804525971413,
0.05807268247008324,
-0.035979121923446655,
-0.02705385908484459,
0.05733826383948326,
0.08922042697668076,
-0.03974153473973274,
-0.03986813500523567,
-0.0011494678910821676,
-0.051271095871925354,
0.037736158818006516,
-0.0009129358222708106,
0.014547870494425297,
0.01802491396665573,
-0.014519299380481243,
0.07487969100475311,
-0.024623505771160126,
0.07019412517547607,
-0.023820383474230766,
0.06193004176020622,
0.021701224148273468,
0.10723473876714706,
0.056654829531908035,
0.026324504986405373,
-0.020592493936419487,
-0.010049271397292614,
0.04586441442370415,
-0.026080455631017685,
0.04787242040038109,
0.029774166643619537,
0.035254064947366714,
-0.05442019924521446,
0.01785334385931492,
0.05724246799945831,
-0.0320475734770298,
0.00004108960638404824,
0.029038766399025917,
-0.020440176129341125,
-0.003055410459637642,
-0.07144540548324585,
0.023674527183175087,
-0.011090941727161407,
-0.026871563866734505,
-0.03175933659076691,
-0.004144679289311171,
0.013010019436478615,
-0.05991546809673309,
-0.006145710591226816,
-0.040509916841983795,
1.5632339175059038e-33,
0.06638963520526886,
0.07460752874612808,
-0.03730107098817825,
0.044891633093357086,
-0.003814397845417261,
-0.05912444368004799,
-0.04344267398118973,
0.00657278997823596,
0.0053920201025903225,
-0.06582563370466232,
-0.04370005801320076,
-0.056805506348609924,
-0.052553776651620865,
-0.02594401128590107,
-0.0003251135640311986,
0.01015244610607624,
0.00360615155659616,
-0.043150629848241806,
-0.0595235712826252,
0.03497428447008133,
0.04459521546959877,
0.08567647635936737,
-0.038650672882795334,
-0.017172690480947495,
-0.037677086889743805,
0.05720139667391777,
-0.013221670873463154,
0.036250170320272446,
0.001302072312682867,
-0.0723288357257843,
0.0722963735461235,
0.05127202719449997,
0.012802514247596264,
-0.004040881991386414,
-0.04925423488020897,
0.014743316918611526,
0.1274765431880951,
-0.03627890348434448,
-0.022262200713157654,
0.06977564841508865,
0.09516777098178864,
-0.05112208053469658,
-0.008167093619704247,
-0.019002897664904594,
0.011100046336650848,
-0.038073353469371796,
-0.1074046641588211,
-0.02291168086230755,
0.017291050404310226,
0.06720936298370361,
-0.0756554827094078,
-0.008002153597772121,
-0.09127433598041534,
0.07119186967611313,
-0.016636693850159645,
-0.1474599838256836,
-0.0939871072769165,
0.0336291640996933,
0.05917716771364212,
-0.03801651671528816,
-0.029591534286737442,
-0.011982740834355354,
-0.003753948723897338,
0.003789982059970498,
0.012519470416009426,
-0.08474303781986237,
-0.00001946573866007384,
-0.05698491260409355,
-0.032867878675460815,
0.033006541430950165,
0.07393348962068558,
-0.054672591388225555,
-0.01656203344464302,
0.02210811898112297,
0.009428516961634159,
0.07676244527101517,
0.09821958094835281,
-0.03238748013973236,
0.009914593771100044,
-0.12987081706523895,
-0.04342265427112579,
0.005254881456494331,
0.04944419115781784,
0.045979224145412445,
0.04828394576907158,
0.09185757488012314,
0.059413932263851166,
0.07652853429317474,
0.019711637869477272,
0.10431751608848572,
-0.06881474703550339,
0.003290668362751603,
0.01889646239578724,
0.05755891278386116,
-0.007514216471463442,
-3.431732409353572e-8,
-0.02196331135928631,
0.04199475795030594,
-0.048067230731248856,
0.030553249642252922,
-0.03735551983118057,
-0.009802606888115406,
-0.008543574251234531,
0.04769575968384743,
0.07570898532867432,
-0.04936232417821884,
0.08275497704744339,
-0.014191903173923492,
-0.03264927491545677,
-0.015299788676202297,
-0.06593156605958939,
0.009596792049705982,
-0.012125394307076931,
0.010721679776906967,
0.025377757847309113,
0.020704608410596848,
-0.03712082654237747,
0.04401403293013573,
-0.00329422065988183,
-0.06314403563737869,
-0.03910910338163376,
0.01408536359667778,
-0.028924597427248955,
0.08959084749221802,
0.043638166040182114,
-0.08709962666034698,
0.020772233605384827,
0.04792370647192001,
0.057297851890325546,
-0.0701342523097992,
0.01998697593808174,
-0.0064527541399002075,
0.02609473653137684,
0.005996199790388346,
0.009469613432884216,
0.0035225667525082827,
0.038168489933013916,
0.04397185519337654,
-0.0037439081352204084,
0.004395745228976011,
0.017897818237543106,
0.026958627626299858,
-0.07188393920660019,
-0.05139674246311188,
0.03164475038647652,
-0.02069058269262314,
-0.04691804200410843,
-0.06899324804544449,
-0.021813897415995598,
0.09030594676733017,
0.06897837668657303,
0.0031607786659151316,
0.01170665118843317,
-0.006810653489083052,
0.07309795171022415,
0.05992148071527481,
0.04514394700527191,
-0.03701300919055939,
-0.02544492483139038,
-0.00373270851559937
] |
Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models | https://openreview.net/forum?id=ZYd5wJSaMs | [
"Shuhong Zheng",
"Zhipeng Bao",
"Ruoyu Zhao",
"Martial Hebert",
"Yu-Xiong Wang"
] | Poster | Beyond high-fidelity image synthesis, diffusion models have recently exhibited promising results in dense visual perception tasks. However, most existing work treats diffusion models as a standalone component for perception tasks, employing them either solely for off-the-shelf data augmentation or as mere feature extractors. In contrast to these isolated and thus sub-optimal efforts, we introduce a integrated, versatile, diffusion-based framework, Diff-2-in-1, that can simultaneously handle both multi-modal data generation and dense visual perception, through a unique exploitation of the diffusion-denoising process. Within this framework, we further enhance discriminative visual perception via multi-modal generation, by utilizing the denoising network to create multi-modal data that mirror the distribution of the original training set. Importantly, Diff-2-in-1 optimizes the utilization of the created diverse and faithful data by leveraging a novel self-improving learning mechanism. Comprehensive experimental evaluations validate the effectiveness of our framework, showcasing consistent performance improvements across various discriminative backbones and high-quality multi-modal data generation characterized by both realism and usefulness. Our project website is available at https://zsh2000.github.io/diff-2-in-1.github.io/. | Diffusion Models, Generation, Dense Perception | null | 635 | null | [
0.00043193259625695646,
-0.19102299213409424,
0.06824412941932678,
0.024685071781277657,
0.07744880765676498,
-0.07747875899076462,
-0.035721275955438614,
-0.07984812557697296,
-0.0011345488019287586,
-0.07619553059339523,
0.005325635429471731,
-0.10206051915884018,
-0.008390427567064762,
0.04854149743914604,
-0.026908310130238533,
0.04605194553732872,
0.0952502191066742,
0.13177089393138885,
-0.1314215213060379,
-0.06225390359759331,
0.015357592143118382,
-0.02216503582894802,
-0.007238344289362431,
0.017518267035484314,
0.07467937469482422,
0.05292871594429016,
0.01306809950619936,
-0.026707634329795837,
0.03815944865345955,
-0.07181262224912643,
0.04741789773106575,
-0.0333271287381649,
-0.08690139651298523,
-0.009651425294578075,
-0.055462297052145004,
0.003516114316880703,
0.011405745521187782,
0.06413877010345459,
0.011606228537857533,
-0.04764679819345474,
0.03206576034426689,
0.09280560910701752,
0.022054294124245644,
0.06458696722984314,
0.017503464594483376,
-0.021444374695420265,
-0.018165893852710724,
-0.03363131731748581,
0.07293182611465454,
0.015325577929615974,
-0.008342900313436985,
0.009167627431452274,
0.017358796671032906,
0.054002851247787476,
-0.00802849605679512,
0.02098941057920456,
0.05513978749513626,
0.06315670907497406,
0.0032857470214366913,
-0.02170870266854763,
-0.013250199146568775,
-0.00816706195473671,
0.03577917814254761,
0.0001863919460447505,
0.003314132569357753,
-0.029358621686697006,
0.07741272449493408,
-0.027481039986014366,
-0.03643343970179558,
-0.005448987241834402,
-0.037596721202135086,
0.01952352561056614,
-0.07496104389429092,
-0.017553044483065605,
-0.006478733383119106,
0.027223549783229828,
0.006806999444961548,
-0.028776416555047035,
0.023278888314962387,
-0.044630974531173706,
0.08210963010787964,
0.04241473600268364,
0.08671547472476959,
0.0007867890526540577,
0.06499350070953369,
0.008054378442466259,
-0.025493603199720383,
-0.001611216808669269,
-0.08015982061624527,
-0.01044177170842886,
-0.09187079221010208,
0.03676370903849602,
-0.04859040677547455,
-0.02910243161022663,
0.008593928068876266,
0.03638245910406113,
0.0030145132914185524,
-0.042170483618974686,
0.01577872410416603,
0.09483224153518677,
-0.01197061687707901,
-0.017037922516465187,
-0.07283181697130203,
0.02575729228556156,
0.032231561839580536,
-0.06286879628896713,
0.09282564371824265,
0.027340753003954887,
-0.03761344030499458,
-0.011227712966501713,
0.06706448644399643,
0.008315181359648705,
0.028330018743872643,
0.046990253031253815,
0.036803971976041794,
-0.02157147414982319,
0.03217505291104317,
0.0428977869451046,
0.12946246564388275,
-0.028047829866409302,
0.01969626545906067,
-0.05295034497976303,
-0.05613580718636513,
0.01239769533276558,
0.05226622149348259,
-0.015259640291333199,
-0.01807355135679245,
1.1044661148643085e-33,
-0.001171222305856645,
-0.012423697859048843,
0.07912436872720718,
0.04604291915893555,
0.0999113917350769,
-0.05790023133158684,
-0.059835053980350494,
-0.015378573909401894,
-0.021277926862239838,
-0.01345871388912201,
-0.04089028015732765,
0.031012248247861862,
-0.07484003156423569,
0.14150695502758026,
0.06210818141698837,
-0.10021012276411057,
-0.037141937762498856,
-0.0021777942311018705,
0.02543632686138153,
0.022411581128835678,
0.005578255280852318,
0.018169930204749107,
-0.021590737625956535,
-0.006427313666790724,
0.000017334505173494108,
-0.017118018120527267,
0.08355126529932022,
-0.008178167045116425,
0.04366801306605339,
0.02204267866909504,
-0.1499273031949997,
0.013508852571249008,
-0.018054060637950897,
-0.020706404000520706,
-0.020685605704784393,
-0.0255966167896986,
0.0166325680911541,
0.01160013023763895,
0.008968344889581203,
-0.018669478595256805,
0.04610377550125122,
0.008503197692334652,
-0.05786372348666191,
-0.01683758944272995,
-0.10961560904979706,
0.023370902985334396,
0.04183929041028023,
0.05339249223470688,
-0.09052728861570358,
-0.03773413598537445,
-0.0006144114886410534,
-0.0185542069375515,
-0.008574181236326694,
-0.14702706038951874,
-0.05592130497097969,
-0.02414865605533123,
0.10197540372610092,
-0.021148990839719772,
0.06958459317684174,
-0.055822353810071945,
-0.01502105500549078,
0.02471196837723255,
-0.015367582440376282,
0.05806835740804672,
0.026346120983362198,
-0.013643302023410797,
-0.04448169469833374,
0.062322914600372314,
0.017948973923921585,
0.01775190979242325,
-0.06170281022787094,
0.06483909487724304,
0.004016493447124958,
-0.10576124489307404,
0.07043784856796265,
-0.01866116188466549,
0.040664542466402054,
0.01451856829226017,
0.06281711906194687,
0.02803066372871399,
0.004276229999959469,
-0.0006816026289016008,
-0.013924642466008663,
-0.08828370273113251,
0.01756453886628151,
0.05061057582497597,
-0.02773902751505375,
-0.050845012068748474,
-0.03164266422390938,
-0.06876469403505325,
-0.0252972599118948,
-0.01133741158992052,
0.021274978294968605,
-0.004636258352547884,
0.03358752280473709,
-9.348625756853034e-34,
-0.012269265949726105,
0.08532467484474182,
0.01392567902803421,
0.05477273836731911,
0.00461910106241703,
0.006556012202054262,
0.018204087391495705,
0.029627375304698944,
0.1018478199839592,
-0.04780760779976845,
-0.0015911048976704478,
-0.01617211475968361,
-0.036771275103092194,
-0.015724439173936844,
0.0030563492327928543,
-0.01995285414159298,
0.04381607845425606,
0.024530043825507164,
-0.10866081714630127,
-0.02108399197459221,
-0.01637384295463562,
0.0725606307387352,
-0.050779152661561966,
0.05339430645108223,
-0.08938278257846832,
0.08077412843704224,
-0.05326167121529579,
0.03562590852379799,
0.002834889804944396,
-0.017982814460992813,
-0.0161559097468853,
0.013951636850833893,
0.07049502432346344,
-0.09227946400642395,
-0.12246353924274445,
0.05403595417737961,
0.032767247408628464,
-0.05811778083443642,
-0.09916017949581146,
0.09259022027254105,
-0.0025421485770493746,
-0.07782195508480072,
-0.04693380743265152,
0.047619663178920746,
0.04619157314300537,
-0.08375542610883713,
-0.01552104577422142,
-0.04450412839651108,
-0.020212918519973755,
0.10212372243404388,
0.008441026322543621,
0.10048963129520416,
-0.11181404441595078,
0.07511502504348755,
-0.018171435222029686,
-0.10065264999866486,
-0.0394081212580204,
0.027950381860136986,
0.10672135651111603,
0.06388281285762787,
-0.0744597464799881,
-0.030888056382536888,
-0.021119626238942146,
-0.03276187926530838,
-0.030895505100488663,
-0.012296219356358051,
-0.06176934763789177,
-0.014357516542077065,
0.07356072217226028,
0.08795931935310364,
-0.05844741314649582,
0.015372234396636486,
0.005143793765455484,
-0.043526265770196915,
0.03620103746652603,
-0.09652628004550934,
0.01395487505942583,
0.029186638072133064,
-0.01731564849615097,
0.06747491657733917,
0.01567041128873825,
-0.05561318248510361,
0.018050547689199448,
0.03616122901439667,
0.05854549631476402,
0.00743458466604352,
-0.03183802589774132,
0.02015974000096321,
0.03393496945500374,
0.01807655207812786,
0.026057902723550797,
-0.018611304461956024,
-0.04264885187149048,
0.05851471424102783,
-0.03901892900466919,
-4.275477394344307e-8,
-0.07608641684055328,
0.020487193018198013,
0.006176697090268135,
0.008324988186359406,
0.00011304493818897754,
-0.08989372849464417,
0.01690194569528103,
0.09800683706998825,
0.009724083356559277,
-0.10049037635326385,
0.06431815773248672,
-0.009335269220173359,
-0.058427657932043076,
-0.04852283373475075,
0.0423179492354393,
0.11662138253450394,
0.0055472515523433685,
0.029952269047498703,
0.035577256232500076,
0.047282520681619644,
0.018237438052892685,
-0.002534835133701563,
0.035840246826410294,
-0.036292027682065964,
0.039061762392520905,
-0.030416300520300865,
-0.09806898981332779,
-0.012645548209547997,
0.04499291256070137,
0.002791615203022957,
0.020004980266094208,
0.022225171327590942,
-0.004505150485783815,
-0.06721368432044983,
-0.02335883304476738,
0.042892131954431534,
0.05215398594737053,
-0.0024645996745675802,
-0.03730609640479088,
-0.038124896585941315,
-0.03382589668035507,
0.03638238459825516,
-0.016925493255257607,
0.056520454585552216,
-0.030328677967190742,
0.03261740133166313,
0.0386081337928772,
-0.016706988215446472,
0.0020646918565034866,
-0.02268144302070141,
0.004980115685611963,
-0.027858642861247063,
0.02247832901775837,
0.08084713667631149,
0.047984521836042404,
-0.053087733685970306,
0.02578761614859104,
0.010277614928781986,
0.031157808378338814,
0.07952083647251129,
0.038652192801237106,
0.02629059925675392,
-0.005142459645867348,
0.014283863827586174
] |
COME: Test-time Adaption by Conservatively Minimizing Entropy | https://openreview.net/forum?id=506BjJ1ziZ | [
"Qingyang Zhang",
"Yatao Bian",
"Xinke Kong",
"Peilin Zhao",
"Changqing Zhang"
] | Poster | Machine learning models must continuously self-adjust themselves for novel data distribution in the open world. As the predominant principle, entropy minimization (EM) has been proven to be a simple yet effective cornerstone in existing test-time adaption (TTA) methods. While unfortunately its fatal limitation (i.e., overconfidence) tends to result in model collapse. For this issue, we propose to \textbf{\texttt{Co}}nservatively \textbf{\texttt{M}}inimize the \textbf{\texttt{E}}ntropy (\texttt{COME}), which is a simple drop-in replacement of traditional EM to elegantly address the limitation. In essence, \texttt{COME} explicitly models the uncertainty by characterizing a Dirichlet prior distribution over model predictions during TTA. By doing so, \texttt{COME} naturally regularizes the model to favor conservative confidence on unreliable samples. Theoretically, we provide a preliminary analysis to reveal the ability of \texttt{COME} in enhancing the optimization stability by introducing a data-adaptive lower bound on the entropy. Empirically, our method achieves state-of-the-art performance on commonly used benchmarks, showing significant improvements in terms of classification accuracy and uncertainty estimation under various settings including standard, life-long and open-world TTA, i.e., up to $34.5\%$ improvement on accuracy and $15.1\%$ on false positive rate. Our code is available at: \href{https://github.com/BlueWhaleLab/COME}{https://github.com/BlueWhaleLab/COME}. | Test-time adaption, Out-of-distribution generalization | We propose an alternative to entropy minimization as a better learning principle for TTA tasks. | 632 | 2410.10894 | [
-0.10181792825460434,
0.0009770292090252042,
0.05183495581150055,
0.07075437903404236,
0.04090658575296402,
-0.05183713510632515,
-0.017441842705011368,
0.02600996382534504,
0.008600008673965931,
-0.008673451840877533,
-0.011979416944086552,
-0.026746120303869247,
0.046950798481702805,
-0.02591235749423504,
-0.0342579111456871,
-0.043390825390815735,
0.13695594668388367,
0.08217798173427582,
-0.08893522620201111,
0.001070803264155984,
-0.017809202894568443,
0.08604824542999268,
0.013089379295706749,
0.03478951379656792,
-0.008931959047913551,
-0.0723562017083168,
-0.011796321719884872,
0.003482689382508397,
0.05440693721175194,
-0.0031637190841138363,
0.021206669509410858,
-0.01764894649386406,
-0.03961245343089104,
0.017135273665189743,
0.021395428106188774,
0.00107397953979671,
0.003575982293114066,
0.0017862968379631639,
0.03231346979737282,
0.043082863092422485,
0.02193707972764969,
0.025502081960439682,
0.033647969365119934,
0.07064665108919144,
0.04381323605775833,
-0.009900855831801891,
0.038063425570726395,
-0.07822787761688232,
-0.052760135382413864,
-0.019737225025892258,
-0.015301491133868694,
0.046291470527648926,
-0.054780494421720505,
-0.0004927967092953622,
-0.0036215076688677073,
0.010586204007267952,
0.007807441521435976,
-0.030830109491944313,
-0.022497570142149925,
0.026285530999302864,
-0.01548092346638441,
-0.09595627337694168,
-0.07197604328393936,
0.020533185452222824,
0.07022122293710709,
-0.03395445644855499,
0.05971984565258026,
0.03717179223895073,
-0.017385462298989296,
0.08470809459686279,
-0.0553857758641243,
0.06981465965509415,
-0.13771924376487732,
0.1064191609621048,
0.017300723120570183,
0.06371057778596878,
0.06439439952373505,
-0.026694688946008682,
0.06501331925392151,
-0.0499429926276207,
-0.010624862276017666,
0.036058686673641205,
0.05041910707950592,
-0.05074107646942139,
0.09710051119327545,
-0.027976585552096367,
0.020712601020932198,
-0.015381795354187489,
0.09692378342151642,
-0.04672783985733986,
-0.062341418117284775,
-0.031423043459653854,
-0.003224120009690523,
0.04781704023480415,
-0.004933646880090237,
0.019578928127884865,
0.049913372844457626,
-0.05012304335832596,
0.015394491143524647,
0.09786083549261093,
0.0014306219527497888,
0.09362050145864487,
0.0015374684007838368,
-0.037446390837430954,
-0.02247762121260166,
-0.09592557698488235,
0.05260941758751869,
0.010784369893372059,
0.09430525451898575,
-0.025237105786800385,
0.05619696527719498,
0.05742759257555008,
-0.028273213654756546,
-0.03869073837995529,
-0.012545722536742687,
-0.03933097422122955,
-0.04604226350784302,
-0.004327388014644384,
-0.0069730207324028015,
0.088466115295887,
-0.02426215074956417,
0.049557141959667206,
0.008212077431380749,
0.011400255374610424,
0.01684369146823883,
0.01307753287255764,
0.017080221325159073,
3.132163472483526e-33,
0.026152800768613815,
0.03430861979722977,
0.02334846369922161,
-0.010700623504817486,
0.025293665006756783,
0.05739668756723404,
-0.062152355909347534,
-0.027031851932406425,
-0.02812960185110569,
0.027111809700727463,
-0.03625266253948212,
-0.05258829519152641,
-0.08792722225189209,
-0.04725942388176918,
0.03948815166950226,
0.04407582804560661,
-0.07411941885948181,
0.02645150199532509,
0.05268799513578415,
-0.00911788735538721,
0.08149344474077225,
-0.0929357185959816,
-0.0005030780448578298,
-0.09337445348501205,
-0.011045056395232677,
-0.03703474998474121,
0.058524444699287415,
0.01180352084338665,
-0.04340988025069237,
0.043323591351509094,
-0.11533357203006744,
0.018489336594939232,
-0.0025410884991288185,
0.050086215138435364,
0.03950507566332817,
-0.06387999653816223,
-0.027055665850639343,
-0.053471703082323074,
-0.05557849630713463,
0.010892098769545555,
-0.06587903946638107,
0.034561123698949814,
0.02451588772237301,
-0.04658213257789612,
-0.001285472884774208,
-0.13118542730808258,
0.07792343199253082,
-0.08044520765542984,
0.004798116162419319,
-0.1363941729068756,
0.015482163988053799,
-0.0012318509398028255,
0.020385554060339928,
-0.03009096160531044,
-0.006356348283588886,
0.026127908378839493,
0.05719190090894699,
-0.021364856511354446,
0.0050772870890796185,
0.05873231217265129,
-0.05755417048931122,
0.012784121558070183,
0.04763774201273918,
-0.004750929772853851,
0.02757837064564228,
0.056465815752744675,
0.010802381671965122,
-0.041999317705631256,
0.027019279077649117,
-0.001481359824538231,
-0.005368590820580721,
-0.06098347157239914,
-0.08740141987800598,
-0.06510300189256668,
0.060601118952035904,
-0.04534591734409332,
0.12257371842861176,
-0.0035757827572524548,
0.03092797100543976,
0.02213742583990097,
0.052733611315488815,
-0.07382922619581223,
0.016524244099855423,
-0.04944562911987305,
0.0036938448902219534,
-0.0431797094643116,
0.00012820154370274395,
-0.04563284292817116,
-0.09748999774456024,
-0.045134104788303375,
-0.07399458438158035,
-0.01564209721982479,
0.020609430968761444,
-0.01478673703968525,
-0.019493799656629562,
-3.499661944800944e-33,
-0.07938503473997116,
-0.07402769476175308,
-0.0061930157244205475,
0.039795078337192535,
-0.06006908044219017,
-0.06810423731803894,
-0.03711610287427902,
0.004237496759742498,
-0.04457179084420204,
-0.10429425537586212,
0.0037721574772149324,
-0.027122339233756065,
0.07023707777261734,
-0.05158497765660286,
-0.0005047729937359691,
0.04342547431588173,
0.014073256403207779,
0.07203339040279388,
0.04491914436221123,
0.01287531666457653,
0.07701389491558075,
-0.000321114290272817,
-0.13871364295482635,
0.033732809126377106,
0.010201320983469486,
0.0013971758307889104,
-0.05744487792253494,
0.11136582493782043,
0.023819733411073685,
-0.06901860982179642,
-0.038667090237140656,
-0.023362917825579643,
-0.047425415366888046,
0.020048867911100388,
-0.06173741817474365,
0.11732834577560425,
0.06856460869312286,
0.023462416604161263,
-0.022125238552689552,
0.14688889682292938,
0.04179475083947182,
0.052721504122018814,
-0.08384596556425095,
-0.05290262773633003,
0.013722593896090984,
-5.3372719577282624e-8,
-0.03795396536588669,
-0.0555182509124279,
0.09241065382957458,
0.060452111065387726,
0.06016605719923973,
0.006308095995336771,
-0.02243863418698311,
0.12267006933689117,
-0.09647179394960403,
-0.03822748735547066,
-0.059697914868593216,
-0.027542565017938614,
0.05063639208674431,
0.04800622537732124,
-0.060745108872652054,
-0.009962489828467369,
0.012848341837525368,
0.01657804846763611,
0.004518083296716213,
-0.032375678420066833,
-0.030026448890566826,
0.07069683074951172,
0.06777970492839813,
0.031239045783877373,
0.032115448266267776,
-0.006892802659422159,
-0.010358978994190693,
-0.0052524167113006115,
0.03912412375211716,
0.0030963949393481016,
0.09254603832960129,
-0.030794555321335793,
-0.017105577513575554,
-0.02573922462761402,
-0.012944328598678112,
0.07634539902210236,
0.035840459167957306,
0.064510278403759,
0.07092171907424927,
0.05807308852672577,
0.08787812292575836,
0.032059911638498306,
0.00483131455257535,
-0.007102985866367817,
-0.061817094683647156,
0.05985653027892113,
0.007545504719018936,
0.04231540113687515,
-0.0955486074090004,
-5.6253920632798327e-8,
-0.01312801893800497,
-0.007574933115392923,
0.034128062427043915,
-0.02721926011145115,
0.11289151757955551,
-0.011231270618736744,
-0.07751012593507767,
-0.014531952328979969,
-0.043425265699625015,
-0.04789695888757706,
0.10500919073820114,
0.06460968405008316,
0.018871916458010674,
-0.038178328424692154,
-0.032749880105257034,
0.07271947711706161,
-0.0688048005104065,
-0.011667635291814804,
-0.01185965072363615,
-0.07457533478736877,
0.017969170585274696,
0.03389876335859299,
0.056204717606306076,
-0.018957536667585373,
0.03384818509221077,
-0.030668111518025398,
0.011792869307100773,
0.08816282451152802,
0.013158163987100124,
0.018542885780334473,
-0.039174873381853104,
-0.027350831776857376,
-0.0182760301977396,
-0.02390017732977867,
-0.04042285680770874,
0.10645215213298798,
0.05055435746908188,
-0.03868312016129494,
-0.0568414032459259,
-0.041472285985946655,
-0.021408362314105034,
0.05064711347222328,
-0.07265972346067429,
0.026791751384735107,
0.04228145629167557,
-0.04005079343914986,
-0.003262139158323407,
-0.0003490733215585351,
-0.0059951301664114,
-0.08511935919523239,
0.03638308867812157,
0.03620052710175514,
-0.02227470651268959,
0.05219189450144768,
0.049008648842573166,
-0.030475536361336708,
-0.004802212584763765,
-0.05737866088747978,
-0.028215227648615837,
0.04789496213197708,
0.05067498981952667,
-0.001724354107864201,
-0.04779897630214691,
-0.024485023692250252
] |
Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology | https://openreview.net/forum?id=trj2Jq8riA | [
"Pei Liu",
"Luping Ji",
"Jiaxiang Gou",
"Bo Fu",
"Mao Ye"
] | Poster | Histopathology Whole-Slide Images (WSIs) provide an important tool to assess cancer prognosis in computational pathology (CPATH). While existing survival analysis (SA) approaches have made exciting progress, they are generally limited to adopting highly-expressive network architectures and only coarse-grained patient-level labels to learn visual prognostic representations from gigapixel WSIs. Such learning paradigm suffers from critical performance bottlenecks, when facing present scarce training data and standard multi-instance learning (MIL) framework in CPATH. To overcome it, this paper, for the first time, proposes a new Vision-Language-based SA (**VLSA**) paradigm. Concretely, (1) VLSA is driven by pathology VL foundation models. It no longer relies on high-capability networks and shows the advantage of *data efficiency*. (2) In vision-end, VLSA encodes textual prognostic prior and then employs it as *auxiliary signals* to guide the aggregating of visual prognostic features at instance level, thereby compensating for the weak supervision in MIL. Moreover, given the characteristics of SA, we propose i) *ordinal survival prompt learning* to transform continuous survival labels into textual prompts; and ii) *ordinal incidence function* as prediction target to make SA compatible with VL-based prediction. Notably, VLSA's predictions can be interpreted intuitively by our Shapley values-based method. The extensive experiments on five datasets confirm the effectiveness of our scheme. Our VLSA could pave a new way for SA in CPATH by offering weakly-supervised MIL an effective means to learn valuable prognostic clues from gigapixel WSIs. Our source code is available at https://github.com/liupei101/VLSA. | Computation Pathology, Survival Analysis, Multi-Instance Learning, Whole-Slide Images, Vision-Language Modes | This paper first introduces Vision-Language Survival Analysis (VLSA) paradigm for computational pathology. | 631 | 2409.09369 | [
0.033990528434515,
-0.05670858174562454,
-0.015526553615927696,
-0.052940733730793,
0.1017567366361618,
0.039790090173482895,
-0.06903612613677979,
0.02122381329536438,
0.039417047053575516,
-0.10806474089622498,
-0.0813605934381485,
-0.039207860827445984,
0.043263990432024,
0.09011373668909073,
-0.0876021757721901,
0.03151809424161911,
-0.006460609845817089,
0.03736637905240059,
-0.10065314918756485,
-0.069622702896595,
0.005514970514923334,
0.02464321255683899,
0.0321420393884182,
0.027834132313728333,
0.0629790872335434,
-0.012508400715887547,
-0.026999471709132195,
-0.025491392239928246,
0.014707397669553757,
0.0014290474355220795,
0.010980430990457535,
-0.07300107181072235,
-0.046473219990730286,
0.07368621975183487,
-0.028464538976550102,
0.06638990342617035,
-0.03586476296186447,
0.03257910907268524,
-0.11537455767393112,
0.01194562017917633,
-0.03310292214155197,
0.017983675003051758,
-0.007487450260668993,
0.0030626009684056044,
0.10632327944040298,
-0.04620688036084175,
-0.07466000318527222,
-0.055963438004255295,
-0.012327187694609165,
0.06150559335947037,
-0.11884717643260956,
-0.053528305143117905,
-0.04995940625667572,
0.07330112159252167,
0.017349958419799805,
-0.007230993825942278,
0.0021715068724006414,
-0.04394654557108879,
-0.06303825229406357,
-0.02287214994430542,
-0.02775115892291069,
-0.09816207736730576,
0.0006079397862777114,
0.05943711847066879,
-0.007217656821012497,
-0.025639966130256653,
0.05876721814274788,
0.006531471852213144,
-0.08005959540605545,
0.057235196232795715,
-0.04453573748469353,
0.03721795976161957,
0.029563134536147118,
0.12796208262443542,
-0.026846762746572495,
0.03200152516365051,
0.09491768479347229,
0.014963568188250065,
0.11109162867069244,
-0.09775899350643158,
0.036315273493528366,
0.09694555401802063,
0.059698689728975296,
0.02165563963353634,
0.11200440675020218,
0.020801501348614693,
-0.0203542560338974,
0.0076072621159255505,
-0.022352661937475204,
0.03854327276349068,
-0.0013534054160118103,
-0.015468787401914597,
0.0373600572347641,
-0.008341076783835888,
0.053203027695417404,
0.0027928941417485476,
-0.04741727188229561,
-0.07403597235679626,
0.013163426890969276,
0.012564627453684807,
-0.07419929653406143,
-0.0391581729054451,
0.04616319388151169,
-0.04115694761276245,
0.00750227365642786,
-0.022498207166790962,
0.014100540429353714,
0.04317239671945572,
0.053007226437330246,
-0.049068547785282135,
0.03984016925096512,
0.004372214898467064,
0.0150926997885108,
-0.013870025984942913,
0.004640215542167425,
0.019220737740397453,
-0.07598785310983658,
-0.049674246460199356,
0.0646158829331398,
0.02757973223924637,
-0.033364761620759964,
-0.030727175995707512,
-0.0033974635880440474,
0.028297780081629753,
0.060069430619478226,
0.026616135612130165,
0.003156655468046665,
3.0177965046851565e-33,
0.07657036185264587,
-0.039393745362758636,
0.03459775075316429,
-0.021250495687127113,
0.07284045964479446,
-0.03079369105398655,
0.005696370732039213,
-0.0011069432366639376,
-0.05477456748485565,
0.0010443347273394465,
0.015513213351368904,
0.03595299646258354,
-0.019781917333602905,
0.03437822312116623,
-0.007407315541058779,
-0.0046887886710464954,
0.07097794860601425,
0.06283923238515854,
-0.11891139298677444,
-0.04530381038784981,
-0.005602727644145489,
-0.03982187435030937,
-0.0022261571139097214,
-0.07110531628131866,
0.006658399943262339,
0.046508174389600754,
0.017978621646761894,
0.020570475608110428,
0.013797263614833355,
0.015304259955883026,
-0.11479706317186356,
0.023034920915961266,
0.042335547506809235,
-0.04262113943696022,
0.016742674633860588,
-0.08013300597667694,
-0.050769586116075516,
-0.028839783743023872,
0.1069064810872078,
0.03541552275419235,
-0.010525557212531567,
0.01642688736319542,
0.02291155606508255,
-0.05728123337030411,
-0.06626752018928528,
-0.019084507599473,
0.036678213626146317,
0.11299015581607819,
-0.10701275616884232,
-0.05866542458534241,
0.04324334114789963,
-0.03442142531275749,
-0.028644800186157227,
-0.031054487451910973,
0.04847201704978943,
0.053352151066064835,
-0.0024185669608414173,
0.053692370653152466,
0.009686128236353397,
-0.021230066195130348,
0.04697318375110626,
0.06548133492469788,
0.00641945656388998,
0.1030752956867218,
-0.013980408199131489,
-0.010195259936153889,
-0.051683709025382996,
-0.03253825381398201,
0.023851027712225914,
0.009359505027532578,
-0.029201898723840714,
0.09556522220373154,
-0.03485269099473953,
-0.056926652789115906,
0.03467487543821335,
-0.037398725748062134,
0.03587512671947479,
-0.03568155691027641,
-0.010113682597875595,
0.11615104973316193,
-0.05754782631993294,
-0.008408374153077602,
-0.025311574339866638,
-0.02472868748009205,
-0.032049763947725296,
-0.01513117365539074,
0.04036330059170723,
-0.04138665273785591,
-0.06600940227508545,
-0.006798725575208664,
-0.06402130424976349,
-0.02106095850467682,
0.01809757575392723,
-0.004316236358135939,
0.006744649726897478,
-3.0168792517495304e-33,
-0.004406095016747713,
0.09555812925100327,
0.059238824993371964,
-0.0019023262429982424,
-0.01482423022389412,
-0.061703164130449295,
-0.049454592168331146,
0.018550727516412735,
0.015305990353226662,
-0.029311062768101692,
0.06212979182600975,
-0.00201913109049201,
-0.018387168645858765,
0.0028150014113634825,
0.012504490092396736,
-0.0031048657838255167,
-0.046512313187122345,
-0.014482605271041393,
-0.0071167959831655025,
0.015518909320235252,
-0.005603667814284563,
0.13877515494823456,
-0.1479528695344925,
0.05657055601477623,
-0.06489226967096329,
0.06707996875047684,
0.04603719338774681,
0.042841631919145584,
-0.009891112335026264,
-0.09095500409603119,
-0.002225202275440097,
-0.032786108553409576,
-0.021552111953496933,
-0.04008667916059494,
-0.04295125976204872,
0.05335504189133644,
0.020909082144498825,
-0.17030467092990875,
-0.06665675342082977,
0.06013598293066025,
0.048919159919023514,
-0.02611551620066166,
-0.08291593194007874,
-0.05942901223897934,
0.003372412407770753,
-0.01101678516715765,
0.0412522628903389,
0.0490485280752182,
0.08366090804338455,
0.05915412679314613,
-0.08081186562776566,
0.013704697601497173,
-0.07668997347354889,
0.02424287609755993,
-0.059100937098264694,
-0.07195109874010086,
-0.07633212953805923,
-0.003597603179514408,
0.04343170300126076,
-0.09134628623723984,
-0.0004426738014444709,
-0.03757406771183014,
-0.00598621042445302,
0.043149568140506744,
-0.020737117156386375,
0.06883977353572845,
-0.018310334533452988,
-0.011457040905952454,
-0.026945145800709724,
0.0325947068631649,
0.017824821174144745,
0.03353896737098694,
-0.10070677101612091,
0.005942176561802626,
-0.012177438475191593,
0.006386500783264637,
-0.011465797200798988,
-0.018092062324285507,
-0.011204901151359081,
0.004400501027703285,
0.03977882117033005,
-0.048541583120822906,
0.0357789546251297,
0.05193905904889107,
0.10252924263477325,
0.12170272320508957,
0.027726387605071068,
-0.02535071037709713,
0.06258294731378555,
-0.029470115900039673,
-0.04891485720872879,
-0.03504042327404022,
-0.049908947199583054,
0.05846288800239563,
-0.06986483186483383,
-4.445063339630906e-8,
-0.06264226138591766,
0.02985139936208725,
-0.0322544202208519,
-0.08575980365276337,
0.041568659245967865,
-0.07664728164672852,
0.00277514080516994,
0.06783746927976608,
0.02973540686070919,
0.09002642333507538,
-0.0027666541282087564,
0.012502918019890785,
-0.03619817644357681,
-0.10387209057807922,
0.029748497530817986,
0.07616342604160309,
0.049775876104831696,
0.0415499284863472,
0.015064546838402748,
0.012230363674461842,
0.0058784326538443565,
-0.049737349152565,
0.06912645697593689,
0.02615019492805004,
0.006074453238397837,
-0.06077880039811134,
-0.023885387927293777,
0.07965444773435593,
0.04505104571580887,
0.010214979760348797,
0.008113072253763676,
0.04743924364447594,
0.04892250522971153,
0.012431815266609192,
0.05186276137828827,
0.05073928460478783,
0.01774316094815731,
0.028830187395215034,
-0.01191168837249279,
0.05440234765410423,
0.05614633113145828,
0.05726366117596626,
-0.0007461846107617021,
0.010505574755370617,
-0.05405624210834503,
0.01778051257133484,
0.027963990345597267,
-0.07643058151006699,
-0.015006649307906628,
-0.019982805475592613,
-0.03602487966418266,
-0.058293044567108154,
-0.0024088153149932623,
0.08707331120967865,
0.03987230360507965,
-0.014574789442121983,
0.08043602108955383,
-0.045787714421749115,
0.013103538192808628,
0.11498094350099564,
0.01789359375834465,
-0.026067113503813744,
0.03442298620939255,
0.012876770459115505
] |
SleepSMC: Ubiquitous Sleep Staging via Supervised Multimodal Coordination | https://openreview.net/forum?id=B5VEi5d3p2 | [
"Shuo Ma",
"Yingwei Zhang",
"Yiqiang Chen",
"Hualei Wang",
"Yuan Jin",
"Wei Zhang",
"Ziyu Jia"
] | Poster | Sleep staging is critical for assessing sleep quality and tracking health. Polysomnography (PSG) provides comprehensive multimodal sleep-related information, but its complexity and impracticality limit its practical use in daily and ubiquitous monitoring. Conversely, unimodal devices offer more convenience but less accuracy. Existing multimodal learning paradigms typically assume that the data types remain consistent between the training and testing phases. This makes it challenging to leverage information from other modalities in ubiquitous scenarios (e.g., at home) where only one modality is available. To address this issue, we introduce a novel framework for ubiquitous Sleep staging via Supervised Multimodal Coordination, called SleepSMC. To capture category-related consistency and complementarity across modality-level instances, we propose supervised modality-level instance contrastive coordination. Specifically, modality-level instances within the same category are considered positive pairs, while those from different categories are considered negative pairs. To explore the varying reliability of auxiliary modalities, we calculate uncertainty estimates based on the variance in confidence scores for correct predictions during multiple rounds of random masks. These uncertainty estimates are employed to assign adaptive weights to multiple auxiliary modalities during contrastive learning, ensuring that the primary modality learns from high-quality, category-related features. Experimental results on four public datasets, ISRUC-S3, MASS-SS3, Sleep-EDF-78, and ISRUC-S1, show that SleepSMC achieves state-of-the-art cross-subject performance. SleepSMC significantly improves performance when only one modality is present during testing, making it suitable for ubiquitous sleep monitoring. | Sleep staging, Multimodal coordination, Ubiquitous computing | null | 626 | null | [
0.037508390843868256,
-0.08723380416631699,
0.05476311221718788,
0.060069598257541656,
0.11422321945428848,
-0.006469360087066889,
0.025025855749845505,
-0.08543235808610916,
0.023540766909718513,
-0.06494113057851791,
-0.06250805407762527,
-0.1023510992527008,
0.04008333757519722,
0.06222741678357124,
0.0324530266225338,
-0.04708775505423546,
0.1101989895105362,
0.10525477677583694,
-0.031125036999583244,
0.048682693392038345,
0.08615560829639435,
0.0314137302339077,
0.08193281292915344,
0.02522202581167221,
0.07702522724866867,
0.027845142409205437,
-0.004151054657995701,
-0.0773463323712349,
0.057072240859270096,
-0.02601194940507412,
-0.038755446672439575,
0.023446835577487946,
-0.04288675636053085,
0.020797329023480415,
-0.00020552730711642653,
0.03487803786993027,
0.018371835350990295,
0.003947513177990913,
-0.07624545693397522,
0.00790600385516882,
0.05429358035326004,
0.03201160952448845,
0.033898890018463135,
-0.027212994173169136,
-0.035056643187999725,
0.039660435169935226,
-0.04201967269182205,
-0.097280353307724,
-0.0037110811099410057,
0.030615605413913727,
-0.0054049016907811165,
-0.05619766190648079,
0.027296729385852814,
0.041205521672964096,
-0.05368867143988609,
0.061205554753541946,
0.011060820892453194,
-0.02857878804206848,
0.06181346997618675,
0.026260534301400185,
0.0016467192908748984,
-0.045731548219919205,
0.0024002897553145885,
-0.029750525951385498,
0.006644543260335922,
0.05086594820022583,
-0.013464311137795448,
-0.028222789987921715,
0.0826159119606018,
-0.022527964785695076,
-0.10956604778766632,
0.030505232512950897,
0.012257728725671768,
0.007454841863363981,
-0.04842592030763626,
0.02564062550663948,
0.08489103615283966,
-0.07786854356527328,
0.04142975062131882,
-0.06894749402999878,
0.03506877273321152,
0.03126144036650658,
0.0526251420378685,
-0.05165364220738411,
0.11663525551557541,
-0.00315296882763505,
0.0062919254414737225,
0.10695056617259979,
-0.07139616459608078,
0.0663856714963913,
-0.08459675312042236,
0.03222639858722687,
-0.07056873291730881,
-0.035969655960798264,
0.01389141846448183,
-0.0010054628364741802,
0.0008056826773099601,
-0.02596661075949669,
0.05674722045660019,
0.08124513924121857,
0.01593373343348503,
0.012707379646599293,
-0.02155832201242447,
-0.01590912975370884,
-0.0014532082714140415,
-0.051859356462955475,
0.020044023171067238,
-0.09089917689561844,
0.027125198394060135,
-0.001655835541896522,
-0.026157213374972343,
0.021397622302174568,
0.05711246281862259,
-0.0068019600585103035,
0.06261445581912994,
0.0257648266851902,
-0.008742475882172585,
0.10330364853143692,
0.08087639510631561,
0.016531502828001976,
0.02238922007381916,
-0.06743228435516357,
0.09527336806058884,
-0.08921248465776443,
0.0679924488067627,
-0.005290557164698839,
-0.08238524943590164,
1.139990105159135e-33,
0.07166151702404022,
-0.015352064743638039,
0.046050094068050385,
0.07515337318181992,
0.05398130044341087,
0.002464433666318655,
-0.12175044417381287,
0.038467057049274445,
-0.0045676338486373425,
0.03225870430469513,
-0.00005194326877244748,
0.0650918260216713,
-0.07863930612802505,
0.018261274322867393,
0.05230157449841499,
0.005232689902186394,
-0.028854215517640114,
0.05024857074022293,
-0.00879936944693327,
-0.010965756140649319,
-0.03093033656477928,
-0.07209186255931854,
-0.04336972162127495,
0.03799042850732803,
-0.01917644403874874,
0.027629181742668152,
0.05271998420357704,
0.021804431453347206,
0.05527530610561371,
0.018723666667938232,
-0.11730597168207169,
-0.01854466088116169,
-0.015706131234765053,
0.0002522082650102675,
0.07795686274766922,
-0.055195216089487076,
-0.00635859090834856,
-0.05100492015480995,
-0.05102413520216942,
-0.026796746999025345,
-0.06806095689535141,
0.02388802543282509,
-0.05053615942597389,
-0.03559238463640213,
-0.042308297008275986,
-0.014082011766731739,
0.06521981954574585,
0.05502405762672424,
-0.06085415184497833,
-0.06153038516640663,
-0.057423558086156845,
0.01924200914800167,
-0.03211621940135956,
-0.15347805619239807,
-0.07710299640893936,
-0.03848159313201904,
-0.013953584246337414,
-0.013943744823336601,
-0.007960067130625248,
0.018560608848929405,
0.03671441599726677,
0.060542359948158264,
-0.035167206078767776,
0.023847656324505806,
0.02713816799223423,
0.05466395616531372,
-0.016057951375842094,
0.009332694113254547,
0.0370151661336422,
-0.09228846430778503,
0.004522962030023336,
0.04033895209431648,
-0.014573495835065842,
0.0008695668075233698,
0.05031539127230644,
0.014242093078792095,
0.02613251283764839,
0.02239581011235714,
-0.040263693779706955,
0.014904915355145931,
0.05353955924510956,
0.021640483289957047,
-0.020044270902872086,
-0.012526885606348515,
-0.05297407507896423,
-0.08286617696285248,
-0.03863818198442459,
-0.051621053367853165,
-0.1348857581615448,
0.07732138782739639,
-0.09913024306297302,
-0.013660836033523083,
0.04003439471125603,
0.02463044971227646,
-0.07178838551044464,
-1.6391603833224615e-33,
0.0028963075019419193,
-0.022731270641088486,
0.00756674213334918,
-0.02911367453634739,
0.043167367577552795,
-0.06856771558523178,
0.03006884828209877,
-0.056436873972415924,
0.00980394147336483,
-0.03955848515033722,
0.021195311099290848,
-0.044656701385974884,
0.03988463431596756,
-0.03773362562060356,
0.035599302500486374,
-0.0010491844732314348,
-0.03716789931058884,
0.03459702804684639,
-0.05220845714211464,
0.056303635239601135,
0.03927256166934967,
-0.009410122409462929,
-0.030629338696599007,
0.023842250928282738,
-0.04268724098801613,
0.07668495178222656,
-0.06572715193033218,
0.0709746927022934,
0.0016030531842261553,
-0.052558936178684235,
-0.013827552087605,
-0.052271194756031036,
0.002652588998898864,
-0.03723306208848953,
0.027172233909368515,
0.03378262743353844,
0.006752613466233015,
-0.06011802703142166,
-0.055348437279462814,
0.01723392866551876,
0.09622619301080704,
0.019266551360487938,
-0.07330838590860367,
-0.044304259121418,
0.0663967877626419,
-0.016580434516072273,
-0.0975288525223732,
-0.04970027506351471,
0.027680575847625732,
0.01672361046075821,
0.018369827419519424,
-0.003193870885297656,
-0.1397109180688858,
-0.04487744718790054,
-0.065585657954216,
-0.056380465626716614,
-0.05419948697090149,
-0.032256804406642914,
0.015267269685864449,
0.014425036497414112,
-0.018037836998701096,
-0.02401748299598694,
-0.10494548082351685,
0.04645276814699173,
0.007899151183664799,
0.06567748636007309,
0.031081760302186012,
-0.06977251917123795,
0.04054193198680878,
0.016765225678682327,
0.023216160014271736,
-0.033238742500543594,
0.030095819383859634,
0.03775309771299362,
0.0049304356798529625,
-0.013014436699450016,
0.047001324594020844,
-0.04636133834719658,
-0.02200302481651306,
-0.05602576583623886,
0.07035382091999054,
-0.13703025877475739,
0.027334148064255714,
0.046313606202602386,
0.03904762864112854,
0.08967187255620956,
0.09827357530593872,
0.02578568272292614,
0.031308095902204514,
0.045472342520952225,
-0.006344091147184372,
0.06402795016765594,
-0.053771767765283585,
0.026955895125865936,
0.0022910935804247856,
-4.272557418971701e-8,
-0.053893670439720154,
-0.018816150724887848,
0.012882528826594353,
-0.07370755821466446,
-0.017824921756982803,
-0.03994586691260338,
-0.0011619937140494585,
-0.07271680980920792,
0.028635839000344276,
0.04122001677751541,
0.01632816717028618,
-0.07106360793113708,
0.05510813370347023,
-0.0556427538394928,
0.037201836705207825,
0.048059556633234024,
-0.019424887374043465,
0.04989282414317131,
0.0045199147425591946,
0.016022536903619766,
0.027088459581136703,
-0.005554052535444498,
-0.008505821228027344,
-0.01780368946492672,
0.10470081865787506,
-0.031012916937470436,
0.0009275272605009377,
0.07039619237184525,
0.054015930742025375,
0.008698026649653912,
0.024485623463988304,
-0.02809493988752365,
-0.025920787826180458,
0.0001489154965383932,
-0.08058237284421921,
0.04184730723500252,
0.058723412454128265,
-0.07338377088308334,
0.06588804721832275,
0.024906424805521965,
0.007242304272949696,
0.024274254217743874,
-0.09832600504159927,
0.06732228398323059,
0.05171173810958862,
-0.022653598338365555,
0.035150766372680664,
-0.1145755723118782,
-0.028741320595145226,
-0.002348795998841524,
-0.041345078498125076,
-0.020430272445082664,
0.029435407370328903,
0.09610014408826828,
0.07453388720750809,
-0.005149722099304199,
0.05099034309387207,
-0.0010259359842166305,
0.01621577888727188,
0.07389048486948013,
0.05137849226593971,
0.03705848380923271,
-0.11656933277845383,
-0.01624392531812191
] |
Towards Generalizable Reinforcement Learning via Causality-Guided Self-Adaptive Representations | https://openreview.net/forum?id=bMvqccRmKD | [
"Yupei Yang",
"Biwei Huang",
"Fan Feng",
"Xinyue Wang",
"Shikui Tu",
"Lei Xu"
] | Poster | General intelligence requires quick adaptation across tasks. While existing reinforcement learning (RL) methods have made progress in generalization, they typically assume only distribution changes between source and target domains. In this paper, we explore a wider range of scenarios where not only the distribution but also the environment spaces may change. For example, in the CoinRun environment, we train agents from easy levels and generalize them to difficulty levels where there could be new enemies that have never occurred before. To address this challenging setting, we introduce a causality-guided self-adaptive representation-based approach, called CSR, that equips the agent to generalize effectively across tasks with evolving dynamics. Specifically, we employ causal representation learning to characterize the latent causal variables within the RL system. Such compact causal representations uncover the structural relationships among variables, enabling the agent to autonomously determine whether changes in the environment stem from distribution shifts or variations in space, and to precisely locate these changes. We then devise a three-step strategy to fine-tune the causal model under different scenarios accordingly. Empirical experiments show that CSR efficiently adapts to the target domains with only a few samples and outperforms state-of-the-art baselines on a wide range of scenarios, including our simulated environments, CartPole, CoinRun and Atari games. | Reinforcement Learning, Transfer Learning | null | 620 | 2407.20651 | [
-0.04152025654911995,
-0.037561044096946716,
0.05275576561689377,
0.06889033317565918,
0.088865727186203,
0.04112634062767029,
-0.029711484909057617,
-0.056239038705825806,
-0.004367895424365997,
-0.02707471139729023,
-0.05058486387133598,
-0.10419177263975143,
0.0424121730029583,
0.09809806197881699,
0.029821911826729774,
-0.06587246060371399,
0.09041284769773483,
-0.01340392790734768,
-0.1058385893702507,
-0.053360167890787125,
-0.04305345192551613,
0.014185757376253605,
0.012334797531366348,
0.021516885608434677,
-0.027525946497917175,
-0.015910090878605843,
0.01465526595711708,
-0.021793019026517868,
-0.006208955310285091,
-0.043334003537893295,
0.10738436877727509,
-0.01663764752447605,
-0.06459394097328186,
-0.08611945062875748,
-0.035464633256196976,
0.049045052379369736,
-0.0744718611240387,
0.09618789702653885,
-0.009962520562112331,
-0.042333368211984634,
0.04655338451266289,
-0.00207764795050025,
-0.025837069377303123,
0.0051237791776657104,
0.026569971814751625,
0.0015488844364881516,
0.009880303405225277,
-0.01689828746020794,
0.024153899401426315,
-0.039192598313093185,
-0.04968741536140442,
-0.0036016798112541437,
0.0041358694434165955,
-0.016476480290293694,
0.04990355670452118,
0.008158530108630657,
0.07573315501213074,
0.04254554212093353,
-0.011583900079131126,
-0.05113062635064125,
0.011145049706101418,
-0.0795932337641716,
0.01844392716884613,
-0.0525394082069397,
-0.046697117388248444,
0.02433672547340393,
0.025344979017972946,
0.15447898209095,
-0.00451111514121294,
-0.0002539509441703558,
-0.01658111810684204,
0.06880515813827515,
-0.020735016092658043,
-0.03981498256325722,
0.06439566612243652,
-0.00581120140850544,
-0.03827517107129097,
0.004133153241127729,
0.06534653156995773,
-0.01904400810599327,
-0.0070008933544158936,
0.037412844598293304,
-0.039121050387620926,
0.07545807957649231,
0.1068013459444046,
-0.023217761889100075,
0.03340373560786247,
0.03333507478237152,
0.04630492255091667,
0.05178743600845337,
0.07152965664863586,
-0.030043726786971092,
0.04403598606586456,
0.051666345447301865,
0.03197665512561798,
0.041550252586603165,
0.03345108404755592,
-0.08778885006904602,
0.04653032124042511,
0.08321238309144974,
0.040952567011117935,
-0.035447973757982254,
-0.002062261803075671,
0.04331579431891441,
0.013779586181044579,
-0.015704583376646042,
-0.01595900021493435,
0.05133504420518875,
0.011384671553969383,
-0.05209469422698021,
0.0040627713315188885,
-0.0038799033500254154,
0.03622836992144585,
-0.014768684282898903,
0.012598581612110138,
-0.02612047828733921,
-0.023439668118953705,
0.019752414897084236,
-0.03226536512374878,
-0.00046285163261927664,
0.056762948632240295,
-0.08008214086294174,
-0.02819676138460636,
0.04125522822141647,
0.07652021199464798,
-0.05834212154150009,
-0.02557092159986496,
7.488981525648934e-35,
0.07026173919439316,
0.043482039123773575,
0.00785859115421772,
-0.011042383499443531,
0.0038759810850024223,
-0.01260980125516653,
-0.009254693984985352,
0.06086600571870804,
0.0035231439396739006,
0.039578743278980255,
0.021190175786614418,
0.052229490131139755,
-0.05489600449800491,
0.04417702928185463,
0.05571797117590904,
0.028035104274749756,
-0.13397246599197388,
0.09717346727848053,
-0.01151029858738184,
-0.09393275529146194,
-0.01238606870174408,
-0.0357782244682312,
-0.052876066416502,
-0.0740160420536995,
0.09305588901042938,
0.0006773017812520266,
0.052466556429862976,
-0.043269313871860504,
-0.022066786885261536,
-0.02665424719452858,
-0.008725786581635475,
0.027268974110484123,
0.022354817017912865,
0.09226906299591064,
0.015256349928677082,
-0.04677484184503555,
-0.11968815326690674,
-0.05145089700818062,
-0.032150667160749435,
0.009112107567489147,
-0.0033563943579792976,
0.022377707064151764,
0.009637315757572651,
0.03715822100639343,
-0.08333639800548553,
-0.12070558220148087,
-0.003808673471212387,
0.014236188493669033,
-0.13141708076000214,
-0.08244602382183075,
0.004386320244520903,
0.057468216866254807,
0.020434681326150894,
-0.07432334870100021,
0.006636950187385082,
0.029771964997053146,
-0.01990625448524952,
0.029985859990119934,
-0.034880317747592926,
-0.04060190171003342,
0.0955883339047432,
0.04478509724140167,
-0.041666023433208466,
0.12027989327907562,
0.05982498079538345,
0.013723710551857948,
-0.02506852149963379,
0.05358192324638367,
0.029311414808034897,
0.04785716161131859,
-0.008962806314229965,
0.04216020926833153,
-0.02140778675675392,
-0.01752145029604435,
0.08357232064008713,
-0.041021689772605896,
0.024736599996685982,
-0.06557932496070862,
-0.025715555995702744,
-0.05624178797006607,
-0.053442638367414474,
0.006326283328235149,
-0.0469956211745739,
0.05667409673333168,
-0.04220966249704361,
-0.011946354992687702,
-0.011030329391360283,
-0.07875615358352661,
-0.006075792945921421,
-0.09763824194669724,
-0.005735273472964764,
-0.08172131329774857,
-0.022503767162561417,
0.05614485591650009,
-0.05549296736717224,
-5.970905187902578e-34,
-0.00815568957477808,
-0.03715025633573532,
0.04066530987620354,
-0.038467008620500565,
0.02680760622024536,
-0.026576809585094452,
-0.046013589948415756,
-0.012855373322963715,
0.060381438583135605,
-0.039035994559526443,
-0.12914159893989563,
0.029612679034471512,
0.028681814670562744,
0.08219405263662338,
-0.05807406082749367,
-0.016984576359391212,
0.04275493323802948,
0.052560970187187195,
-0.010990881361067295,
-0.06430242210626602,
-0.030816124752163887,
0.05507748946547508,
-0.014677472412586212,
-0.002223180141299963,
0.01841973327100277,
0.0021046085748821497,
0.038304444402456284,
0.09448512643575668,
0.03618117421865463,
0.0692671537399292,
-0.06365857273340225,
0.012660062871873379,
-0.01597016490995884,
-0.041861820966005325,
-0.05528651922941208,
0.12065577507019043,
0.051052454859018326,
-0.03341836854815483,
-0.040773674845695496,
0.06764528155326843,
0.037090666592121124,
-0.05244126170873642,
-0.01174605917185545,
0.01819254457950592,
0.03096984513103962,
-0.0121284956112504,
-0.04592899605631828,
0.09722404181957245,
0.01793079264461994,
0.05091332271695137,
-0.045243363827466965,
0.03677455708384514,
-0.04092627763748169,
-0.008262930437922478,
-0.04180146008729935,
-0.011956091970205307,
0.017722390592098236,
-0.05475545674562454,
0.055226072669029236,
-0.02329934760928154,
-0.10658901929855347,
-0.020117461681365967,
0.09068957716226578,
0.05854696035385132,
-0.09678684920072556,
-0.022554585710167885,
-0.0595170333981514,
0.02880292758345604,
0.05254008620977402,
-0.006121184676885605,
0.05232594534754753,
0.01985243707895279,
-0.006845972966402769,
-0.07257012277841568,
0.014692199416458607,
-0.03959942236542702,
-0.0288523118942976,
-0.025721095502376556,
-0.05465307459235191,
-0.0364074669778347,
0.06411387771368027,
0.013305475935339928,
0.06234585493803024,
0.00828190054744482,
-0.02855244278907776,
0.11849195510149002,
0.06454218924045563,
0.09675777703523636,
0.0914383977651596,
0.0033031501807272434,
-0.0658893957734108,
-0.010240181349217892,
-0.04249246418476105,
0.0531357079744339,
-0.10460463911294937,
-4.330987835032829e-8,
-0.05524180084466934,
0.08301311731338501,
-0.0012706825509667397,
0.04754070192575455,
0.047038041055202484,
0.02818823605775833,
-0.05137081816792488,
-0.0034569192212074995,
0.0023130876943469048,
0.012727592140436172,
0.04567651450634003,
0.01399149838835001,
0.02249569073319435,
-0.02012542076408863,
-0.041851624846458435,
0.055195797234773636,
0.05485932156443596,
-0.02494768425822258,
-0.04232025519013405,
-0.025982974097132683,
0.04668981209397316,
0.04448005184531212,
-0.032320860773324966,
-0.0765659362077713,
0.062424227595329285,
-0.11406949907541275,
-0.021817415952682495,
0.03807603195309639,
0.044208623468875885,
-0.015522254630923271,
0.019797006621956825,
0.053751058876514435,
0.009582346305251122,
0.01647775247693062,
-0.004616260528564453,
0.10753124207258224,
0.029552031308412552,
-0.028546154499053955,
-0.02319924719631672,
-0.058695241808891296,
-0.023326372727751732,
0.10852549225091934,
-0.04710588604211807,
0.03012765757739544,
0.034442611038684845,
-0.05669564753770828,
-0.03875293210148811,
-0.06318362057209015,
0.017191123217344284,
-0.10253974050283432,
0.046970006078481674,
-0.032812803983688354,
-0.048474542796611786,
0.04287848621606827,
0.14169497787952423,
0.02151348628103733,
0.034161314368247986,
-0.03840818256139755,
0.056014399975538254,
0.08691516518592834,
-0.016923323273658752,
-0.016393672674894333,
-0.008485558442771435,
-0.01989581808447838
] |
From Artificial Needles to Real Haystacks: Improving Retrieval Capabilities in LLMs by Finetuning on Synthetic Data | https://openreview.net/forum?id=8m7p4k6Zeb | [
"Zheyang Xiong",
"Vasilis Papageorgiou",
"Kangwook Lee",
"Dimitris Papailiopoulos"
] | Poster | Recent studies have shown that Large Language Models (LLMs) struggle to accurately retrieve information and maintain reasoning capabilities when processing long-context inputs. To address these limitations, we propose a finetuning approach utilizing a carefully designed synthetic dataset comprising numerical key-value retrieval tasks. Our experiments on models like GPT-3.5 Turbo and Mistral 7B demonstrate that finetuning LLMs on this dataset significantly improves LLMs' information retrieval and reasoning capabilities in longer-context settings. We present an analysis of the finetuned models, illustrating the transfer of skills from synthetic to real task evaluations (e.g., $10.5\%$ improvement on $20$ documents MDQA at position $10$ for GPT-3.5 Turbo). We also find that finetuned LLMs' performance on general benchmarks remains almost constant while LLMs finetuned on other baseline long-context augmentation data can encourage hallucination (e.g., on TriviaQA, Mistral 7B finetuned on our synthetic data cause no performance drop while other baseline data can cause a drop that ranges from $2.33\%$ to $6.19\%$). Our study highlights the potential of finetuning on synthetic data for improving the performance of LLMs on longer-context tasks. | Synthetic Data, LLM finetuning, Long Context, Retrieval | null | 615 | 2406.19292 | [
-0.043613024055957794,
-0.0721491351723671,
-0.0016942860092967749,
0.034751396626234055,
0.05503740534186363,
-0.02774546854197979,
0.045477572828531265,
0.04269761964678764,
-0.04197061061859131,
-0.05715272203087807,
-0.09151706099510193,
0.007603974547237158,
0.04339735954999924,
0.022917727008461952,
0.03305230289697647,
-0.0242482777684927,
0.11446797102689743,
0.0762133076786995,
-0.12147128582000732,
-0.06994237005710602,
0.010609243996441364,
-0.019838042557239532,
0.07719089835882187,
-0.03377911448478699,
0.04159159958362579,
-0.0460229329764843,
-0.0396428145468235,
-0.032265786081552505,
0.08142875134944916,
-0.041291382163763046,
0.013733774423599243,
0.1057945042848587,
-0.045979686081409454,
0.049121249467134476,
-0.06615476310253143,
0.07280344516038895,
-0.07684571295976639,
0.04779180511832237,
0.005578276235610247,
-0.0685044676065445,
-0.03577180206775665,
-0.03852565586566925,
-0.0047047617845237255,
0.014366761781275272,
0.06878901273012161,
-0.009103117510676384,
0.020064394921064377,
0.007174991071224213,
-0.004981813486665487,
0.0331965796649456,
-0.14514875411987305,
0.007293528411537409,
0.006266482174396515,
0.0273173488676548,
-0.03905056416988373,
0.022966930642724037,
0.035492926836013794,
0.032182078808546066,
-0.042366962879896164,
-0.05608106404542923,
-0.056299008429050446,
-0.10063991695642471,
-0.054350126534700394,
-0.03616832196712494,
-0.0455428808927536,
-0.00765607412904501,
-0.03896699845790863,
0.05855202674865723,
0.007956860587000847,
-0.014890322461724281,
0.028822004795074463,
0.04625232145190239,
-0.06819155812263489,
0.049669716507196426,
0.005290944594889879,
0.017849698662757874,
-0.031099185347557068,
-0.03718479350209236,
0.06691426038742065,
-0.020679257810115814,
0.031090890988707542,
0.007902001962065697,
0.029780441895127296,
0.014285729266703129,
0.06279589235782623,
-0.04633897915482521,
0.01402528490871191,
0.05039548873901367,
0.023431353271007538,
0.02447853982448578,
0.043197426944971085,
-0.07075603306293488,
0.011353343725204468,
0.011477909050881863,
0.10112246125936508,
0.0470123328268528,
0.062258388847112656,
-0.03710673376917839,
-0.006292782723903656,
0.06878320127725601,
0.022921711206436157,
0.07933970540761948,
-0.07427557557821274,
-0.06840454787015915,
-0.01663094013929367,
0.006270002573728561,
0.03480454161763191,
0.09349171072244644,
0.017396917566657066,
-0.13049782812595367,
0.0088603924959898,
0.0890868678689003,
0.015121597796678543,
-0.028848566114902496,
-0.028862599283456802,
-0.002451599109917879,
-0.03949245437979698,
-0.051693808287382126,
-0.00538799911737442,
0.03712358698248863,
-0.015070565976202488,
0.05229782685637474,
0.017255330458283424,
-0.00916670635342598,
-0.008417907170951366,
-0.02520037069916725,
-0.08401226252317429,
4.303964566992324e-33,
0.04758390039205551,
0.06747236102819443,
-0.03267113119363785,
0.028834493830800056,
0.008696505799889565,
0.011007699184119701,
0.00551204290241003,
0.0758373886346817,
-0.029120909050107002,
0.017973799258470535,
-0.0480458065867424,
0.04911624640226364,
-0.10274157673120499,
0.007563445251435041,
0.05647251009941101,
0.012226172722876072,
-0.0951143205165863,
0.02441246435046196,
-0.08718898147344589,
-0.004713531117886305,
0.06408338248729706,
-0.03167711943387985,
0.024605853483080864,
-0.04103948920965195,
0.04509243369102478,
0.07986866682767868,
0.042571719735860825,
-0.0493316575884819,
-0.027695003896951675,
0.008251434192061424,
-0.106859490275383,
-0.020254557952284813,
-0.0058852070942521095,
0.06785595417022705,
0.03938686475157738,
-0.015370223671197891,
-0.014172568917274475,
-0.08071162551641464,
0.07526453584432602,
0.055364686995744705,
0.0008660639869049191,
0.018863560631871223,
0.09073108434677124,
0.011067312210798264,
-0.08443903923034668,
-0.03861675411462784,
-0.01253802515566349,
0.03767548128962517,
0.000732871238142252,
0.03735518082976341,
0.05280051752924919,
0.0631079375743866,
-0.07349471002817154,
-0.05437171086668968,
-0.044649068266153336,
0.023181354627013206,
0.011473708786070347,
0.05441169813275337,
-0.0026061483658850193,
0.07600606977939606,
0.03263068571686745,
0.024062419310212135,
0.009184329770505428,
0.08989770710468292,
-0.031186390668153763,
0.03206015005707741,
-0.036056745797395706,
0.08600860089063644,
0.10847657173871994,
-0.018727997317910194,
-0.05019737407565117,
-0.03202182054519653,
0.015311617404222488,
-0.017484107986092567,
0.07478424161672592,
-0.01979226805269718,
0.03843594342470169,
-0.08449089527130127,
0.021174777299165726,
-0.03047955594956875,
0.08010219037532806,
0.04753383994102478,
-0.01235996838659048,
-0.039572812616825104,
-0.012542525306344032,
-0.03776849806308746,
-0.03892853856086731,
-0.15060143172740936,
0.02551858499646187,
-0.023303719237446785,
-0.032299771904945374,
-0.017473863437771797,
-0.11572679132223129,
-0.03119238279759884,
-0.05990578606724739,
-3.7966443477224564e-33,
0.0017666174098849297,
-0.06913910806179047,
-0.011414440348744392,
0.11588562279939651,
-0.004111737012863159,
-0.029101183637976646,
0.026824522763490677,
0.01879647932946682,
-0.05539660528302193,
-0.04111166670918465,
-0.035733502358198166,
-0.03889990225434303,
0.029637763276696205,
0.0027609223034232855,
0.003449446754530072,
-0.06103208288550377,
-0.02711416967213154,
-0.021869443356990814,
-0.0427236407995224,
0.04963811859488487,
0.004359617363661528,
0.058290377259254456,
-0.12320642173290253,
0.002060337457805872,
-0.05355318263173103,
0.05232976749539375,
-0.10509226471185684,
-0.011404798366129398,
0.04325416311621666,
0.06416021287441254,
-0.005388670600950718,
-0.007770817261189222,
-0.03831235319375992,
-0.018913554027676582,
-0.038467198610305786,
0.02185545489192009,
0.050195805728435516,
-0.03989620879292488,
-0.08574949949979782,
0.10121322423219681,
0.06480496376752853,
-0.005736089311540127,
-0.03731236979365349,
0.03441261127591133,
0.0004140694800298661,
0.007395761553198099,
-0.14562854170799255,
-0.052780065685510635,
0.06544079631567001,
0.0032071005553007126,
0.08162342011928558,
-0.022156599909067154,
-0.025771642103791237,
0.02732393518090248,
-0.005626431666314602,
-0.09171222895383835,
-0.05003915727138519,
-0.09715808928012848,
-0.02736607752740383,
0.007044552359730005,
-0.05429268255829811,
-0.00929158367216587,
0.02838425524532795,
-0.053983304649591446,
0.02654133550822735,
-0.03848082572221756,
-0.04468831419944763,
-0.03421090170741081,
-0.03966551646590233,
-0.004329529125243425,
0.02584274858236313,
-0.09191915392875671,
0.01874610222876072,
-0.01680411584675312,
0.011168176308274269,
0.05630508437752724,
-0.017445595934987068,
-0.001528294524177909,
-0.003207120345905423,
-0.02454906515777111,
-0.02669786661863327,
-0.0005410099402070045,
0.014670460484921932,
0.046845052391290665,
0.0017325265798717737,
0.07483458518981934,
0.020765885710716248,
0.11377716809511185,
0.014361524023115635,
0.00310608115978539,
-0.03256954997777939,
0.03223658725619316,
-0.02803177759051323,
0.055193983018398285,
-0.03761332482099533,
-4.891715121857487e-8,
-0.047327157109975815,
0.06285330653190613,
-0.06255888193845749,
0.07755310833454132,
0.0013273474760353565,
-0.02027755416929722,
-0.080869160592556,
0.14651164412498474,
-0.034617822617292404,
0.01163031067699194,
0.07382646948099136,
-0.05954539775848389,
-0.059408269822597504,
-0.07027419656515121,
0.08429469913244247,
0.07525797933340073,
0.05693019926548004,
0.016064146533608437,
-0.025698108598589897,
0.005572410766035318,
0.07310504466295242,
0.08724726736545563,
-0.019657352939248085,
0.01794619858264923,
0.028294488787651062,
-0.037044983357191086,
-0.049749378114938736,
0.139470636844635,
0.09242033213376999,
0.04343155771493912,
0.019116193056106567,
-0.01746128872036934,
-0.00951310619711876,
-0.04907940700650215,
0.04587908834218979,
0.023763351142406464,
-0.012120187282562256,
-0.012543690390884876,
0.037079229950904846,
-0.014510349370539188,
0.06012712046504021,
0.07010998576879501,
-0.06349175423383713,
0.07958415150642395,
0.01660943403840065,
-0.024292241781949997,
-0.05834685266017914,
-0.05390588194131851,
0.03300872817635536,
-0.019578324630856514,
0.028792088851332664,
-0.04221506789326668,
-0.015094848349690437,
0.023931102827191353,
0.035142526030540466,
0.06416485458612442,
-0.017481215298175812,
0.004524810239672661,
-0.003431206103414297,
-0.01637614704668522,
-0.0075791338458657265,
-0.03287043049931526,
-0.0720464438199997,
0.040364667773246765
] |
Enhancing Document Understanding with Group Position Embedding: A Novel Approach to Incorporate Layout Information | https://openreview.net/forum?id=Dj9a4zQsSl | [
"Yuke Zhu",
"Yue Zhang",
"Dongdong Liu",
"Chi Xie",
"Zihua Xiong",
"Bo Zheng",
"Sheng Guo"
] | Poster | Recent advancements in document understanding have been dominated by leveraging large language models (LLMs) and multimodal large models. However, enabling LLMs to comprehend complex document layouts and structural information often necessitates intricate network modifications or costly pre-training, limiting their practical applicability. In this paper, we introduce Group Position Embedding (GPE), a novel and efficient technique to enhance the layout understanding capabilities of LLMs without architectural changes or additional pre-training. GPE achieves this by strategically grouping the attention heads and feeding each group with distinct positional embeddings, effectively encoding layout information relevant to document comprehension. This simple yet powerful method allows for effective integration of layout information within the existing LLM framework. We evaluate GPE against several competitive baselines across five mainstream document tasks. We also introduce a challenging benchmark called BLADE, specifically designed to assess layout comprehension.
Extensive experiments on both established and BLADE benchmarks confirm the efficacy of GPE in significantly advancing the state-of-the-art in document understanding. Our code is available at https://github.com/antgroup/GroupPositionEmbedding.git | DocAI, LLM, Position Embedding | This paper proposes a layout-aware position embedding that enable LLMs to comprehend complex documents. | 612 | null | [
0.026603590697050095,
-0.0873875543475151,
0.0024152735713869333,
0.019591890275478363,
0.09893271327018738,
0.0020259872544556856,
-0.08739472180604935,
0.0028524205554276705,
0.05579385906457901,
-0.01873520016670227,
-0.05494547262787819,
0.07381463795900345,
0.0634567141532898,
0.045465126633644104,
-0.03367815911769867,
-0.04669870808720589,
0.0659489631652832,
0.0554618276655674,
-0.07120365649461746,
-0.0005089943879283965,
-0.01918424852192402,
-0.007188555784523487,
0.10981966555118561,
-0.08523362874984741,
0.030852628871798515,
0.07532089948654175,
-0.061385978013277054,
-0.05257127806544304,
0.004783147014677525,
-0.0786651149392128,
0.07668384909629822,
0.013412327505648136,
0.0235846396535635,
0.11383862793445587,
-0.04070950672030449,
0.07968377321958542,
-0.0646597295999527,
0.002655484713613987,
0.005032995715737343,
-0.06448154896497726,
-0.027005985379219055,
0.03201347962021828,
-0.007331302855163813,
0.07074322551488876,
0.10633178800344467,
-0.033012762665748596,
-0.02950502745807171,
-0.06789235025644302,
-0.030339671298861504,
0.0387628898024559,
-0.021964741870760918,
-0.03411038964986801,
-0.011644258163869381,
0.014317772351205349,
-0.0182797871530056,
0.07179997116327286,
0.04156571999192238,
-0.0091194286942482,
-0.05051174387335777,
-0.02973937802016735,
0.015078911557793617,
-0.10930680483579636,
0.02273211069405079,
-0.01291149202734232,
-0.0396183580160141,
-0.028675109148025513,
0.05705336853861809,
0.048375196754932404,
-0.058604273945093155,
0.010058993473649025,
0.0047958060167729855,
0.052704934030771255,
-0.004359790123999119,
-0.0038943905383348465,
0.032633181661367416,
0.07024085521697998,
0.023525850847363472,
-0.0706993117928505,
0.05779053270816803,
-0.07931689918041229,
0.019292132928967476,
0.0032962297555059195,
0.1143658235669136,
0.049287598580121994,
0.08989265561103821,
-0.027879968285560608,
0.049265213310718536,
0.04838666319847107,
-0.05136916786432266,
0.03321265056729317,
-0.04951260983943939,
-0.05248298496007919,
0.0580657422542572,
-0.005559538956731558,
0.0420181080698967,
0.08716768771409988,
0.0530107244849205,
-0.08371122181415558,
-0.08977330476045609,
0.06776637583971024,
0.0036983813624829054,
0.055290840566158295,
-0.00636796560138464,
-0.06648117303848267,
-0.041314881294965744,
0.010456843301653862,
-0.03360147401690483,
0.02180570363998413,
-0.0814475491642952,
-0.0877612829208374,
0.042394861578941345,
0.028946222737431526,
-0.038823261857032776,
0.01646404340863228,
-0.019779207184910774,
-0.0805080384016037,
0.08737252652645111,
-0.013995631597936153,
0.08003847301006317,
0.03279956430196762,
0.003917219582945108,
0.055618464946746826,
-0.00014012621250003576,
-0.03298259526491165,
0.0009303570841439068,
-0.0036915778182446957,
-0.017918173223733902,
1.4327346672749405e-33,
0.055763762444257736,
-0.006563562899827957,
-0.04247196018695831,
0.07813062518835068,
0.12018657475709915,
0.01341949962079525,
-0.057540759444236755,
0.10285384207963943,
-0.0757976844906807,
-0.027129877358675003,
0.010829528793692589,
-0.03361523896455765,
-0.020532455295324326,
0.10017387568950653,
-0.02181115746498108,
-0.05723707750439644,
-0.10193299502134323,
0.033322352916002274,
-0.1072356179356575,
-0.011416672728955746,
-0.050708476454019547,
-0.024998420849442482,
-0.006496333982795477,
0.017521997913718224,
-0.016386019065976143,
0.06410730630159378,
0.06904008239507675,
-0.1127467229962349,
-0.11503814905881882,
-0.01847013644874096,
-0.13411366939544678,
-0.08788163214921951,
-0.01315582636743784,
-0.02480759285390377,
-0.01111032348126173,
-0.045132920145988464,
0.0010441559134051204,
-0.06762710958719254,
0.05755896866321564,
0.010249871760606766,
0.00021385177387855947,
-0.03763081505894661,
0.04261760413646698,
-0.051331013441085815,
-0.10579842329025269,
0.027013985440135002,
-0.017462443560361862,
0.07938360422849655,
0.009651163592934608,
-0.01033084187656641,
0.02766045182943344,
0.04231877624988556,
-0.011014234274625778,
-0.010326586663722992,
0.0373569093644619,
-0.0746922492980957,
0.03366091102361679,
0.025933636352419853,
0.06749255210161209,
-0.02944653481245041,
0.07878393679857254,
0.08256014436483383,
-0.05546770989894867,
0.08960629999637604,
0.002283256035298109,
-0.005089444573968649,
-0.02162361890077591,
0.08774320036172867,
0.02516593597829342,
0.030772291123867035,
-0.030700191855430603,
0.016020704060792923,
0.027662258595228195,
-0.02216622419655323,
-0.02079310081899166,
0.05259660631418228,
-0.005946455989032984,
-0.09876136481761932,
0.015697145834565163,
-0.0003814358206000179,
-0.00923160370439291,
-0.0018448695773258805,
0.09762666374444962,
-0.10162688791751862,
-0.11006821691989899,
-0.0716114267706871,
0.09287802129983902,
-0.1237933486700058,
0.023127393797039986,
0.019244978204369545,
-0.024465886875987053,
-0.06979969143867493,
0.013730202801525593,
-0.017413418740034103,
0.056959740817546844,
-1.045628490008326e-33,
0.012876641936600208,
-0.0019438952440395951,
0.03037281520664692,
-0.009531703777611256,
0.009223533794283867,
0.01744997128844261,
-0.026710599660873413,
0.01685301773250103,
-0.00630448292940855,
-0.05377382040023804,
-0.046245962381362915,
0.02388189546763897,
0.01486019603908062,
0.0022048847749829292,
0.06697038561105728,
0.003467209404334426,
0.004488533828407526,
0.0006835219683125615,
0.005456971470266581,
0.008429237641394138,
-0.036305900663137436,
0.03781179338693619,
-0.009614899754524231,
0.07161823660135269,
0.003168174298480153,
0.04982965812087059,
0.004228879231959581,
-0.009734539315104485,
-0.08670471608638763,
-0.03728330135345459,
-0.02032579481601715,
-0.029025578871369362,
-0.020750325173139572,
0.02147398702800274,
-0.05551585555076599,
0.10755547136068344,
-0.01732935570180416,
0.003523202147334814,
-0.039978355169296265,
0.07739613205194473,
0.04797659441828728,
-0.03172028437256813,
-0.029158486053347588,
0.01743241958320141,
-0.046791836619377136,
0.0622766837477684,
-0.05754119157791138,
-0.0017730490071699023,
0.01706724800169468,
0.045656658709049225,
-0.045905496925115585,
0.08823695778846741,
-0.051767390221357346,
-0.06184590980410576,
-0.049443479627370834,
-0.014136487618088722,
-0.07568869739770889,
-0.06914272159337997,
-0.013086974620819092,
-0.03459206968545914,
0.017578575760126114,
-0.013468993827700615,
-0.009684471413493156,
-0.03576858341693878,
-0.021222293376922607,
-0.05389401316642761,
-0.03519877418875694,
-0.034614305943250656,
-0.028520939871668816,
-0.01612056791782379,
0.023754410445690155,
-0.08303969353437424,
-0.019431190565228462,
0.010818907991051674,
0.03704944998025894,
0.04900224879384041,
0.05535266920924187,
-0.03260032460093498,
-0.01335542369633913,
-0.09187029302120209,
0.023657435551285744,
0.0002832627505995333,
-0.025102920830249786,
0.02853137068450451,
0.0895073413848877,
0.07254626601934433,
0.014858633279800415,
0.09648772329092026,
0.04645722359418869,
0.0299379862844944,
-0.058064911514520645,
0.025463860481977463,
0.021480992436408997,
0.06349866092205048,
-0.011682240292429924,
-3.810573545592888e-8,
-0.0862990990281105,
0.0015159156173467636,
-0.040361437946558,
0.00961257517337799,
-0.03092685155570507,
-0.05205203592777252,
0.012783175334334373,
0.10469157248735428,
-0.03533073887228966,
0.0011859958758577704,
0.06091023609042168,
-0.03862302750349045,
-0.09882305562496185,
-0.061397600919008255,
0.007415484171360731,
0.09822818636894226,
-0.011897217482328415,
0.025278819724917412,
-0.02827487327158451,
-0.04397403076291084,
0.03366360068321228,
0.0867801085114479,
0.027361605316400528,
-0.011036409065127373,
-0.020141838118433952,
-0.0016390071250498295,
-0.022968001663684845,
0.09743660688400269,
0.0722716897726059,
-0.061763517558574677,
0.00041339005110785365,
0.0005287340609356761,
-0.011213724501430988,
0.022328374907374382,
0.02075125463306904,
0.04969675838947296,
0.06354576349258423,
0.06865577399730682,
0.02836369350552559,
-0.03597009554505348,
0.05738786607980728,
-0.05215558409690857,
0.01811613328754902,
0.09038719534873962,
0.012039626948535442,
-0.0020161850843578577,
-0.05734962970018387,
-0.03242241218686104,
0.047761689871549606,
-0.09659150242805481,
-0.019177276641130447,
-0.02991357073187828,
-0.005226591601967812,
0.04421494901180267,
-0.02760779857635498,
0.030854154378175735,
0.00799567624926567,
0.07208231091499329,
0.042421773076057434,
0.014892671257257462,
-0.03047112189233303,
0.06769611686468124,
-0.025918900966644287,
0.014080097898840904
] |
Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology | https://openreview.net/forum?id=rUvCIvI4eB | [
"Xiangyu Wang",
"Donglin Yang",
"Ziqin Wang",
"Hohin Kwan",
"Jinyu Chen",
"Wenjun Wu",
"Hongsheng Li",
"Yue Liao",
"Si Liu"
] | Poster | Developing agents capable of navigating to a target location based on language instructions and visual information, known as vision-language navigation (VLN), has attracted widespread interest. Most research has focused on ground-based agents, while UAV-based VLN remains relatively underexplored. Recent efforts in UAV vision-language navigation predominantly adopt ground-based VLN settings, relying on predefined discrete action spaces and neglecting the inherent disparities in agent movement dynamics and the complexity of navigation tasks between ground and aerial environments. To address these disparities and challenges, we propose solutions from three perspectives: platform, benchmark, and methodology. To enable realistic UAV trajectory simulation in VLN tasks, we propose the OpenUAV platform, which features diverse environments, realistic flight control, and extensive algorithmic support. We further construct a target-oriented VLN dataset consisting of approximately 12k trajectories on this platform, serving as the first dataset specifically designed for realistic UAV VLN tasks. To tackle the challenges posed by complex aerial environments, we propose an assistant-guided UAV object search benchmark called UAV-Need-Help, which provides varying levels of guidance information to help UAVs better accomplish realistic VLN tasks. We also propose a UAV navigation LLM that, given multi-view images, task descriptions, and assistant instructions, leverages the multimodal understanding capabilities of the MLLM to jointly process visual and textual information, and performs hierarchical trajectory generation. The evaluation results of our method significantly outperform the baseline models, while there remains a considerable gap between our results and those achieved by human operators, underscoring the challenge presented by the UAV-Need-Help task. | Unmanned Aerial Vehicle, Drone, Vision-Language Navigation | We propose a UAV simulation platform, an assistant-guided realistic UAV VLN benchmark, and an MLLM-based method to address the challenges in realistic UAV vision-language navigation. | 611 | 2410.07087 | [
0.0667024478316307,
-0.05263620987534523,
-0.03905392810702324,
-0.045914411544799805,
0.02650243230164051,
0.025686223059892654,
-0.02828725054860115,
-0.04149132966995239,
-0.024438222870230675,
-0.010845240205526352,
-0.009857257828116417,
-0.0825023427605629,
-0.02196565456688404,
0.07005751132965088,
0.04296383634209633,
0.010160399600863457,
0.0694437101483345,
0.06821020692586899,
-0.08743633329868317,
0.0002977451658807695,
0.041080541908741,
0.06664945185184479,
0.05029553920030594,
-0.014739715494215488,
-0.011840218678116798,
0.05203694850206375,
-0.000717261282261461,
-0.10848309844732285,
-0.041498105973005295,
-0.07094478607177734,
0.033227596431970596,
-0.00003749822644749656,
-0.0056596421636641026,
0.014052997343242168,
-0.03919477015733719,
0.07193402945995331,
-0.02503785490989685,
-0.0375245101749897,
-0.0451400950551033,
-0.015383667312562466,
-0.0357629619538784,
0.07835225015878677,
0.06761491298675537,
0.015209449455142021,
0.06138341501355171,
0.06747730076313019,
0.06871852278709412,
-0.0055139511823654175,
0.09089060127735138,
0.030789341777563095,
-0.12154995650053024,
-0.08105584979057312,
0.05422980710864067,
-0.031217578798532486,
-0.028841862455010414,
0.019226929172873497,
-0.006470385007560253,
-0.006902881897985935,
0.07843314111232758,
-0.05477038025856018,
0.037106551229953766,
0.0019239626126363873,
0.03357362002134323,
-0.002596682868897915,
-0.09753747284412384,
-0.04376528412103653,
0.06395161896944046,
-0.04770737886428833,
0.04912426322698593,
-0.0259157195687294,
-0.0643487200140953,
0.05294106900691986,
0.10082866251468658,
0.027280887588858604,
-0.06024717167019844,
0.03575877472758293,
0.05794727802276611,
-0.03261549770832062,
0.0009907748317345977,
-0.08683411777019501,
0.08907102048397064,
0.0029070768505334854,
-0.07770617306232452,
0.08505332469940186,
0.11917687207460403,
0.006220936309546232,
-0.030660098418593407,
0.0743275135755539,
0.06493251770734787,
0.013505973853170872,
-0.01875494047999382,
-0.09261301904916763,
0.05185454338788986,
-0.041029687970876694,
0.03708980977535248,
-0.017431247979402542,
-0.01832733303308487,
-0.1343597173690796,
-0.01426792424172163,
0.029786303639411926,
0.005329701583832502,
-0.06318903714418411,
0.038961924612522125,
0.05191807821393013,
0.007665636483579874,
0.02250027284026146,
0.08741440623998642,
0.01463522668927908,
-0.007749828044325113,
-0.02999488078057766,
0.023428212851285934,
-0.04254471883177757,
0.0006373634678311646,
-0.003312547691166401,
0.0029357420280575752,
0.03738205134868622,
-0.06737908720970154,
-0.025282206013798714,
-0.03721557557582855,
0.022606171667575836,
-0.04542560875415802,
-0.015923039987683296,
0.0952288955450058,
-0.025815516710281372,
0.058258429169654846,
0.007132627069950104,
0.01531002763658762,
-2.0163908927564823e-34,
-0.0035144106950610876,
0.038639772683382034,
-0.009049915708601475,
0.012455889023840427,
0.06781435012817383,
-0.04469460994005203,
0.03304678201675415,
0.007478289771825075,
-0.02426745370030403,
0.04457857459783554,
-0.013693022541701794,
-0.035786066204309464,
-0.012638501822948456,
0.05148860812187195,
0.08459645509719849,
-0.017467457801103592,
-0.02445569448173046,
-0.03945259004831314,
-0.0938287228345871,
0.03671778738498688,
0.03772605583071709,
-0.046225544065237045,
-0.05049121752381325,
-0.0471811406314373,
0.05798608064651489,
0.004786756820976734,
-0.02214031107723713,
-0.040088072419166565,
-0.0020555383525788784,
0.011451217345893383,
-0.06469976902008057,
0.06036481633782387,
-0.02877623215317726,
0.036797087639570236,
0.019281573593616486,
-0.0019915925804525614,
-0.07110569626092911,
-0.07845361530780792,
-0.04075058922171593,
0.05642818659543991,
-0.04590427875518799,
0.07820189744234085,
-0.002974896226078272,
0.0037206048145890236,
-0.059701550751924515,
-0.05929849296808243,
-0.0040431395173072815,
0.06299956142902374,
-0.06275034695863724,
0.008910611271858215,
-0.005743944086134434,
0.05961243063211441,
-0.018266476690769196,
-0.07793105393648148,
-0.0181561391800642,
-0.058191124349832535,
0.02683146297931671,
0.04255443811416626,
-0.006678071804344654,
0.07582700997591019,
0.010269095189869404,
-0.0398486852645874,
0.014093330129981041,
0.07260715216398239,
0.0030539531726390123,
-0.015965985134243965,
0.005486040376126766,
0.08206751942634583,
0.08447618037462234,
0.0014506609877571464,
-0.016982391476631165,
0.06033888831734657,
0.03395594283938408,
-0.07333570718765259,
0.04882316663861275,
0.0342973917722702,
0.025228578597307205,
-0.1388888955116272,
0.05898496136069298,
0.02472720481455326,
-0.019938290119171143,
-0.002146261977031827,
0.04259270802140236,
0.02631542645394802,
0.022537732496857643,
-0.030535731464624405,
-0.0014605632750317454,
-0.05809035152196884,
-0.0003431370132602751,
-0.050117045640945435,
-0.08890438079833984,
-0.025823373347520828,
-0.05458124727010727,
-0.03754467889666557,
0.06434530019760132,
-4.025649840881335e-34,
0.025235891342163086,
-0.04548458009958267,
0.007297554984688759,
0.058091092854738235,
-0.0760112926363945,
-0.05840931460261345,
0.07359345257282257,
0.00754587072879076,
0.0004136370262131095,
-0.007684459909796715,
-0.12554584443569183,
0.02091223932802677,
0.06332716345787048,
0.03538477420806885,
0.016558093950152397,
-0.04043218120932579,
0.014348839409649372,
0.022862177342176437,
0.054345980286598206,
-0.01619877852499485,
-0.024308273568749428,
0.0390237420797348,
-0.04787611588835716,
-0.0392441563308239,
0.013136523775756359,
0.02067858912050724,
0.0296246949583292,
0.0049463012255728245,
-0.035856571048498154,
0.0016051032580435276,
0.03307013958692551,
-0.033791955560445786,
-0.009394645690917969,
-0.05709104239940643,
-0.053128719329833984,
0.04874648526310921,
0.04464282467961311,
0.009998273104429245,
-0.06472592800855637,
0.07758834213018417,
0.0975780263543129,
-0.03739781305193901,
-0.01294493768364191,
-0.03675643354654312,
-0.04420400410890579,
0.0833621472120285,
-0.02491648495197296,
0.07321060448884964,
0.0023369009140878916,
-0.011650017462670803,
-0.027443690225481987,
0.004186105914413929,
-0.06958898901939392,
0.03478062152862549,
0.01851477287709713,
-0.040671493858098984,
-0.06317436695098877,
0.0278087817132473,
0.07437384873628616,
-0.0516006238758564,
-0.026951001957058907,
0.022484609857201576,
-0.09501371532678604,
0.04822919890284538,
-0.009579315781593323,
0.008726065047085285,
-0.023223089054226875,
-0.008533861488103867,
-0.012971486896276474,
0.0018160255858674645,
0.0010024167131632566,
-0.0954623892903328,
0.013648612424731255,
0.041606493294239044,
-0.050111379474401474,
-0.018737604841589928,
0.04596830531954765,
-0.031218964606523514,
0.07810473442077637,
-0.0030807259026914835,
0.09410704672336578,
-0.03692462295293808,
-0.0068467119708657265,
0.06310968846082687,
0.010742957703769207,
0.17471463978290558,
-0.04721207171678543,
0.007681793998926878,
-0.006039534229785204,
0.05956323444843292,
-0.02030646800994873,
-0.032294828444719315,
-0.01834777742624283,
0.04929233342409134,
-0.0933586061000824,
-3.618792376869351e-8,
-0.09151513129472733,
0.07688066363334656,
0.020538663491606712,
-0.033224496990442276,
-0.012093416415154934,
0.07602176815271378,
0.005683504976332188,
0.029940025880932808,
0.05274255946278572,
-0.012216444127261639,
-0.006969086360186338,
-0.007656377274543047,
-0.033941928297281265,
-0.022512713447213173,
-0.029091449454426765,
0.09316706657409668,
-0.026869498193264008,
-0.04173293337225914,
0.016951503232121468,
0.010734831914305687,
-0.03732407093048096,
0.0056817918084561825,
-0.026307880878448486,
0.011630550026893616,
-0.08431081473827362,
0.03841413930058479,
-0.09313508123159409,
0.03979624807834625,
0.10812775045633316,
-0.023542366921901703,
0.0488775260746479,
0.035445522516965866,
0.023110281676054,
-0.03823604807257652,
-0.02926693670451641,
0.010064192116260529,
-0.050878413021564484,
-0.0332200825214386,
0.009065496735274792,
-0.05029499530792236,
0.0456656888127327,
0.057837776839733124,
0.03627239540219307,
-0.007754209917038679,
0.005176806356757879,
0.04012706130743027,
-0.04398578405380249,
-0.14496059715747833,
-0.010839248076081276,
-0.09202370792627335,
0.006133177317678928,
-0.016606900840997696,
-0.04806562140583992,
0.17953968048095703,
0.0405275896191597,
0.059634678065776825,
0.04119258001446724,
-0.18850301206111908,
0.04447189345955849,
0.0799013078212738,
0.030165720731019974,
0.021124862134456635,
-0.07820862531661987,
-0.006502920296043158
] |
Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment | https://openreview.net/forum?id=cJQ1K2fjpD | [
"Chenhang Cui",
"An Zhang",
"Yiyang Zhou",
"Zhaorun Chen",
"Gelei Deng",
"Huaxiu Yao",
"Tat-Seng Chua"
] | Poster | The recent advancements in large language models (LLMs) and pre-trained vision models have accelerated the development of vision-language large models (VLLMs), enhancing the interaction between visual and linguistic modalities. Despite their notable success across various domains, VLLMs face challenges in modality alignment, which can lead to issues like hallucinations and unsafe content generation. Current alignment techniques often rely on coarse feedback and external datasets, limiting scalability and performance. In this paper, we propose FiSAO (Fine-Grained Self-Alignment Optimization), a novel self-alignment method that utilizes the model’s own visual encoder as a fine-grained verifier to improve vision-language alignment without the need for additional data. By leveraging token-level feedback from the vision encoder, FiSAO significantly improves vision-language alignment, even surpassing traditional preference tuning methods that require additional data. Through both theoretical analysis and experimental validation, we demonstrate that FiSAO effectively addresses the misalignment problem in VLLMs, marking the first instance of token-level rewards being applied to such models. Our code is avaliable at \url{https://anonymous.4open.science/r/FISAO-57F0/}. | Large Models; Alignment; Hallucination | A self-alignment method that utilizes a fine-grained verifier to improve vision-language alignment without the need for additional data. | 606 | 2410.14148 | [
-0.021978793665766716,
-0.09606382995843887,
0.017981214448809624,
-0.05259376019239426,
-0.004657735116779804,
0.017796112224459648,
-0.06140471249818802,
0.029563741758465767,
0.06397038698196411,
-0.0476144477725029,
0.02465365268290043,
-0.1471208781003952,
0.020866286009550095,
0.04651649668812752,
-0.004465486388653517,
-0.024683523923158646,
0.0598878338932991,
0.1274823546409607,
-0.08926467597484589,
-0.06673648953437805,
0.02991248294711113,
0.002998818177729845,
0.045392103493213654,
0.03535329923033714,
0.03089355118572712,
0.0861867293715477,
0.0414845272898674,
-0.013697827234864235,
0.0033263061195611954,
-0.1071312427520752,
0.005838302429765463,
0.0034632659517228603,
0.006159553769975901,
0.06225018948316574,
-0.04878461733460426,
0.07195153087377548,
-0.102567657828331,
-0.018049631267786026,
0.00034045844222418964,
-0.1173463985323906,
-0.062952920794487,
-0.006625123787671328,
0.008265385404229164,
-0.0021388663444668055,
0.0426439568400383,
0.03297705948352814,
0.05852971225976944,
0.007154286839067936,
0.00787759106606245,
0.0005390375154092908,
-0.11747603863477707,
0.017795898020267487,
-0.043453365564346313,
0.03887632489204407,
0.011541388928890228,
0.07286430895328522,
0.0991676077246666,
0.013940392062067986,
0.03815881535410881,
0.011533121578395367,
0.030374985188245773,
-0.051301371306180954,
0.04920795559883118,
0.053202614188194275,
-0.03674350678920746,
0.0020450439769774675,
0.017896153032779694,
0.04833149164915085,
-0.06677861511707306,
-0.00027558731380850077,
-0.0019438170129433274,
0.05758465826511383,
0.03121056780219078,
0.074772909283638,
0.026422010734677315,
0.10398945957422256,
0.048707470297813416,
-0.010576378554105759,
0.07366520166397095,
-0.07185478508472443,
0.0074708242900669575,
0.045852672308683395,
0.12818026542663574,
0.02394275926053524,
0.07740061730146408,
-0.0031890994869172573,
-0.030906325206160545,
0.005763852037489414,
0.05504448339343071,
0.0005747879622504115,
-0.06794941425323486,
-0.0970698669552803,
-0.015992049127817154,
-0.04158291220664978,
0.07600916177034378,
0.015553400851786137,
-0.013426687568426132,
-0.07800479233264923,
-0.07950224727392197,
0.015481388196349144,
0.04076823219656944,
0.06539967656135559,
-0.00042966348701156676,
-0.03587537631392479,
0.06857036054134369,
-0.0243508480489254,
0.08048325031995773,
0.043614067137241364,
0.050960179418325424,
-0.065462127327919,
0.08947225660085678,
0.0035736949648708105,
-0.022088464349508286,
0.03634343296289444,
-0.02252572402358055,
0.0008416458149440587,
-0.02110823057591915,
-0.015033296309411526,
0.044868238270282745,
-0.01371221337467432,
-0.029864847660064697,
-0.057006336748600006,
-0.03423970192670822,
-0.013675368390977383,
-0.003376706037670374,
-0.0660669133067131,
-0.014415323734283447,
1.1703849911580415e-33,
-0.017446696758270264,
0.08888037502765656,
0.03839642181992531,
-0.03524453565478325,
0.06896533071994781,
-0.06336354464292526,
-0.01769626885652542,
0.010866967961192131,
-0.0611264705657959,
-0.021023917943239212,
-0.029253408312797546,
-0.044912680983543396,
-0.06197725608944893,
0.12436287850141525,
0.02717488817870617,
-0.050537727773189545,
-0.080861896276474,
0.028103282675147057,
-0.10896064341068268,
0.018350131809711456,
0.038776978850364685,
0.0190726388245821,
-0.01987328752875328,
-0.05228332430124283,
0.0013826516224071383,
0.03422907367348671,
0.06537826359272003,
-0.1073857769370079,
-0.0777212381362915,
-0.007196461316198111,
-0.057943761348724365,
0.02845243364572525,
0.06081853061914444,
-0.029486004263162613,
0.0889713391661644,
-0.02166515216231346,
0.008402676321566105,
-0.046803053468465805,
0.007689106278121471,
-0.014052016660571098,
0.05702613294124603,
0.11172761768102646,
-0.016331611201167107,
-0.019424978643655777,
-0.060804594308137894,
-0.08616865426301956,
0.010513346642255783,
0.09900718182325363,
-0.06236964836716652,
-0.062454935163259506,
0.020864205434918404,
0.033123694360256195,
-0.06566138565540314,
-0.043448228389024734,
-0.04915091395378113,
-0.014995180070400238,
0.018083641305565834,
0.016249878332018852,
0.024992840364575386,
0.013907847926020622,
-0.026521822437644005,
-0.000737649155780673,
0.017225168645381927,
0.04659987613558769,
0.05020584166049957,
-0.07153001427650452,
0.053941503167152405,
0.010786131955683231,
0.004453958012163639,
-0.021284950897097588,
-0.021469440311193466,
0.023679077625274658,
-0.05936656892299652,
-0.043768495321273804,
-0.016447002068161964,
-0.009901982732117176,
0.0670815259218216,
-0.07352891564369202,
-0.012797773815691471,
0.023526858538389206,
0.017305204644799232,
0.04314510524272919,
0.004405634943395853,
-0.047765348106622696,
-0.05384670943021774,
-0.0620475709438324,
0.09195015579462051,
-0.056017983704805374,
-0.01110017765313387,
-0.003611018182709813,
0.05289769917726517,
-0.03490075096487999,
0.03029804676771164,
-0.0008229473605751991,
-0.03332178667187691,
-3.563172481125682e-34,
0.0028644087724387646,
0.016526905819773674,
0.041009578853845596,
0.08171451091766357,
-0.02811460942029953,
-0.06610783189535141,
0.08696464449167252,
0.034589022397994995,
0.007365179713815451,
-0.04264764115214348,
-0.037974968552589417,
0.004778012167662382,
-0.00323862349614501,
-0.005880893208086491,
0.043219193816185,
0.019814880564808846,
-0.00005952925857855007,
0.06194918602705002,
0.019808275625109673,
-0.026970969513058662,
0.01860654726624489,
0.08868568390607834,
-0.12145524471998215,
0.046818770468235016,
-0.037606120109558105,
0.0826621800661087,
-0.04049861058592796,
0.022268280386924744,
-0.008856324478983879,
0.008260427974164486,
0.020005015656352043,
0.040421511977910995,
-0.03845438361167908,
-0.030515125021338463,
-0.02662418968975544,
0.04184017702937126,
-0.016998959705233574,
-0.0696311742067337,
-0.0075546810403466225,
0.07739794999361038,
0.0410439632833004,
0.013437063433229923,
-0.09883683174848557,
-0.028748149052262306,
-0.05990651249885559,
-0.07269030809402466,
-0.031373530626297,
-0.0023025143891572952,
0.01255543902516365,
0.05096293240785599,
-0.06994275003671646,
0.04785602539777756,
-0.10169731825590134,
-0.007061284501105547,
-0.05533350259065628,
-0.049805380403995514,
-0.06705141812562943,
0.014630964025855064,
0.05912945792078972,
-0.011061078868806362,
-0.0181029811501503,
0.010742231272161007,
0.09525353461503983,
-0.08371619880199432,
0.01925995759665966,
-0.018570084124803543,
-0.03525576740503311,
-0.005379764828830957,
0.03751485049724579,
-0.06778720766305923,
0.05918126180768013,
-0.06824984401464462,
-0.024886267259716988,
0.07234703749418259,
0.012613646686077118,
-0.05091925710439682,
0.06946386396884918,
-0.03884163871407509,
0.022979171946644783,
-0.031832505017519,
-0.004096883814781904,
0.006681945640593767,
-0.004660011734813452,
0.11049459874629974,
0.03787316381931305,
0.07293600589036942,
0.09224971383810043,
0.08953594416379929,
0.02044888399541378,
0.04243730381131172,
-0.03207626938819885,
-0.00134561606682837,
0.05016806349158287,
0.08810164034366608,
-0.051355987787246704,
-3.741846299476492e-8,
-0.13261285424232483,
-0.014744359999895096,
0.0039954353123903275,
0.07284710556268692,
-0.09175608307123184,
-0.05551998317241669,
0.042758285999298096,
0.0771661326289177,
-0.017734462395310402,
-0.05987248942255974,
0.02648550644516945,
0.003329069819301367,
-0.06642038375139236,
-0.08416172116994858,
-0.007191442418843508,
0.05668790638446808,
0.016349932178854942,
0.08225705474615097,
-0.024415036663413048,
0.027696961537003517,
-0.032778896391391754,
0.037649910897016525,
0.020877670496702194,
-0.06919297575950623,
-0.0037690780591219664,
0.013318668119609356,
-0.06029024347662926,
0.09407605975866318,
0.04263073205947876,
-0.028436144813895226,
0.011914630420506,
0.051220592111349106,
0.08111757040023804,
-0.05237223580479622,
-0.03917159140110016,
0.03236369788646698,
-0.021386820822954178,
0.0565098375082016,
0.02326342836022377,
-0.03633181378245354,
0.03510419651865959,
0.08591366559267044,
0.01051480695605278,
0.04041776433587074,
-0.01034311018884182,
-0.0232698917388916,
0.05239307880401611,
-0.07754745334386826,
0.017249085009098053,
-0.06450481712818146,
-0.0051935696974396706,
-0.010389878414571285,
0.022861458361148834,
0.05809127539396286,
-0.005870821885764599,
-0.007659775670617819,
0.010570984333753586,
-0.00333694601431489,
0.10839737951755524,
0.08563461154699326,
-0.017728421837091446,
0.011541717685759068,
-0.0426432229578495,
-0.04135027155280113
] |
Minimal Impact ControlNet: Advancing Multi-ControlNet Integration | https://openreview.net/forum?id=rzbSNDXgGD | [
"Shikun Sun",
"Min Zhou",
"Zixuan Wang",
"Xubin Li",
"Tiezheng Ge",
"Zijie Ye",
"Xiaoyu Qin",
"Junliang Xing",
"Bo Zheng",
"Jia Jia"
] | Poster | With the advancement of diffusion models, there is a growing demand for high-quality, controllable image generation, particularly through methods that utilize one or multiple control signals based on ControlNet. However, in current ControlNet training, each control is designed to influence all areas of an image, which can lead to conflicts when different control signals are expected to manage different parts of the image in practical applications. This issue is especially pronounced with edge-type control conditions, where regions lacking boundary information often represent low-frequency signals, referred to as silent control signals. When combining multiple ControlNets, these silent control signals can suppress the generation of textures in related areas, resulting in suboptimal outcomes. To address this problem, we propose Minimal Impact ControlNet. Our approach mitigates conflicts through three key strategies: constructing a balanced dataset, combining and injecting feature signals in a balanced manner, and addressing the asymmetry in the score function’s Jacobian matrix induced by ControlNet. These improvements enhance the compatibility of control signals, allowing for freer and more harmonious generation in areas with silent control signals. | generative models, ControlNet | null | 600 | null | [
0.01670602522790432,
-0.05711441487073898,
0.0654207319021225,
0.026908498257398605,
0.07996267825365067,
0.009216910228133202,
0.009237061254680157,
-0.09511126577854156,
0.0018195462180301547,
-0.05979340523481369,
0.0414576530456543,
-0.025677166879177094,
0.04737146943807602,
-0.003201634157449007,
0.024300966411828995,
0.04619283229112625,
0.05740434303879738,
0.04469743371009827,
-0.12540586292743683,
-0.037665437906980515,
0.024584917351603508,
0.006194750312715769,
0.05244798958301544,
-0.044823627918958664,
-0.02738693729043007,
-0.0007782569155097008,
0.003333751577883959,
-0.02597353607416153,
0.06564810872077942,
-0.08281829208135605,
0.08057934045791626,
0.006356117315590382,
0.046144578605890274,
-0.020975187420845032,
-0.09383773803710938,
0.006483838427811861,
-0.005742828361690044,
0.04254095256328583,
0.008986095897853374,
-0.03055727295577526,
0.01226877048611641,
0.06233055144548416,
0.0086509445682168,
-0.045505110174417496,
0.00908542051911354,
-0.003384718904271722,
0.000533548416569829,
-0.03230101242661476,
0.03366949409246445,
-0.03373507037758827,
0.04405978322029114,
-0.06789430975914001,
-0.026719819754362106,
0.03700491413474083,
0.03348521143198013,
-0.019059397280216217,
0.031086081638932228,
0.018353130668401718,
0.0025537298060953617,
-0.005160633474588394,
-0.06076733395457268,
-0.06429217010736465,
-0.0576702281832695,
-0.023201068863272667,
0.010006298311054707,
0.004893011413514614,
0.04805508255958557,
0.0032267370261251926,
-0.06705383211374283,
-0.0007913021836429834,
-0.03607650101184845,
0.0195385180413723,
-0.038914524018764496,
0.008891621604561806,
0.01700327731668949,
0.09988755732774734,
0.012078188359737396,
0.0020136123057454824,
-0.033505767583847046,
-0.047643858939409256,
0.04208061471581459,
-0.05239638686180115,
-0.008698685094714165,
-0.03296288102865219,
0.02787245623767376,
0.01155133917927742,
-0.02906258963048458,
0.03866519778966904,
0.07298380136489868,
0.03535407409071922,
-0.10555190593004227,
0.03798938915133476,
0.06291969120502472,
-0.011425233446061611,
0.00821487046778202,
0.009090964682400227,
-0.01704007387161255,
-0.11594852805137634,
-0.04366239532828331,
0.1040513813495636,
-0.016545075923204422,
-0.13553224503993988,
-0.03044172003865242,
0.0368187353014946,
0.12755152583122253,
-0.0467606857419014,
0.04197268933057785,
-0.014102956280112267,
0.01119882520288229,
-0.03787409886717796,
0.0011142157018184662,
-0.0005881624529138207,
-0.02257498726248741,
-0.017959274351596832,
-0.002040937077254057,
-0.054968226701021194,
0.060378435999155045,
0.007783723063766956,
0.09013992547988892,
-0.00804918259382248,
-0.014276924543082714,
-0.08956605195999146,
-0.042270563542842865,
0.03639112785458565,
0.04841294139623642,
-0.01754937320947647,
-0.09934277087450027,
1.5554958585996314e-33,
-0.005050486419349909,
-0.030926335602998734,
-0.005769323091953993,
-0.0011143237352371216,
0.10857648402452469,
0.0276506170630455,
0.016204608604311943,
-0.07748854905366898,
-0.013087013736367226,
0.003450193675234914,
-0.04650856554508209,
-0.07262551784515381,
-0.07473486661911011,
0.0881582498550415,
0.07323913276195526,
-0.04678184539079666,
0.062485698610544205,
0.034503694623708725,
0.02697724848985672,
0.0035220531281083822,
-0.04237719252705574,
-0.039074189960956573,
-0.0365326888859272,
0.0002868918818421662,
-0.004578129854053259,
-0.03143823891878128,
0.05109475925564766,
-0.03155376389622688,
0.037997353821992874,
0.00596813578158617,
-0.10569829493761063,
-0.016879433766007423,
0.06236419454216957,
0.04455600306391716,
0.02745702490210533,
0.0042314897291362286,
-0.09506739675998688,
0.011957468464970589,
0.04657886177301407,
0.057996660470962524,
-0.024570707231760025,
0.013516032136976719,
-0.11042284965515137,
-0.02572537586092949,
-0.08176623284816742,
0.007174883037805557,
0.08002421259880066,
0.09808759391307831,
-0.09267934411764145,
-0.015069611370563507,
0.03322521969676018,
0.05997692048549652,
0.008407742716372013,
-0.08027192205190659,
-0.05157947540283203,
-0.03957216814160347,
0.041207149624824524,
-0.06163807958364487,
0.03226475417613983,
0.034073587507009506,
0.0006834253435954452,
0.007970619015395641,
0.00418790802359581,
0.010205637663602829,
0.06145121157169342,
0.0035824147053062916,
0.0273825004696846,
0.0152963325381279,
-0.0169803649187088,
-0.06131839007139206,
-0.020684177055954933,
0.06125304847955704,
-0.05086804926395416,
-0.06862114369869232,
0.07192191481590271,
-0.02295556664466858,
0.007766794878989458,
-0.003835872048512101,
0.04645588621497154,
0.050363991409540176,
-0.05361601710319519,
0.03832361102104187,
-0.05808773264288902,
-0.07771489024162292,
0.008424886502325535,
0.03182933107018471,
-0.012157655321061611,
-0.03431052342057228,
-0.0732257217168808,
-0.07869216054677963,
-0.09508710354566574,
0.0018846961902454495,
-0.0247858427464962,
0.011391360312700272,
-0.01633632369339466,
-2.4965437386221678e-33,
-0.04264497011899948,
0.13519254326820374,
-0.019903132691979408,
0.010633910074830055,
-0.05359090864658356,
0.01898178458213806,
0.05428357422351837,
-0.0025832687970250845,
-0.018872179090976715,
-0.027515169233083725,
0.05899191275238991,
-0.03976530209183693,
-0.0674465075135231,
-0.10292106866836548,
-0.08716624230146408,
-0.07759510725736618,
-0.03287023305892944,
-0.0077355289831757545,
-0.03892100974917412,
0.009957949630916119,
-0.018352799117565155,
0.05551477149128914,
-0.024633944034576416,
0.06827738136053085,
-0.1484600305557251,
0.002523100469261408,
-0.09369511157274246,
0.10059718042612076,
0.035312436521053314,
0.0003535586583893746,
-0.10024090111255646,
-0.004907926544547081,
0.07686867564916611,
-0.07536198943853378,
-0.020246172323822975,
0.06647952646017075,
0.08902819454669952,
0.028933148831129074,
-0.060203567147254944,
0.10532378405332565,
0.04864741489291191,
-0.0388033464550972,
-0.0820767804980278,
0.10678646713495255,
-0.04817932844161987,
-0.032398272305727005,
-0.028003130108118057,
-0.006463439203798771,
-0.06282173097133636,
0.06402146816253662,
-0.019635528326034546,
0.06937452405691147,
-0.04089251533150673,
0.071173295378685,
0.012887719087302685,
-0.024538835510611534,
-0.010473229922354221,
0.003217487595975399,
0.09553148597478867,
0.047185927629470825,
-0.040586404502391815,
-0.06848429143428802,
-0.0881558358669281,
-0.0698317363858223,
0.022915668785572052,
0.06292640417814255,
0.003323709825053811,
0.017743894830346107,
0.07929305732250214,
0.10129374265670776,
-0.013541760854423046,
0.02146315388381481,
0.06313124299049377,
-0.02304651215672493,
0.020675914362072945,
-0.05777609348297119,
0.007980584166944027,
-0.10872160643339157,
-0.012913119047880173,
0.005800880957394838,
0.009130902588367462,
-0.018461165949702263,
0.014880320057272911,
0.029453013092279434,
0.04210259020328522,
0.07126374542713165,
0.003326292848214507,
-0.05038547143340111,
0.06352833658456802,
0.029025079682469368,
-0.027180440723896027,
0.07637649774551392,
0.01594702899456024,
0.09900271892547607,
-0.009926526807248592,
-5.1611518614436136e-8,
-0.021071504801511765,
0.028792522847652435,
0.009487119503319263,
-0.04555577039718628,
0.00016609269368927926,
-0.04502269625663757,
0.03651375323534012,
-0.0030311860609799623,
0.051745135337114334,
-0.00587456813082099,
0.031177638098597527,
0.013504132628440857,
-0.06620226055383682,
-0.0640048235654831,
-0.03436122089624405,
0.02529503032565117,
-0.01195741631090641,
0.08910390734672546,
0.00975107867270708,
-0.08429528027772903,
0.007865170016884804,
0.0014623462921008468,
0.021565701812505722,
-0.009929810650646687,
0.11390551924705505,
-0.065969318151474,
0.010213055647909641,
0.06917640566825867,
0.0539122000336647,
0.016373256221413612,
0.009044429287314415,
-0.03910878300666809,
0.01769864745438099,
0.031526342034339905,
0.026695093140006065,
0.045305896550416946,
-0.04003777727484703,
-0.0158414114266634,
-0.008742469362914562,
0.013293338939547539,
-0.04817762225866318,
0.07629360258579254,
-0.04740508273243904,
0.018672216683626175,
0.020442744717001915,
0.04461364820599556,
0.07862897217273712,
-0.08642951399087906,
0.04452821984887123,
0.03199475631117821,
0.0665249228477478,
0.05069738253951073,
0.03132101520895958,
0.12902219593524933,
0.0649777352809906,
-0.014246413484215736,
0.06245755776762962,
0.007263947743922472,
0.025564825162291527,
0.11935947090387344,
-0.01547297090291977,
0.019391829147934914,
-0.04873634874820709,
-0.04815332219004631
] |
TASAR: Transfer-based Attack on Skeletal Action Recognition | https://openreview.net/forum?id=I393kV3bz4 | [
"Yunfeng Diao",
"Baiqi Wu",
"Ruixuan Zhang",
"Ajian Liu",
"Xiaoshuai Hao",
"Xingxing Wei",
"Meng Wang",
"He Wang"
] | Poster | Skeletal sequence data, as a widely employed representation of human actions, are crucial in Human Activity Recognition (HAR). Recently, adversarial attacks have been proposed in this area, which exposes potential security concerns, and more importantly provides a good tool for model robustness test. Within this research, transfer-based attack is an important tool as it mimics the real-world scenario where an attacker has no knowledge of the target model, but is under-explored in Skeleton-based HAR (S-HAR). Consequently, existing S-HAR attacks exhibit weak adversarial transferability and the reason remains largely unknown. In this paper, we investigate this phenomenon via the characterization of the loss function. We find that one prominent indicator of poor transferability is the low smoothness of the loss function. Led by this observation, we improve the transferability by properly smoothening the loss when computing the adversarial examples. This leads to the first Transfer-based Attack on Skeletal Action Recognition, TASAR. TASAR explores the smoothened model posterior of pre-trained surrogates, which is achieved by a new post-train Dual Bayesian optimization strategy. Furthermore, unlike existing transfer-based methods which overlook the temporal coherence within sequences, TASAR incorporates motion dynamics into the Bayesian attack, effectively disrupting the spatial-temporal coherence of S-HARs. For exhaustive evaluation, we build the first large-scale robust S-HAR benchmark, comprising 7 S-HAR models, 10 attack methods, 3 S-HAR datasets and 2 defense models. Extensive results demonstrate the superiority of TASAR. Our benchmark enables easy comparisons for future studies, with the code available in the https://github.com/yunfengdiao/Skeleton-Robustness-Benchmark. | Human Activity Recognition, Transfer-based adversarial attack | This work proposes a new transfer-based attack on skeletal action recognition and builds a large-scale robust evaluation benchmark for this task. | 597 | 2409.02483 | [
-0.1550985723733902,
-0.05317842587828636,
0.03493581339716911,
0.011573557741940022,
0.070930615067482,
0.03024430014193058,
0.04970994591712952,
-0.06718580424785614,
0.02039502002298832,
-0.0030161896720528603,
0.008305689319968224,
-0.0375714935362339,
0.007047565188258886,
0.010626135393977165,
-0.1117580384016037,
-0.05358513817191124,
0.08207884430885315,
0.15400677919387817,
-0.04559462517499924,
0.06317464262247086,
-0.00626542791724205,
0.028115758672356606,
0.016716929152607918,
0.011771922931075096,
0.005406072363257408,
-0.06396492570638657,
0.016113484278321266,
-0.0667724683880806,
-0.05007099732756615,
-0.041702888906002045,
0.055723950266838074,
-0.06481122970581055,
0.00413716584444046,
-0.02046889252960682,
-0.1336195319890976,
0.09536044299602509,
-0.0020467666909098625,
-0.00833483412861824,
0.002187300706282258,
-0.03953263536095619,
-0.03407209739089012,
-0.055176910012960434,
-0.012684240937232971,
0.014483648352324963,
-0.025418756529688835,
-0.03125760704278946,
0.05269843712449074,
-0.05164536461234093,
0.039717815816402435,
0.019249429926276207,
-0.01894804835319519,
-0.0068910629488527775,
-0.04828938841819763,
0.04920671507716179,
-0.060128893703222275,
-0.05459369346499443,
0.013642527163028717,
0.017826469615101814,
-0.022046947851777077,
0.0539361909031868,
0.0248488150537014,
-0.05460211634635925,
0.009160718880593777,
-0.033286478370428085,
0.034130219370126724,
-0.0636996179819107,
0.10364347696304321,
0.037753380835056305,
-0.018188482150435448,
0.060798246413469315,
0.028710676357150078,
0.07705118507146835,
-0.06357157975435257,
-0.030632512643933296,
0.011595336720347404,
0.026253975927829742,
0.013865821994841099,
-0.01808217726647854,
0.03096853569149971,
-0.08462641388177872,
0.051339153200387955,
-0.017868435010313988,
0.029660332947969437,
-0.007640970405191183,
0.0956142470240593,
0.10422559082508087,
-0.12536992132663727,
0.036305204033851624,
0.09219958633184433,
0.018653573468327522,
-0.029733192175626755,
0.04819512367248535,
0.03734380006790161,
-0.024812059476971626,
0.06959621608257294,
0.010003740899264812,
-0.020354900509119034,
0.07238949835300446,
-0.01684868521988392,
0.11001261323690414,
-0.02278304472565651,
0.031145624816417694,
0.039697881788015366,
-0.004673470743000507,
0.11291699856519699,
-0.009311864152550697,
0.06901349872350693,
-0.06278305500745773,
0.048050522804260254,
-0.005710376426577568,
0.06086770445108414,
0.008524935692548752,
-0.008699408732354641,
0.017244290560483932,
0.028410116210579872,
-0.01180948968976736,
-0.03089018724858761,
-0.013428212143480778,
-0.0907386764883995,
0.031789250671863556,
0.014208908192813396,
0.019830020144581795,
0.021701911464333534,
0.014496664516627789,
0.029139319434762,
-0.0338897705078125,
-0.07166597247123718,
-1.643613853873147e-33,
0.04587899148464203,
0.02971133030951023,
0.10069417953491211,
-0.018606770783662796,
-0.039270564913749695,
-0.09602513909339905,
-0.03538380563259125,
0.0200409684330225,
-0.02725936472415924,
0.044170062988996506,
-0.06308788061141968,
-0.00710693746805191,
-0.035578567534685135,
0.09095761924982071,
0.07654596120119095,
0.11537433415651321,
-0.008472156710922718,
-0.017122531309723854,
0.04024919494986534,
0.009944166988134384,
0.04916961491107941,
-0.020175108686089516,
-0.03944575414061546,
-0.031643759459257126,
0.014379281550645828,
-0.0412016361951828,
0.01766461879014969,
-0.016220295801758766,
-0.03039085678756237,
0.02723543904721737,
-0.13272733986377716,
-0.02167142927646637,
0.03588215634226799,
0.01897474192082882,
0.07093542069196701,
-0.06128755584359169,
-0.016118552535772324,
-0.032875921577215195,
-0.04045838490128517,
-0.020685316994786263,
0.06971969455480576,
0.037206631153821945,
0.015979750081896782,
-0.003292362205684185,
-0.019168337807059288,
-0.053187333047389984,
0.07196754217147827,
0.0037500488106161356,
-0.0888042002916336,
-0.015237477608025074,
0.003097115783020854,
0.012426392175257206,
0.009412279352545738,
-0.014831372536718845,
-0.10750436037778854,
-0.019603831693530083,
-0.006529590580612421,
0.03221515938639641,
-0.010941917076706886,
0.10611827671527863,
-0.034838370978832245,
0.01855182833969593,
0.03934375196695328,
0.05740814656019211,
-0.0647214725613594,
-0.03965120017528534,
-0.0304256584495306,
-0.002672832692041993,
0.0210264865309,
0.053568001836538315,
-0.02306045964360237,
0.004621540661901236,
-0.005698359105736017,
-0.017039287835359573,
0.012905909679830074,
-0.02319289557635784,
0.04262074455618858,
-0.03259317949414253,
-0.016094103455543518,
-0.0017934105126187205,
-0.009587953798472881,
-0.05160599574446678,
0.039251673966646194,
0.019171323627233505,
-0.08850529789924622,
-0.0031282939016819,
0.04498012363910675,
-0.09965014457702637,
-0.02369556576013565,
0.0632808580994606,
-0.06977821886539459,
-0.01785185933113098,
-0.0915265902876854,
-0.007771489676088095,
0.02709081396460533,
9.155778561760945e-34,
-0.005231924820691347,
0.03781537339091301,
-0.04736042022705078,
0.04565632715821266,
-0.0668197050690651,
-0.01737079955637455,
0.01009995024651289,
0.10556688159704208,
-0.05925481766462326,
0.0077294111251831055,
-0.007755969651043415,
-0.00966616254299879,
0.007594722788780928,
-0.06267895549535751,
0.03747740760445595,
-0.045646075159311295,
0.010918963700532913,
0.058030929416418076,
-0.07780887186527252,
-0.029643472284078598,
0.07769303023815155,
0.12471309304237366,
-0.026707550510764122,
0.06777266412973404,
-0.04562396928668022,
-0.012303873896598816,
-0.019201679155230522,
0.06611495465040207,
0.017775313928723335,
0.05538282170891762,
-0.03960783779621124,
0.0006719918455928564,
-0.036080438643693924,
0.025400258600711823,
-0.08839623630046844,
0.08600297570228577,
0.04786656051874161,
0.06614530086517334,
-0.010664226487278938,
0.10738130658864975,
0.06869462877511978,
0.003132670186460018,
-0.026075609028339386,
0.038920655846595764,
-0.04825550690293312,
-0.06855440139770508,
-0.021172625944018364,
-0.005342739634215832,
0.012041407637298107,
-0.06515002995729446,
0.0751357302069664,
0.009062067605555058,
-0.0070894923992455006,
0.020160647109150887,
0.006399722769856453,
-0.017300887033343315,
0.016199661418795586,
-0.06587778776884079,
0.08246055990457535,
0.017333069816231728,
-0.011727062985301018,
0.038617875427007675,
-0.06871067732572556,
-0.017239544540643692,
0.04645133018493652,
-0.03174172714352608,
-0.020446637645363808,
0.011444609612226486,
-0.0009603602229617536,
0.023452062159776688,
0.022502833977341652,
-0.026179593056440353,
0.0333503857254982,
0.06929630786180496,
-0.03601790592074394,
-0.07385960966348648,
-0.02553953230381012,
-0.07395444810390472,
0.013143758289515972,
-0.0017372516449540854,
0.05640656501054764,
-0.13477571308612823,
0.08311807364225388,
0.024992208927869797,
0.05262618511915207,
0.11784642934799194,
0.04214295744895935,
0.023635033518075943,
-0.04658820480108261,
-0.023823332041502,
-0.02901998721063137,
0.007352668792009354,
-0.02414040081202984,
0.022649748250842094,
-0.04220317676663399,
-4.063878478177685e-8,
-0.06978853791952133,
0.026425497606396675,
0.08677681535482407,
-0.029259534552693367,
-0.019252119585871696,
0.05340323969721794,
-0.07691475749015808,
0.06063873693346977,
0.00579278077930212,
0.00822689849883318,
0.03292324021458626,
-0.006339166313409805,
0.021009277552366257,
-0.08182898908853531,
-0.05464686453342438,
0.06953921169042587,
-0.004730205982923508,
0.03596111759543419,
-0.06914033740758896,
-0.07092829048633575,
-0.02340889908373356,
0.04401664063334465,
-0.03267305716872215,
-0.025055933743715286,
0.03823374584317207,
-0.06042543798685074,
-0.056164126843214035,
0.09057888388633728,
0.05270840600132942,
0.056894466280937195,
-0.033447787165641785,
-0.011949298903346062,
-0.030197929590940475,
0.038614045828580856,
-0.002174904802814126,
0.13520199060440063,
0.021858317777514458,
-0.05335204303264618,
-0.011260807514190674,
-0.009130422957241535,
-0.04900448024272919,
0.04720718786120415,
-0.023849043995141983,
-0.029877211898565292,
-0.03972359374165535,
0.057064130902290344,
0.035680484026670456,
-0.12735489010810852,
0.04841189831495285,
-0.0057105859741568565,
0.1073475107550621,
-0.008433982729911804,
-0.0006946030771359801,
0.10823074728250504,
-0.021402936428785324,
0.027428895235061646,
0.05689229443669319,
-0.050812721252441406,
0.03467182070016861,
0.07115321606397629,
0.09519404917955399,
-0.025974053889513016,
-0.08873851597309113,
-0.0007221712730824947
] |
HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts | https://openreview.net/forum?id=2eFq6S35iB | [
"Hongjun Wang",
"Sagar Vaze",
"Kai Han"
] | Poster | Generalized Category Discovery (GCD) is a challenging task in which, given a partially labelled dataset, models must categorize all unlabelled instances, regardless of whether they come from labelled categories or from new ones. In this paper, we challenge a remaining assumption in this task: that all images share the same domain. Specifically, we introduce a new task and method to handle GCD when the unlabelled data also contains images from different domains to the labelled set. Our proposed `HiLo' networks extract High-level semantic and Low-level domain features, before minimizing the mutual information between the representations. Our intuition is that the clusterings based on domain information and semantic information should be independent. We further extend our method with a specialized domain augmentation tailored for the GCD task, as well as a curriculum learning approach. Finally, we construct a benchmark from corrupted fine-grained datasets as well as a large-scale evaluation on DomainNet with real-world domain shifts, reimplementing a number of GCD baselines in this setting. We demonstrate that HiLo outperforms SoTA category discovery models by a large margin on all evaluations. | Generalized Category Discovery | we extracts domain and semantic information independently and minimize their mutual information while incorporating contrastive learning for robust representations with pseudo-labelling strategies | 592 | 2408.04591 | [
-0.025997981429100037,
-0.08239499479532242,
-0.051028452813625336,
0.04577971249818802,
0.08794884383678436,
-0.04696023091673851,
-0.02566353790462017,
-0.05915429815649986,
-0.026807833462953568,
-0.05824032053351402,
0.0316922701895237,
-0.05466584488749504,
0.03343807905912399,
0.08910144865512848,
0.007067111786454916,
-0.00237662298604846,
0.13953758776187897,
0.045512985438108444,
-0.09422546625137329,
-0.008280974812805653,
0.019658543169498444,
0.09781491756439209,
0.1283726841211319,
0.04396840184926987,
0.016831059008836746,
0.04642799496650696,
0.07665080577135086,
-0.01945514790713787,
0.010288644582033157,
-0.03157704696059227,
-0.001747236936353147,
0.020433200523257256,
-0.08912505209445953,
0.016599029302597046,
0.03507299721240997,
0.10182276368141174,
-0.0002555752289481461,
0.09665822237730026,
-0.013236785307526588,
-0.027094507589936256,
0.00159426499158144,
-0.012377126142382622,
0.07983613014221191,
-0.008154153823852539,
-0.012456337921321392,
0.07182799279689789,
-0.0767086073756218,
-0.045030172914266586,
-0.00383581779897213,
-0.008760383352637291,
0.0024129189550876617,
0.0009068128420040011,
-0.05373609811067581,
0.08326868712902069,
-0.06961356848478317,
0.02756846882402897,
0.02630203776061535,
-0.05939377844333649,
0.014614745043218136,
0.01850672997534275,
-0.06309831887483597,
-0.06200580671429634,
0.017339885234832764,
-0.011007010005414486,
-0.047066353261470795,
-0.017733491957187653,
0.0031264694407582283,
0.01867969147861004,
0.0160421933978796,
-0.013821294531226158,
0.04715574532747269,
0.06132274866104126,
-0.019955193623900414,
0.04256138950586319,
0.053068824112415314,
0.028918856754899025,
0.07290573418140411,
0.02005390077829361,
0.06259163469076157,
-0.09797137975692749,
0.05123581364750862,
0.05793912336230278,
0.07609793543815613,
0.04124386981129646,
0.14644326269626617,
0.025098443031311035,
0.010167638771235943,
-0.0006655527977272868,
-0.05657565966248512,
0.010952789336442947,
-0.04964720085263252,
-0.09918484836816788,
0.10844174027442932,
-0.012581857852637768,
0.020276419818401337,
-0.037342458963394165,
0.02710345946252346,
-0.11219880729913712,
-0.012543190270662308,
0.07515986263751984,
-0.0529618039727211,
0.06198268383741379,
0.0013913862640038133,
-0.047415852546691895,
-0.004886956419795752,
0.031488917768001556,
-0.013429813086986542,
0.05141603574156761,
0.10204585641622543,
-0.07354767620563507,
-0.004375026561319828,
0.05914771184325218,
-0.026852808892726898,
-0.03874712809920311,
-0.00701125618070364,
0.0026221831794828176,
0.06359852850437164,
0.025362389162182808,
0.0025698384270071983,
-0.029521875083446503,
-0.026642393320798874,
-0.007435876876115799,
0.0020858272910118103,
-0.04604678228497505,
0.01992279477417469,
-0.054977599531412125,
-0.11281514167785645,
1.4755906199120287e-33,
0.09104424715042114,
-0.045547597110271454,
0.0727810263633728,
-0.0011980962008237839,
-0.016099976375699043,
-0.05025516450405121,
-0.061174169182777405,
0.01708626002073288,
-0.02428988739848137,
0.05785444378852844,
-0.004027410410344601,
0.06460422277450562,
-0.07157989591360092,
0.08746322244405746,
0.0034243734553456306,
0.0334542840719223,
-0.030940866097807884,
0.05521583557128906,
-0.001306296675466001,
-0.05938573554158211,
-0.06047554314136505,
0.005607482511550188,
0.008212908171117306,
-0.035213351249694824,
-0.011109239421784878,
0.012590189464390278,
0.013127762824296951,
-0.09094633907079697,
0.02627212181687355,
0.01969721168279648,
-0.08577004820108414,
-0.0046598045155406,
0.05379496142268181,
0.09106370061635971,
0.024599673226475716,
-0.011589867062866688,
0.010201320052146912,
-0.06235693767666817,
0.04918584227561951,
-0.015049725770950317,
-0.025403818115592003,
0.02531575970351696,
0.007345454301685095,
-0.060496579855680466,
0.026558605954051018,
0.008126715198159218,
0.01587495394051075,
0.017115509137511253,
-0.034733857959508896,
0.04290149733424187,
0.08830289542675018,
-0.05651769042015076,
-0.02397606335580349,
0.027227938175201416,
-0.12843340635299683,
0.01696786843240261,
0.03609057515859604,
0.00019302104192320257,
0.04115716740489006,
0.016584495082497597,
0.045932456851005554,
0.045080121606588364,
-0.01916811242699623,
0.101638063788414,
-0.055083535611629486,
0.046419233083724976,
0.03545414283871651,
0.055828679352998734,
0.020483726635575294,
0.0717688500881195,
-0.017110638320446014,
0.07558868080377579,
-0.020311560481786728,
-0.05866653844714165,
0.05944841355085373,
0.06983627378940582,
0.05784997344017029,
-0.06790584325790405,
-0.06807560473680496,
0.07302762567996979,
0.05352277681231499,
-0.05819191783666611,
-0.008415241725742817,
-0.03531019762158394,
-0.04920144006609917,
0.004090177360922098,
0.040463972836732864,
-0.07986708730459213,
-0.008589867502450943,
0.005995067302137613,
-0.09141507744789124,
0.029780078679323196,
-0.051611289381980896,
-0.0003497809811960906,
0.07994487881660461,
-3.5650897766967204e-34,
-0.0029765188228338957,
0.015816478058695793,
-0.022963356226682663,
-0.0014393349410966039,
-0.006305206101387739,
-0.06543999165296555,
0.026362840086221695,
0.040020983666181564,
-0.054504476487636566,
-0.08430448919534683,
-0.01966226100921631,
-0.04406299442052841,
-0.08970936387777328,
-0.05064088851213455,
-0.003159251296892762,
-0.029948577284812927,
-0.05192239210009575,
0.06247002258896828,
-0.00014642500900663435,
0.1026286855340004,
0.016506649553775787,
0.07338889688253403,
-0.14335833489894867,
0.09013374149799347,
0.00656320434063673,
0.06968077272176743,
-0.04409470781683922,
0.013257449492812157,
0.004014085046947002,
0.0037750916089862585,
-0.06923505663871765,
0.00453434232622385,
-0.016050783917307854,
-0.07541628926992416,
-0.0438605435192585,
0.037085551768541336,
-0.017186174169182777,
-0.030667895451188087,
-0.10059358924627304,
-0.028902122750878334,
-0.05220598727464676,
-0.015414546243846416,
-0.1150597408413887,
0.05470156669616699,
0.04225651174783707,
-0.05992761626839638,
-0.024406373500823975,
0.015673942863941193,
0.0357925146818161,
0.07102340459823608,
-0.06153087690472603,
-0.03454222157597542,
-0.013506483286619186,
0.03163069859147072,
0.02879367023706436,
-0.06614947319030762,
-0.026688102632761,
-0.024232834577560425,
0.018386775627732277,
0.018584752455353737,
-0.03510364890098572,
0.008674025535583496,
-0.040101923048496246,
0.010200458578765392,
-0.0046912976540625095,
-0.06959463655948639,
0.003873611567541957,
-0.05370551720261574,
-0.02716115117073059,
0.022073136642575264,
0.005307618994265795,
0.0379062257707119,
-0.00041634385706856847,
-0.05494078993797302,
-0.046407122164964676,
-0.037278518080711365,
0.06235550716519356,
-0.044651564210653305,
0.039960406720638275,
0.019443785771727562,
0.0659588873386383,
-0.019268188625574112,
0.022817619144916534,
0.0792820155620575,
0.06831145286560059,
0.06529016047716141,
0.056395288556814194,
0.058632779866456985,
0.03754888474941254,
0.014702407643198967,
-0.03868473321199417,
-0.004435892216861248,
0.030461611226201057,
0.015627650544047356,
-0.039714179933071136,
-4.061555358703117e-8,
-0.08086072653532028,
0.020513197407126427,
-0.07881458848714828,
-0.07002861797809601,
0.015311191789805889,
0.0018340778769925237,
-0.0326833538711071,
0.12869571149349213,
-0.06990011781454086,
0.03207894414663315,
0.027837708592414856,
0.03306906670331955,
-0.04818661883473396,
-0.023134658113121986,
-0.07110048830509186,
0.04008784517645836,
0.07754294574260712,
0.05709419399499893,
0.016764817759394646,
0.0022683830466121435,
0.03257625177502632,
-0.02724512666463852,
0.08046455681324005,
-0.01074873935431242,
0.006452145520597696,
-0.1481265276670456,
0.004252339247614145,
0.054169245064258575,
0.08367994427680969,
-0.022759774699807167,
-0.08407606184482574,
-0.023983769118785858,
0.020659130066633224,
-0.01809946447610855,
0.0338609479367733,
0.06348518282175064,
-0.03071068786084652,
0.007643742486834526,
0.012383034452795982,
-0.0867743045091629,
0.004886055830866098,
0.05555101856589317,
-0.013210845179855824,
0.0007918934570625424,
-0.027581069618463516,
0.018850844353437424,
0.028398791328072548,
0.015967009589076042,
0.019925646483898163,
0.012024237774312496,
-0.08634375035762787,
-0.004756311886012554,
-0.024975527077913284,
0.07960616052150726,
0.0557841993868351,
-0.008879099041223526,
-0.04441923648118973,
-0.06876235455274582,
0.06708068400621414,
0.03312421962618828,
0.02933095581829548,
-0.020701423287391663,
-0.07796239852905273,
0.05595957860350609
] |
TSVD: Bridging Theory and Practice in Continual Learning with Pre-trained Models | https://openreview.net/forum?id=bqv7M0wc4x | [
"Liangzu Peng",
"Juan Elenter",
"Joshua Agterberg",
"Alejandro Ribeiro",
"Rene Vidal"
] | Poster | The goal of continual learning (CL) is to train a model that can solve multiple
tasks presented sequentially. Recent CL approaches have achieved strong performance
by leveraging large pre-trained models that generalize well to downstream
tasks. However, such methods lack theoretical guarantees, making them prone to
unexpected failures. Conversely, principled CL approaches often fail to achieve
competitive performance. In this work, we aim to bridge this gap between theory
and practice by designing a simple CL method that is theoretically sound and
highly performant. Specifically, we lift pre-trained features into a higher dimensional
space and formulate an over-parametrized minimum-norm least-squares
problem. We find that the lifted features are highly ill-conditioned, potentially
leading to large training errors (numerical instability) and increased generalization
errors. We address these challenges by continually truncating the singular value
decomposition (SVD) of the lifted features. Our approach, termed TSVD, is stable
with respect to the choice of hyperparameters, can handle hundreds of tasks, and
outperforms state-of-the-art CL methods on multiple datasets. Importantly, our
method satisfies a recurrence relation throughout its continual learning process,
which allows us to prove it maintains small training and generalization errors by
appropriately truncating a fraction of SVD factors. This results in a stable continual
learning method with strong empirical performance and theoretical guarantees.
Code available: https://github.com/liangzu/tsvd. | Continual Learning; Pretrained Models; Overparameterization; Generalization; Random Feature Models | The paper proposes a simple method that delivers stable and strong performance with theoretical guarantees for continual learning with pre-trained models. | 585 | 2410.00645 | [
-0.04630282521247864,
-0.06563667207956314,
0.07670918852090836,
0.03680634871125221,
0.04897134006023407,
0.028326038271188736,
-0.031068643555045128,
-0.03883519023656845,
-0.027932871133089066,
0.015836959704756737,
-0.03368731588125229,
0.08339080214500427,
-0.012707089073956013,
0.03999050706624985,
-0.06893156468868256,
0.04019315540790558,
0.049212418496608734,
0.04261224716901779,
-0.03144938126206398,
-0.006962418556213379,
-0.004063433036208153,
0.02990833856165409,
-0.04443565011024475,
0.044492918998003006,
0.051052525639534,
-0.03847475349903107,
-0.02124347724020481,
-0.033181410282850266,
0.05695309489965439,
-0.0050065829418599606,
0.009336072951555252,
-0.015028053894639015,
-0.07976718991994858,
-0.03580280393362045,
-0.09428486227989197,
0.07700838893651962,
-0.03583709895610809,
0.057143889367580414,
-0.0660783126950264,
0.005808060988783836,
0.043638259172439575,
-0.03815455734729767,
0.0015779207460582256,
0.026244519278407097,
0.060920704156160355,
0.039995335042476654,
0.013492143712937832,
-0.1007850170135498,
0.055420368909835815,
0.019904926419258118,
-0.03857850283384323,
0.0034150294959545135,
-0.016391610726714134,
0.07729553431272507,
0.00027877528918907046,
-0.06172528862953186,
0.014104777947068214,
-0.01157135609537363,
0.023872006684541702,
-0.010314015671610832,
-0.005136914551258087,
-0.10069984197616577,
0.008807630278170109,
-0.0662958100438118,
-0.013089490123093128,
0.0028290164191275835,
0.028343943879008293,
0.07132810354232788,
-0.014805956743657589,
0.07916917651891708,
-0.03880109637975693,
0.0847100242972374,
-0.05223626643419266,
0.08162549138069153,
0.0013684035511687398,
0.024242449551820755,
0.07091377675533295,
-0.037967294454574585,
0.03225671127438545,
0.015627693384885788,
-0.010362597182393074,
0.041203755885362625,
0.003810526104643941,
-0.10374270379543304,
0.10027752816677094,
0.00025061587803065777,
0.03814934939146042,
-0.03734830394387245,
0.07458079606294632,
0.00926961749792099,
0.03157895803451538,
0.007285481318831444,
0.020611798390746117,
-0.01534949243068695,
-0.02680186741054058,
-0.013504592701792717,
0.030787495896220207,
-0.07461857050657272,
-0.0007555746124126017,
0.10002591460943222,
-0.02243587002158165,
0.0567275695502758,
0.005326543934643269,
-0.026500791311264038,
0.01572209782898426,
-0.06828377395868301,
0.006366265472024679,
0.059780992567539215,
0.059455834329128265,
-0.09521371871232986,
0.04358472675085068,
-0.060139838606119156,
-0.007682885508984327,
0.012101178988814354,
0.05613570660352707,
0.028377458453178406,
0.014357408508658409,
-0.004680270794779062,
0.013540707528591156,
0.06986794620752335,
-0.02674609236419201,
-0.007856748066842556,
0.02459656074643135,
0.07609930634498596,
0.048562224954366684,
0.005955365486443043,
-0.06966546922922134,
2.0717858344502422e-33,
0.0015156902372837067,
0.037713512778282166,
0.04155221953988075,
0.0000658127901260741,
-0.018824318423867226,
-0.0051945885643363,
0.04955136030912399,
0.09023886173963547,
-0.013352124951779842,
0.024893200024962425,
-0.0517711378633976,
-0.02430904470384121,
-0.07966607809066772,
0.05411533638834953,
0.09205757826566696,
-0.042499177157878876,
-0.03386428952217102,
0.04016440734267235,
0.0023777754977345467,
-0.09376981854438782,
0.08774002641439438,
-0.06677397340536118,
-0.00805643480271101,
-0.07205763459205627,
-0.02826322242617607,
-0.009679773822426796,
0.04652263969182968,
-0.013442629016935825,
-0.03841153904795647,
0.021098272874951363,
-0.07653597742319107,
0.047895241528749466,
-0.0017767908284440637,
0.0748976394534111,
0.00825788825750351,
-0.0486711785197258,
0.03950409218668938,
-0.017276886850595474,
0.08004956692457199,
-0.027429752051830292,
-0.0008541172719560564,
-0.003602469339966774,
-0.02410704270005226,
0.013050669804215431,
-0.0385018065571785,
-0.12573285400867462,
0.046582918614149094,
-0.005875785835087299,
-0.08418276160955429,
-0.034740906208753586,
-0.009416390210390091,
0.026983538642525673,
-0.08841025084257126,
-0.05213002860546112,
-0.10289609432220459,
0.004122581332921982,
-0.026680178940296173,
-0.03406102955341339,
0.016924189403653145,
-0.010718705132603645,
0.017899690195918083,
0.013293785974383354,
-0.04294820874929428,
0.06946367770433426,
-0.003600971307605505,
0.02139865793287754,
0.05759793519973755,
0.05841841921210289,
0.028991637751460075,
-0.0221770778298378,
-0.020018232986330986,
-0.08451332151889801,
0.012494085356593132,
-0.12951365113258362,
0.13213396072387695,
-0.011101176962256432,
0.12347380071878433,
-0.09940999001264572,
0.02905164286494255,
-0.013707851059734821,
0.003493081545457244,
0.05402380973100662,
-0.024681171402335167,
-0.07392299920320511,
0.02859634906053543,
-0.04561459645628929,
0.04711534082889557,
-0.11766711622476578,
-0.04852534458041191,
-0.027304071933031082,
-0.10919512063264847,
0.03832676634192467,
-0.032967619597911835,
-0.012864826247096062,
-0.046714719384908676,
-1.2700312799352622e-33,
0.04403627663850784,
0.03584148362278938,
-0.04085354134440422,
0.04634514078497887,
0.02307630330324173,
-0.05682499334216118,
-0.0629933699965477,
-0.017078740522265434,
-0.09879560768604279,
-0.1317702829837799,
0.00981169380247593,
-0.1425294727087021,
-0.039487458765506744,
0.020140385255217552,
0.012983568012714386,
0.05147823318839073,
-0.027645135298371315,
0.0033514571841806173,
-0.0548294261097908,
-0.00322141544893384,
-0.040713511407375336,
0.020754260942339897,
-0.04429777339100838,
0.0887824222445488,
-0.0581536628305912,
0.02530262991786003,
0.03718695789575577,
0.08278341591358185,
0.021228637546300888,
0.020539456978440285,
-0.08842283487319946,
-0.0272371768951416,
-0.03741630166769028,
-0.01039302721619606,
-0.0606299564242363,
0.07820706814527512,
0.034112997353076935,
0.003462212858721614,
0.017902769148349762,
0.12096702307462692,
0.13154584169387817,
-0.05177774652838707,
-0.07352384179830551,
0.020987575873732567,
-0.008398967795073986,
-0.05482151731848717,
-0.012712475843727589,
0.02665911801159382,
0.013410044834017754,
0.007501100655645132,
0.007535297889262438,
0.019138988107442856,
-0.10690904408693314,
0.03597263991832733,
-0.028987327590584755,
-0.025721369311213493,
-0.03926787152886391,
0.0004067412228323519,
0.022614702582359314,
-0.034551896154880524,
-0.02454487420618534,
-0.03607027977705002,
0.03272967040538788,
-0.007456842809915543,
0.06443238258361816,
-0.01078943070024252,
0.05208808556199074,
0.07930710166692734,
-0.05239476263523102,
0.04560882970690727,
-0.07084950804710388,
0.024244818836450577,
0.031613077968358994,
-0.061755161732435226,
-0.02484903670847416,
-0.030712246894836426,
0.006038271822035313,
-0.0706276148557663,
-0.019636699929833412,
-0.014305138029158115,
0.013849712908267975,
0.016641724854707718,
0.04663214460015297,
0.037222668528556824,
0.05101163685321808,
0.14853554964065552,
0.05656428262591362,
0.1059824749827385,
0.061503149569034576,
-0.015244666486978531,
-0.02445237897336483,
0.013052502647042274,
0.0319797620177269,
0.033173806965351105,
0.011441323906183243,
-4.4759030259911015e-8,
-0.08406943827867508,
0.06289851665496826,
0.006712277419865131,
-0.011681957170367241,
0.06975144892930984,
-0.02962949313223362,
-0.011830094270408154,
0.12333008646965027,
-0.060769449919462204,
-0.041833244264125824,
0.03781256079673767,
0.05086519569158554,
-0.025084231048822403,
-0.010896025225520134,
0.0008032185724005103,
0.06190887838602066,
-0.024208538234233856,
0.020050562918186188,
-0.016453079879283905,
-0.006853920873254538,
0.033626217395067215,
-0.011354574002325535,
0.05082244798541069,
-0.03978816047310829,
0.002613865537568927,
-0.08359857648611069,
-0.01600494235754013,
0.1266513615846634,
0.13829109072685242,
0.02338971570134163,
-0.021975846961140633,
0.02410910092294216,
0.026835648342967033,
0.0027874326333403587,
0.05579223111271858,
0.047408584505319595,
0.07524433732032776,
-0.03136391192674637,
-0.07659316807985306,
-0.028683047741651535,
-0.010500977747142315,
0.09403245151042938,
-0.0622011199593544,
0.02028058096766472,
0.028367966413497925,
0.05357581377029419,
-0.009758037514984608,
-0.039184022694826126,
0.007801618427038193,
-0.027391385287046432,
0.023418156430125237,
-0.003846385283395648,
-0.01800336129963398,
0.13451622426509857,
0.05140767619013786,
-0.02593342587351799,
-0.031560592353343964,
-0.02957938238978386,
-0.025877803564071655,
-0.005492235533893108,
0.030675062909722328,
-0.01579606533050537,
-0.04465298354625702,
0.0581139475107193
] |
I Can Hear You: Selective Robust Training for Deepfake Audio Detection | https://openreview.net/forum?id=2GcR9bO620 | [
"Zirui Zhang",
"Wei Hao",
"Aroon Sankoh",
"William Lin",
"Emanuel Mendiola-Ortiz",
"Junfeng Yang",
"Chengzhi Mao"
] | Poster | Recent advances in AI-generated voices have intensified the challenge of detecting deepfake audio, posing risks for scams and the spread of disinformation. To tackle this issue, we establish the largest public voice dataset to date, named DeepFakeVox-HQ, comprising 1.3 million samples, including 270,000 high-quality deepfake samples from 14 diverse sources. Despite previously reported high accuracy, existing deepfake voice detectors struggle with our diversely collected dataset, and their detection success rates drop even further under realistic corruptions and adversarial attacks. We conduct a holistic investigation into factors that enhance model robustness and show that incorporating a diversified set of voice augmentations is beneficial. Moreover, we find that the best detection models often rely on high-frequency features, which are imperceptible to humans and can be easily manipulated by an attacker. To address this, we propose the F-SAT: Frequency-Selective Adversarial Training method focusing on high-frequency components. Empirical results demonstrate that using our training dataset boosts baseline model performance (without robust training) by 33%, and our robust training further improves accuracy by 7.7% on clean samples and by 29.3% on corrupted and attacked samples, over the state-of-the-art RawNet3 model. | Deepfake audio detection, Audio augmentations, Frequency-Selective Adversarial Training | null | 584 | 2411.00121 | [
-0.08904995769262314,
-0.08339129388332367,
-0.021494684740900993,
0.0727013498544693,
0.07996364682912827,
0.06614244729280472,
-0.052086830139160156,
-0.0915154442191124,
-0.01721794530749321,
-0.10115557909011841,
-0.0722385048866272,
-0.08081244677305222,
-0.004029023926705122,
-0.014297145418822765,
-0.02506503090262413,
-0.08073337376117706,
0.12101723998785019,
0.11762568354606628,
-0.08105140179395676,
0.00188286358024925,
-0.00659613823518157,
0.04710013046860695,
0.06771228462457657,
-0.020018326118588448,
0.024298544973134995,
-0.05848400667309761,
-0.03165622055530548,
-0.08239348977804184,
-0.012673682533204556,
-0.06374497711658478,
0.07652026414871216,
0.03284109756350517,
0.0499090813100338,
0.008867381140589714,
-0.021251002326607704,
0.023172900080680847,
0.01993727684020996,
-0.006922121625393629,
0.06132595241069794,
-0.06452367454767227,
0.05167032778263092,
0.027751166373491287,
0.026223525404930115,
-0.017387570813298225,
-0.04060882329940796,
-0.043711934238672256,
0.0009555038996040821,
-0.009206289425492287,
0.05457863584160805,
0.024912068620324135,
-0.01393506396561861,
0.0016766971675679088,
0.06379161775112152,
0.13129965960979462,
-0.060240987688302994,
-0.009378349408507347,
0.08720316737890244,
0.04796880483627319,
0.012373557314276695,
0.030074942857027054,
-0.0127926180139184,
-0.06633393466472626,
-0.033036988228559494,
-0.05137402564287186,
-0.039934080094099045,
-0.0308990441262722,
-0.044540565460920334,
-0.03588996082544327,
-0.01750815287232399,
0.09916980564594269,
0.040404219180345535,
0.09950034320354462,
-0.006658445578068495,
0.015296188183128834,
0.09562228620052338,
0.027432886883616447,
-0.005645617842674255,
-0.041169945150613785,
0.12043935805559158,
-0.004348372109234333,
0.06973181664943695,
-0.03219141438603401,
0.0912548154592514,
-0.033000700175762177,
0.12496679276227951,
-0.0007225820445455611,
-0.04339445382356644,
-0.03181044012308121,
-0.011586753651499748,
0.05330934375524521,
-0.1245718002319336,
-0.007245094049721956,
0.07533243298530579,
0.017110660672187805,
0.03853924944996834,
0.0696413591504097,
-0.04184972494840622,
-0.020590966567397118,
-0.042376890778541565,
0.06535282731056213,
-0.053938329219818115,
-0.016330156475305557,
-0.0009188472758978605,
-0.03468699753284454,
0.009453198872506618,
-0.029815278947353363,
0.014045319519937038,
0.026014771312475204,
0.0037958009634166956,
0.00548629742115736,
0.08452023565769196,
0.023799773305654526,
0.05157582089304924,
-0.04904460161924362,
0.07312566787004471,
0.023107359185814857,
-0.03319285809993744,
-0.00926753506064415,
-0.0019846376962959766,
-0.012533997185528278,
-0.009120051749050617,
0.07851026952266693,
0.013392085209488869,
0.025672079995274544,
0.03849472105503082,
0.004727844148874283,
-0.062260132282972336,
2.3960130740169457e-33,
0.0244898721575737,
0.09478981792926788,
0.06664983183145523,
-0.006989677436649799,
0.0493081733584404,
-0.09104154258966446,
-0.055265624076128006,
0.08836149424314499,
0.014774393290281296,
0.06671102344989777,
-0.03873918578028679,
-0.002985437400639057,
-0.1110573410987854,
0.049503713846206665,
-0.018613923341035843,
0.04538417607545853,
-0.06411328166723251,
0.009896164759993553,
-0.03875754773616791,
-0.012264495715498924,
0.07913953810930252,
-0.03383730724453926,
0.04423041641712189,
-0.00012118295853724703,
0.06995713710784912,
-0.033259835094213486,
0.06539768725633621,
-0.01896112971007824,
0.0022688840981572866,
0.03695126250386238,
-0.11538927257061005,
-0.03666268661618233,
0.06397903710603714,
0.07883454859256744,
-0.009007087908685207,
-0.0029546625446528196,
-0.02429552748799324,
-0.00648336298763752,
-0.0292819831520319,
0.02360827662050724,
-0.010341847315430641,
0.030281875282526016,
-0.01306985504925251,
-0.03365558758378029,
-0.0327233150601387,
-0.029293982312083244,
-0.023013385012745857,
0.0003914125554729253,
-0.04139356687664986,
0.03529173880815506,
0.014160345308482647,
0.04516288638114929,
-0.013327673077583313,
0.0049615600146353245,
-0.05688655003905296,
-0.05087531358003616,
0.03975321352481842,
0.05583877116441727,
0.028962112963199615,
0.021499885246157646,
0.005180399864912033,
0.003893180750310421,
-0.005766923539340496,
0.07336178421974182,
-0.0378447100520134,
-0.023879054933786392,
-0.009526153095066547,
0.0708516538143158,
-0.029186012223362923,
0.014432240277528763,
-0.03806130215525627,
-0.008231161162257195,
0.010458855889737606,
-0.03342605009675026,
-0.038076214492321014,
0.010014061816036701,
0.05861753225326538,
-0.022888941690325737,
0.04001755639910698,
0.026569031178951263,
0.0508294515311718,
0.003414228791370988,
0.001106182229705155,
0.010016851127147675,
-0.09510806947946548,
0.023662185296416283,
-0.01597447507083416,
-0.08013008534908295,
0.05101507902145386,
0.017233818769454956,
-0.03686925023794174,
-0.06256144493818283,
-0.028009377419948578,
-0.004551309626549482,
-0.03483017534017563,
-1.3684925234700602e-33,
-0.008752759546041489,
0.09091934561729431,
-0.024101344868540764,
0.033374007791280746,
-0.04107161983847618,
-0.009760575369000435,
0.014326147735118866,
0.07435594499111176,
0.008715208619832993,
-0.041306983679533005,
-0.04070642590522766,
-0.044190339744091034,
0.09763328731060028,
-0.08489666134119034,
0.04207015410065651,
-0.07424037158489227,
-0.033024221658706665,
0.07476496696472168,
-0.017954083159565926,
0.02041245810687542,
0.041398774832487106,
0.025977736338973045,
-0.02597433514893055,
0.04037972539663315,
-0.016946470364928246,
-0.05489838495850563,
-0.10057612508535385,
0.032328635454177856,
0.08459948003292084,
-0.0247892364859581,
-0.08555901050567627,
0.03338257223367691,
-0.027563806623220444,
-0.019075775519013405,
-0.06312557309865952,
0.09241779893636703,
0.09475908428430557,
0.013863480649888515,
0.013951212167739868,
0.027919447049498558,
0.04896404594182968,
-0.004895129706710577,
-0.0742778480052948,
0.005891443230211735,
-0.026363670825958252,
-0.02254682406783104,
-0.04596281051635742,
0.02500949613749981,
0.015032687224447727,
-0.03531815856695175,
0.006362771615386009,
-0.02776973322033882,
0.02245795913040638,
0.07612167298793793,
-0.07661855965852737,
-0.0514543317258358,
0.03213009238243103,
0.0016072007128968835,
0.0811384990811348,
0.06616731733083725,
-0.0868559405207634,
0.004572606645524502,
-0.08605783432722092,
-0.055060721933841705,
0.04784202575683594,
-0.034685924649238586,
-0.023668326437473297,
0.048080410808324814,
0.051904164254665375,
0.015310601331293583,
0.035356082022190094,
-0.058449484407901764,
0.024208692833781242,
0.06356951594352722,
-0.07654330879449844,
0.020940471440553665,
-0.04274262115359306,
-0.1142251193523407,
-0.056767113506793976,
0.04908224195241928,
0.033664070069789886,
-0.04606744274497032,
-0.012699621729552746,
0.047767046838998795,
0.0941326841711998,
0.11991926282644272,
0.03411043807864189,
-0.030458997935056686,
-0.04937455803155899,
0.07130289077758789,
-0.1383219063282013,
0.05596509948372841,
-0.05054212361574173,
0.03851541876792908,
0.011643050238490105,
-4.0148201208012324e-8,
-0.055282946676015854,
0.031770847737789154,
-0.02202003262937069,
-0.04025958850979805,
0.017059262841939926,
-0.06354255974292755,
-0.04834608733654022,
0.037720318883657455,
0.0009260911028832197,
-0.04674505442380905,
0.05942725017666817,
-0.0005798318888992071,
0.05443896725773811,
-0.049854688346385956,
-0.08034868538379669,
0.006444546394050121,
-0.041218265891075134,
0.11746730655431747,
0.002517781686037779,
-0.05625734478235245,
0.022003360092639923,
0.08719320595264435,
0.07678745687007904,
-0.048535965383052826,
0.04137915000319481,
-0.06246680021286011,
0.019262218847870827,
0.06291534751653671,
0.021057361736893654,
0.08148586750030518,
-0.11709229648113251,
-0.01477015856653452,
-0.002701265038922429,
-0.01583985798060894,
-0.06495379656553268,
0.11621087044477463,
-0.05979084596037865,
-0.030172301456332207,
-0.007254160009324551,
-0.03340955451130867,
0.016461672261357307,
0.06555162370204926,
-0.05235474184155464,
-0.025651834905147552,
-0.0786590725183487,
-0.001888343133032322,
0.018971489742398262,
-0.05908582732081413,
0.06164585426449776,
-0.02121526375412941,
-0.020874591544270515,
0.00993789080530405,
0.018802862614393234,
0.040716394782066345,
-0.007094633765518665,
0.06851569563150406,
-0.006672923918813467,
-0.02830703742802143,
0.03698991984128952,
0.06132819131016731,
-0.002481078030541539,
-0.0009810281917452812,
-0.0002493294596206397,
-0.014211276546120644
] |
CoInD: Enabling Logical Compositions in Diffusion Models | https://openreview.net/forum?id=cCRlEvjrx4 | [
"Sachit Gaudi",
"Gautam Sreekumar",
"Vishnu Boddeti"
] | Poster | How can we learn generative models to sample data with arbitrary logical compositions of statistically independent attributes? The prevailing solution is to sample from distributions expressed as a composition of attributes' conditional marginal distributions under the assumption that they are statistically independent. This paper shows that standard conditional diffusion models violate this assumption, even when all attribute compositions are observed during training. And, this violation is significantly more severe when only a subset of the compositions is observed. We propose CoInD to address this problem. It explicitly enforces statistical independence between the conditional marginal distributions by minimizing Fisher’s divergence between the joint and marginal distributions. The theoretical advantages of CoInD are reflected in both qualitative and quantitative experiments, demonstrating a significantly more faithful and controlled generation of samples for arbitrary logical compositions of attributes. The benefit is more pronounced for scenarios that current solutions relying on the assumption of conditionally independent marginals struggle with, namely, logical compositions involving the NOT operation and when only a subset of compositions are observed during training. | logical compositionality, generative models, diffusion models, causality | Diffusion models violate the conditional independence required for compositionality. We introduce a training objective, CoInD, to mitigate this violation and significantly enhance compositionality. | 578 | 2503.01145 | [
-0.00712368031963706,
-0.08854755014181137,
0.08994867652654648,
0.049586109817028046,
0.09024696797132492,
0.01762538217008114,
0.004766650497913361,
-0.08699195832014084,
0.07258424907922745,
-0.023176342248916626,
0.052483584731817245,
-0.07719187438488007,
0.03224901854991913,
0.029974857345223427,
0.07131455838680267,
-0.03798575699329376,
0.07201916724443436,
0.08075328171253204,
-0.08421361446380615,
0.009676392190158367,
-0.06161018833518028,
0.044599566608667374,
-0.001644982141442597,
0.02767360396683216,
-0.001007735263556242,
0.037006255239248276,
0.06769298762083054,
-0.04534566402435303,
0.03195379674434662,
-0.0037250849418342113,
0.06425843387842178,
0.026510829105973244,
-0.04609276354312897,
0.019213810563087463,
0.008177613839507103,
0.03441213443875313,
-0.039702534675598145,
0.0682343989610672,
0.032553739845752716,
0.03393439203500748,
-0.038000695407390594,
-0.036307234317064285,
0.01567176915705204,
-0.021446261554956436,
0.04830643907189369,
0.021555600687861443,
0.019124710932374,
0.003499226178973913,
-0.06717993319034576,
0.04038573056459427,
0.032201360911130905,
0.02050020545721054,
-0.0389169305562973,
-0.013190478086471558,
0.02467425912618637,
-0.02901664935052395,
0.01130322553217411,
-0.020247340202331543,
-0.029282813891768456,
0.015149757266044617,
-0.10729777812957764,
-0.08779419213533401,
-0.042555708438158035,
-0.006485846824944019,
0.023393068462610245,
-0.058066535741090775,
-0.03123500943183899,
0.11884193867444992,
-0.06805188208818436,
0.051794495433568954,
0.010742108337581158,
0.07874072343111038,
-0.14463931322097778,
0.032916225492954254,
0.014167209155857563,
0.08614099770784378,
0.09805898368358612,
-0.07867878675460815,
-0.008077194914221764,
-0.1284162402153015,
-0.04550367593765259,
0.03955911844968796,
0.07895293086767197,
0.0109288040548563,
0.08160290122032166,
-0.019925326108932495,
0.008699191734194756,
0.09516941010951996,
-0.020301051437854767,
-0.04079388827085495,
-0.04522286355495453,
0.07313855737447739,
0.06701737642288208,
-0.0035706546623259783,
0.08637852966785431,
0.044485077261924744,
0.0072207278572022915,
-0.06022113561630249,
0.09558076411485672,
0.09667801111936569,
0.003828398184850812,
0.08454596251249313,
0.02203940786421299,
-0.09593910723924637,
0.0342024564743042,
-0.08448468893766403,
0.03134062886238098,
-0.03216474503278732,
0.03964310511946678,
-0.03044278547167778,
0.05550854653120041,
0.06328950077295303,
0.015771668404340744,
-0.07050567865371704,
-0.07170131802558899,
-0.10168640315532684,
0.006404383108019829,
0.0018940622685477138,
0.008649636059999466,
-0.004927405156195164,
-0.005955311935395002,
-0.01175323873758316,
-0.009683887474238873,
0.04832494258880615,
0.010201619938015938,
-0.09089059382677078,
0.018946362659335136,
2.9450238534776564e-33,
0.007683939766138792,
-0.07650180160999298,
-0.013407262973487377,
0.04099612310528755,
0.04568835347890854,
0.030253170058131218,
-0.027395514771342278,
-0.04556342586874962,
0.02606150321662426,
-0.013859326019883156,
0.018583640456199646,
0.002876558108255267,
-0.1214836984872818,
0.07877373695373535,
0.015626754611730576,
0.032262783497571945,
-0.06505170464515686,
0.1099075973033905,
0.007153409533202648,
-0.04524649307131767,
0.04649514704942703,
-0.010060328990221024,
-0.0009476309642195702,
0.08859703689813614,
-0.045117221772670746,
0.010380595922470093,
0.022463036701083183,
-0.0435149148106575,
0.0036782685201615095,
0.0032679191790521145,
-0.06434669345617294,
-0.053492605686187744,
-0.0028646052815020084,
-0.023497987538576126,
0.019948069006204605,
0.045071519911289215,
-0.08305686712265015,
-0.023655295372009277,
-0.02041737362742424,
-0.006653021089732647,
-0.011593624949455261,
0.01758670061826706,
0.029703151434659958,
-0.010256164707243443,
-0.12697501480579376,
-0.013739093206822872,
0.015284618362784386,
0.028164124116301537,
-0.006328464951366186,
-0.07961693406105042,
0.022737981751561165,
0.003530490444973111,
0.03427627310156822,
-0.11768589168787003,
-0.023692036047577858,
0.05539451912045479,
-0.017180021852254868,
-0.0021436302922666073,
0.004697839729487896,
-0.040059078484773636,
-0.020153602585196495,
0.13326376676559448,
-0.06001777946949005,
0.003751560114324093,
-0.03882213681936264,
0.06841127574443817,
-0.05621643364429474,
-0.0166862141340971,
0.06721536070108414,
-0.004146336577832699,
0.019162053242325783,
0.016694912686944008,
-0.11159675568342209,
-0.08983073383569717,
0.09113829582929611,
-0.020129943266510963,
0.04486962407827377,
-0.002904568798840046,
0.042353980243206024,
-0.0278059970587492,
-0.021937808021903038,
-0.01018083468079567,
-0.025755563750863075,
-0.04615815728902817,
-0.0012495162663981318,
0.01495666615664959,
-0.006291012279689312,
-0.053227100521326065,
-0.12275366485118866,
-0.04494274780154228,
0.0003719998640008271,
-0.05176911875605583,
-0.06611350178718567,
-0.02634320594370365,
0.06226758658885956,
-4.793689472540305e-33,
-0.13517878949642181,
0.07743921875953674,
0.06943192332983017,
0.024643519893288612,
-0.02850056253373623,
-0.02483336254954338,
-0.01302995067089796,
0.011105257086455822,
0.01868758350610733,
-0.10748005658388138,
-0.06822779774665833,
-0.028628844767808914,
-0.016116024926304817,
-0.05507943406701088,
-0.09624546766281128,
0.06102911755442619,
-0.03676244616508484,
0.06338100135326385,
-0.029094599187374115,
0.03982749581336975,
-0.015050172805786133,
0.1276848316192627,
-0.015845682471990585,
0.0030988967046141624,
-0.030341466888785362,
0.0354304313659668,
-0.025539007037878036,
0.09606791287660599,
-0.0384359173476696,
0.020729847252368927,
-0.010880433022975922,
-0.04916870966553688,
-0.03197941184043884,
-0.08309382200241089,
-0.14108407497406006,
-0.0057278405874967575,
0.0020688611548393965,
-0.0112537806853652,
-0.0724380612373352,
0.0832919105887413,
-0.03605709224939346,
-0.0400652177631855,
-0.06299164146184921,
0.027291204780340195,
-0.01251917239278555,
-0.06579120457172394,
-0.007176629733294249,
-0.03418859466910362,
0.041018541902303696,
0.006272591184824705,
-0.007392659783363342,
0.06205494701862335,
-0.05066835880279541,
0.028176680207252502,
-0.0377001017332077,
-0.053630392998456955,
-0.0056621539406478405,
-0.061618853360414505,
0.041571587324142456,
-0.009916551411151886,
-0.056741345673799515,
0.015317493118345737,
0.06069614365696907,
0.008617714047431946,
-0.022068273276090622,
-0.050983164459466934,
-0.048547111451625824,
-0.019015172496438026,
0.0641651377081871,
-0.007637791335582733,
0.023509209975600243,
0.04963024705648422,
0.025780972093343735,
-0.043309636414051056,
0.03616541251540184,
0.021056335419416428,
0.03030853159725666,
0.0007862953352741897,
-0.008203504607081413,
-0.0007204872090369463,
-0.01769036427140236,
-0.06358864903450012,
0.06688450276851654,
0.02750752680003643,
0.04394683986902237,
0.0165169145911932,
0.03782003000378609,
-0.027574561536312103,
-0.0009198600891977549,
0.019144538789987564,
-0.060548730194568634,
-0.004616140387952328,
0.11050856858491898,
0.1253431737422943,
-0.10257788747549057,
-4.673393050325103e-8,
-0.061462804675102234,
0.038261983543634415,
0.016614118590950966,
0.028061287477612495,
0.06578250229358673,
0.046404678374528885,
-0.00480084540322423,
0.014201524667441845,
-0.04022771120071411,
0.03401355445384979,
0.03896671533584595,
0.034691113978624344,
-0.06015510857105255,
-0.01725255511701107,
0.015192773193120956,
0.06614919751882553,
0.030534064397215843,
0.013672184199094772,
-0.051132604479789734,
0.011402290314435959,
0.07343634217977524,
0.06808531284332275,
-0.006143687758594751,
0.012109524570405483,
0.08091362565755844,
-0.02727738581597805,
-0.02243751846253872,
-0.011409466154873371,
0.06911257654428482,
0.025770729407668114,
-0.007924260571599007,
0.016286710277199745,
-0.06232542544603348,
0.01563233509659767,
0.02948872372508049,
0.10348068922758102,
-0.00670278025791049,
-0.02769814059138298,
-0.08566723018884659,
-0.024055762216448784,
-0.0450172983109951,
0.006457443814724684,
-0.07820765674114227,
0.03955700621008873,
0.06491497904062271,
0.012052270583808422,
-0.02050560526549816,
0.015301693230867386,
-0.009286675602197647,
0.0037043862976133823,
-0.018293321132659912,
-0.08648862689733505,
-0.011730332858860493,
0.02809876576066017,
0.08774586021900177,
-0.0013050815323367715,
0.0222172774374485,
0.03300391882658005,
0.01614764891564846,
-0.006815291475504637,
0.012901474721729755,
0.0035847225226461887,
0.012411990202963352,
-0.04825273156166077
] |
HAMSTER: Hierarchical Action Models for Open-World Robot Manipulation | https://openreview.net/forum?id=h7aQxzKbq6 | [
"Yi Li",
"Yuquan Deng",
"Jesse Zhang",
"Joel Jang",
"Marius Memmel",
"Caelan Reed Garrett",
"Fabio Ramos",
"Dieter Fox",
"Anqi Li",
"Abhishek Gupta",
"Ankit Goyal"
] | Poster | Large foundation models have shown strong open-world generalization to complex problems in vision and language, but similar levels of generalization have yet to be achieved in robotics. One fundamental challenge is the lack of robotic data, which are typically obtained through expensive on-robot operation. A promising remedy is to leverage cheaper, *off-domain* data such as action-free videos, hand-drawn sketches, or simulation data. In this work, we posit that *hierarchical* vision-language-action (VLA) models can be more effective in utilizing off-domain data than standard monolithic VLA models that directly finetune vision-language models (VLMs) to predict actions.
In particular, we study a class of hierarchical VLA models, where the high-level VLM is finetuned to produce a coarse 2D path indicating the desired robot end-effector trajectory given an RGB image and a task description. The intermediate 2D path prediction is then served as guidance to the low-level, 3D-aware control policy capable of precise manipulation. Doing so alleviates the high-level VLM from fine-grained action prediction, while reducing the low-level policy's burden on complex task-level reasoning.
We show that, with the hierarchical design, the high-level VLM can transfer across significant domain gaps between the off-domain finetuning data and real-robot testing scenarios, including differences in embodiments, dynamics, visual appearances, and task semantics, etc.
In the real-robot experiments, we observe an average of 20% improvement in success rate across seven different axes of generalization over OpenVLA, representing a 50% relative gain.
Visual results are provided at: [https://hamster-robot.github.io/](https://hamster-robot.github.io/) | vision language model; cross-domain generalization; sim-to-real transfer; robot manipulation; vision language action model | Hierarchical VLA architectures can enable robotic manipulation with semantic, visual, and geometric generalization after trained on cheap off-domain data | 573 | 2502.05485 | [
0.030360765755176544,
-0.13864299654960632,
0.05378096178174019,
-0.0022975022438913584,
0.06297248601913452,
-0.03133954852819443,
-0.07138657569885254,
0.018759889528155327,
0.0209736879914999,
0.03429132327437401,
-0.007087575271725655,
-0.050839345902204514,
-0.01480960100889206,
0.11807136982679367,
0.03658392280340195,
0.023969141766428947,
0.08987104147672653,
0.04384557530283928,
-0.11450035125017166,
-0.01690238155424595,
0.08893334120512009,
0.0010729762725532055,
0.09893982112407684,
-0.0027730066794902086,
-0.016860388219356537,
0.03729339689016342,
0.022873181849718094,
-0.08262810856103897,
-0.045791588723659515,
-0.07823590934276581,
0.021949218586087227,
-0.027436161413788795,
0.005866479128599167,
-0.053100794553756714,
-0.006078224163502455,
0.09518253058195114,
-0.03499501198530197,
0.015387166291475296,
-0.07936827093362808,
-0.0744997039437294,
-0.04142579063773155,
0.010866843163967133,
0.04823091998696327,
0.010101163759827614,
0.08496639877557755,
0.05584835633635521,
0.07133949548006058,
-0.10503444820642471,
0.02542993240058422,
0.029484160244464874,
-0.05665656551718712,
-0.11024509370326996,
0.049362506717443466,
-0.017912307754158974,
-0.02846156246960163,
0.008856217376887798,
0.04913189262151718,
-0.06640666723251343,
0.07439854741096497,
-0.02234719879925251,
-0.004289176780730486,
-0.09595751762390137,
-0.04329086095094681,
-0.056081317365169525,
-0.12344400584697723,
-0.09043852984905243,
0.05224330350756645,
-0.04168029502034187,
-0.033321477472782135,
-0.004699405748397112,
-0.008389156311750412,
0.00225151632912457,
0.042531415820121765,
-0.007288664113730192,
-0.010603372938930988,
-0.034055229276418686,
0.01529724895954132,
-0.007750126998871565,
0.0512474924325943,
-0.056267984211444855,
0.04553587734699249,
-0.017888061702251434,
0.003804369829595089,
0.06000495329499245,
0.04247165098786354,
0.06796380132436752,
-0.023253919556736946,
0.0682249516248703,
0.05021573230624199,
0.030716607347130775,
-0.027991430833935738,
-0.07399044930934906,
0.008280692622065544,
-0.029481234028935432,
-0.0065779234282672405,
-0.10193942487239838,
0.020104773342609406,
-0.1422743797302246,
-0.032643578946590424,
0.04012409225106239,
-0.011894271709024906,
0.006536432076245546,
0.021549297496676445,
-0.00657240953296423,
0.06860653311014175,
0.02804604358971119,
0.03945592790842056,
0.038089051842689514,
-0.011730422265827656,
-0.017778346315026283,
-0.001401987741701305,
-0.054502587765455246,
0.043141841888427734,
0.03936174884438515,
0.024861231446266174,
-0.013655520044267178,
0.017556771636009216,
-0.041739627718925476,
-0.04179343208670616,
-0.005183378700166941,
-0.0037862167228013277,
-0.037585314363241196,
0.035741034895181656,
0.020157288759946823,
0.028729191049933434,
-0.023697448894381523,
-0.04957744851708412,
-1.0467913110606784e-33,
0.0880742147564888,
0.0398576483130455,
0.06201060488820076,
0.05233697593212128,
0.0853302925825119,
-0.030960913747549057,
0.050427671521902084,
0.025484347715973854,
-0.04504616931080818,
0.021550845354795456,
-0.0018600658513605595,
-0.06141474470496178,
-0.0809992179274559,
0.1200459748506546,
0.03436889871954918,
0.026975363492965698,
0.018209122121334076,
0.006926561240106821,
-0.049190156161785126,
0.10711095482110977,
0.02886105142533779,
0.02519609034061432,
-0.03710278868675232,
-0.034584421664476395,
0.036354415118694305,
-0.04215770959854126,
0.017805881798267365,
-0.05952766537666321,
0.005061219912022352,
-0.018135366961359978,
-0.05864797905087471,
0.030433891341090202,
0.009719610214233398,
0.07252391427755356,
0.0031679992098361254,
0.027096502482891083,
-0.03989484906196594,
-0.07360552251338959,
0.019962873309850693,
-0.013082253746688366,
0.01338043250143528,
0.02912878431379795,
-0.004887582268565893,
-0.04234828054904938,
-0.07938432693481445,
-0.0383710078895092,
0.02851894125342369,
0.09958349913358688,
-0.07350071519613266,
0.0015211727004498243,
0.05681806430220604,
0.0669572576880455,
-0.02505049854516983,
-0.06211519241333008,
-0.03787162899971008,
-0.039504311978816986,
0.031026897951960564,
0.04908238351345062,
0.008496553637087345,
0.04615195095539093,
0.00294116442091763,
0.03065742738544941,
-0.019984371960163116,
0.10603517293930054,
-0.019571524113416672,
-0.013727224431931973,
-0.011169211938977242,
0.06725310534238815,
0.04663384333252907,
0.015328715555369854,
-0.039529163390398026,
0.0353483110666275,
0.03304314240813255,
-0.09312529116868973,
0.06017366424202919,
0.018254989758133888,
0.01690681092441082,
-0.1319681853055954,
0.06892319023609161,
0.04187851771712303,
-0.046476930379867554,
-0.010869398713111877,
0.028459561988711357,
0.02599232830107212,
-0.015333517454564571,
-0.020585017278790474,
0.034081973135471344,
-0.016769958660006523,
-0.05423865094780922,
-0.03311429172754288,
-0.11678759008646011,
-0.11457255482673645,
-0.07036030292510986,
0.015808986499905586,
0.04507453739643097,
6.067462865273013e-34,
0.044452402740716934,
-0.013591119088232517,
0.04536975920200348,
0.04704227298498154,
-0.04359864816069603,
-0.0538104847073555,
-0.02023422159254551,
-0.05277947708964348,
0.041776373982429504,
-0.05497989058494568,
-0.04569033533334732,
0.002748525235801935,
-0.0009036267292685807,
0.06411173194646835,
0.08466297388076782,
-0.06771628558635712,
-0.020911693572998047,
0.0026644994504749775,
0.015305561013519764,
-0.01792026311159134,
-0.026998812332749367,
0.08449897915124893,
-0.04347074776887894,
0.0036434554494917393,
-0.048374440521001816,
0.005644759628921747,
-0.054527051746845245,
0.057736948132514954,
-0.02503853850066662,
0.07099892944097519,
-0.0463164784014225,
-0.04059299826622009,
0.05524445325136185,
0.012070474214851856,
-0.026855411008000374,
0.03692188858985901,
-0.0008790328283794224,
-0.005782353691756725,
-0.012726493179798126,
0.02195923775434494,
0.03642507269978523,
-0.06828676909208298,
-0.07345882058143616,
0.02929772436618805,
0.01475457288324833,
-0.0004128234286326915,
-0.06248568743467331,
0.019556695595383644,
-0.021050002425909042,
0.021876314654946327,
-0.09839855879545212,
0.09517935663461685,
-0.01945570297539234,
-0.02893110178411007,
0.007142117712646723,
-0.060399703681468964,
-0.046762410551309586,
-0.04470791667699814,
0.084254689514637,
-0.06371469050645828,
-0.03308964893221855,
0.03184748813509941,
-0.06512414664030075,
0.054744258522987366,
0.006018226034939289,
-0.00713485199958086,
-0.049387313425540924,
-0.015278041362762451,
0.029877183958888054,
-0.02071886509656906,
-0.004715398419648409,
0.001890056417323649,
0.048903122544288635,
0.027234937995672226,
0.06871926784515381,
-0.06808847934007645,
0.007461413275450468,
-0.0630321279168129,
0.03850705549120903,
-0.02175712026655674,
0.04747987911105156,
-0.029786881059408188,
0.08841171115636826,
0.0803382471203804,
0.0440935455262661,
0.11415921151638031,
-0.007027789484709501,
0.06007422134280205,
0.008804353885352612,
0.02594074420630932,
-0.038629088550806046,
-0.017360273748636246,
0.01986149698495865,
0.07926074415445328,
-0.06507574766874313,
-3.8395242540900654e-8,
-0.0810035988688469,
0.0678718164563179,
0.07362093031406403,
-0.032889656722545624,
0.010765316896140575,
0.035606250166893005,
0.016539087519049644,
0.06160671263933182,
0.03702067583799362,
0.03867657110095024,
0.043638575822114944,
0.04291119799017906,
0.013446265831589699,
-0.0689706802368164,
-0.02510593831539154,
0.08306474983692169,
0.057207562029361725,
0.00001847209205152467,
0.04855813458561897,
0.047286506742239,
0.005103347357362509,
-0.010896170511841774,
0.03797154873609543,
-0.02503196895122528,
0.0047513809986412525,
-0.06725567579269409,
-0.08416838198900223,
0.10512044280767441,
0.04978948086500168,
-0.014946479350328445,
0.07313842326402664,
-0.02193864993751049,
0.07385311275720596,
0.04032288119196892,
0.006947001907974482,
-0.030534477904438972,
-0.011344569735229015,
-0.0028136293403804302,
0.014830220490694046,
-0.013552005402743816,
0.05173536762595177,
0.10823316872119904,
-0.0012378152459859848,
-0.000025337802071589977,
-0.004773502703756094,
0.04717042297124863,
-0.08283186703920364,
-0.19340287148952484,
0.046255581080913544,
-0.019544692710042,
-0.05259452760219574,
0.04601607471704483,
-0.03179898113012314,
0.12132112681865692,
0.06079604849219322,
0.036256927996873856,
0.021428437903523445,
-0.05604523792862892,
0.002414826536551118,
0.057071179151535034,
0.019373035058379173,
-0.02836322784423828,
-0.05314715951681137,
-0.019225310534238815
] |
Mitigating Object Hallucination in MLLMs via Data-augmented Phrase-level Alignment | https://openreview.net/forum?id=yG1fW8igzP | [
"Pritam Sarkar",
"Sayna Ebrahimi",
"Ali Etemad",
"Ahmad Beirami",
"Sercan O Arik",
"Tomas Pfister"
] | Poster | Despite their significant advancements, Multimodal Large Language Models
(MLLMs) often generate factually inaccurate information, referred to as hallucination.
In this work, we address object hallucinations in MLLMs, where information
is generated about an object not present in the input image. We introduce Data-augmented
Phrase-level Alignment (DPA), a novel loss which can be applied to
instruction-tuned off-the-shelf MLLMs to mitigate hallucinations, while preserving
their general vision-language capabilities. To fine-tune MLLMs with DPA, we first
generate a set of 'hallucinated' and 'correct' response pairs through generative data
augmentation by selectively altering the ground-truth information of the correct
responses at a phrase level. The DPA loss is then used to train MLLMs to reduce
the likelihood of hallucinated phrases compared to the correct ones. Our thorough
evaluation on various benchmarks confirms the effectiveness of DPA in mitigating
hallucination while retaining the out-of-the-box performance of the MLLMs on
general tasks. For instance, MLLMs finetuned with DPA, which we refer to as Hallucination
Attenuated Language and Vision Assistant (HALVA), improve F1 by up
to 13.4% on hallucination visual question-answering and reduce the hallucination
rate by up to 4.2% on image description tasks. | Multimodal LLMs, Object Hallucination, Vision-language Models | We introduce phrase-level alignment method that can be applied to off-the-shelf MLLMs for mitigating hallucinations, while preserving their general vision-language capabilities. | 568 | 2405.18654 | [
-0.03900889679789543,
-0.06539260596036911,
0.03252530097961426,
-0.025383245199918747,
0.02568482980132103,
-0.022516902536153793,
-0.055243462324142456,
-0.06931207329034805,
0.04081876948475838,
-0.029030079022049904,
-0.06091455742716789,
-0.05597667396068573,
0.023713288828730583,
0.00041035038884729147,
0.01808803342282772,
-0.011502956971526146,
0.04821532219648361,
0.15278080105781555,
-0.11507035046815872,
-0.048121899366378784,
0.07736614346504211,
0.025921911001205444,
-0.0016055559972301126,
-0.017672056332230568,
0.051928095519542694,
0.07714404165744781,
0.01496280636638403,
-0.060015659779310226,
0.005709839519113302,
-0.11448246240615845,
0.013359946198761463,
0.06146679073572159,
-0.02468218095600605,
0.027518803253769875,
-0.033635158091783524,
0.08161721378564835,
-0.07037381082773209,
0.057640910148620605,
0.02463947981595993,
-0.1094186082482338,
-0.04333729296922684,
0.025047169998288155,
0.033450499176979065,
-0.021135035902261734,
0.0325644426047802,
-0.0034410373773425817,
-0.03216131404042244,
-0.017209716141223907,
0.0004247285833116621,
-0.029038919135928154,
-0.07431413978338242,
-0.004294435493648052,
0.030099166557192802,
0.07114136964082718,
-0.06541866809129715,
0.07075756043195724,
0.07598775625228882,
0.01533837616443634,
-0.0008846899145282805,
0.009451034478843212,
-0.02215546742081642,
-0.030450474470853806,
-0.023471836000680923,
-0.02691558562219143,
-0.0222325399518013,
-0.03518950566649437,
0.030870072543621063,
-0.04333450272679329,
-0.015135104767978191,
0.03925875946879387,
-0.033694300800561905,
0.02086258865892887,
-0.03917183727025986,
0.05797979235649109,
0.05052104592323303,
0.0650407150387764,
0.01600443758070469,
-0.1053861603140831,
0.09484390169382095,
-0.10659746825695038,
0.03799401596188545,
0.051306091248989105,
0.06512587517499924,
0.02788175642490387,
0.09670490771532059,
-0.02543274126946926,
-0.01739371381700039,
0.02862580493092537,
-0.0005993424565531313,
0.019726166501641273,
-0.08362005650997162,
-0.15415044128894806,
0.00905955582857132,
0.004439987242221832,
0.03290446475148201,
0.03018893674015999,
0.020649420097470284,
-0.10221035778522491,
-0.04068165645003319,
0.05162457749247551,
0.006804353557527065,
0.03746221587061882,
-0.05351484194397926,
-0.07715637236833572,
0.0016964769456535578,
-0.06964000314474106,
0.09586023539304733,
0.02833421900868416,
0.016487207263708115,
-0.06501293927431107,
0.035422537475824356,
0.03527480736374855,
-0.025804273784160614,
0.006921358406543732,
-0.016705313697457314,
-0.05862671136856079,
-0.01947988010942936,
-0.0383312851190567,
-0.010596317239105701,
0.00017902102263178676,
-0.017013030126690865,
0.03211470693349838,
0.016976695507764816,
0.03342972695827484,
0.04247934743762016,
-0.11213850975036621,
-0.04567199572920799,
1.8073293602159827e-33,
0.028116503730416298,
0.05208083242177963,
0.016777921468019485,
0.05813341960310936,
0.09069370478391647,
-0.01762166991829872,
-0.05872659385204315,
-0.004081498831510544,
0.052335865795612335,
0.0018683678936213255,
-0.09138697385787964,
-0.0230560265481472,
-0.07780823856592178,
0.07719123363494873,
0.0018054042011499405,
0.04746035113930702,
-0.09123498946428299,
0.05651203915476799,
-0.0691370815038681,
0.05445471778512001,
0.02510937675833702,
0.00034454173874109983,
-0.05041305348277092,
0.037766728550195694,
-0.001018110429868102,
0.023705122992396355,
0.04154008999466896,
-0.0832548514008522,
-0.03746294230222702,
0.027676640078425407,
-0.11937564611434937,
-0.005335416179150343,
0.028876958414912224,
0.02711671032011509,
0.058394260704517365,
-0.025171691551804543,
-0.003381835063919425,
-0.041388172656297684,
0.006900718901306391,
0.013032661750912666,
0.021118858829140663,
0.08408742398023605,
-0.025083027780056,
-0.01813899353146553,
-0.05555545911192894,
0.003606511978432536,
0.04556195065379143,
0.020703688263893127,
-0.08214700222015381,
-0.043117817491292953,
0.05140162631869316,
0.064509317278862,
-0.01193155162036419,
-0.04182824864983559,
-0.07738446444272995,
-0.054696302860975266,
0.013647490181028843,
-0.010729790665209293,
0.09047094732522964,
-0.02305065281689167,
0.0328373946249485,
0.07925357669591904,
-0.0035123592242598534,
0.07044804841279984,
0.011341826990246773,
-0.049458399415016174,
0.030359575524926186,
0.002267502946779132,
-0.010050689801573753,
-0.050770353525877,
-0.04461279138922691,
-0.001855396432802081,
-0.011755733750760555,
0.002338283695280552,
-0.02828053943812847,
-0.0024189455434679985,
0.054431043565273285,
-0.05013106390833855,
0.0299814585596323,
0.02652036026120186,
0.03327920660376549,
0.012659020721912384,
-0.004952666349709034,
-0.06904713809490204,
-0.029276324436068535,
-0.061121828854084015,
0.032027989625930786,
-0.10508133471012115,
-0.09563878178596497,
-0.050420019775629044,
-0.015073388814926147,
-0.016075825318694115,
0.0025357045233249664,
-0.031132493168115616,
0.004118073731660843,
-2.6489258263841457e-33,
0.014802636578679085,
0.0058584497310221195,
0.07367470860481262,
0.055977001786231995,
-0.057012490928173065,
-0.024047158658504486,
0.03884023800492287,
0.004353705327957869,
0.032102975994348526,
-0.06260424107313156,
-0.03274352476000786,
-0.01587478071451187,
-0.03433012589812279,
-0.021433217450976372,
0.029409298673272133,
-0.06591441482305527,
0.007647658698260784,
0.04829220101237297,
-0.03836505115032196,
0.07504204660654068,
0.005852327216416597,
0.12732373178005219,
-0.12831929326057434,
-0.02461814135313034,
-0.03133401274681091,
0.059645529836416245,
-0.03296510502696037,
0.049283452332019806,
0.057688526809215546,
0.003249374683946371,
0.025543106719851494,
0.08263853192329407,
-0.02684941701591015,
-0.02457774430513382,
-0.03414224088191986,
0.05650884285569191,
0.024127909913659096,
-0.032697439193725586,
-0.07206747680902481,
0.04622219502925873,
0.08799704909324646,
-0.006809911224991083,
-0.0959319993853569,
0.006776389665901661,
0.015549358911812305,
-0.10998153686523438,
-0.0822010487318039,
-0.026500651612877846,
0.02361377701163292,
0.05488205328583717,
-0.020149383693933487,
0.0360659621655941,
-0.10711761564016342,
-0.001032728934660554,
-0.03547488525509834,
-0.09087251871824265,
-0.07259662449359894,
0.00184578204061836,
0.05752458795905113,
0.056673675775527954,
-0.07358668744564056,
-0.022661184892058372,
0.020582642406225204,
-0.052726149559020996,
0.025282682850956917,
0.06272508949041367,
-0.014476352371275425,
0.013366000726819038,
0.056331370025873184,
-0.002233498729765415,
0.07913845777511597,
-0.04617205634713173,
-0.0048764729872345924,
0.1073920801281929,
0.0026485812850296497,
-0.08138958364725113,
0.03825075179338455,
-0.03644871711730957,
0.02810332179069519,
-0.07846907526254654,
0.007794871926307678,
-0.06436332315206528,
-0.00037557154428213835,
0.12653782963752747,
0.06726371496915817,
0.005421306937932968,
-0.006786883808672428,
0.07407557964324951,
-0.004174207802861929,
0.0478455014526844,
-0.06357543170452118,
0.028153469786047935,
0.03535344824194908,
0.055612336844205856,
-0.028203412890434265,
-4.24175716773334e-8,
-0.056099336594343185,
0.014978207647800446,
0.0014227682258933783,
0.05830175057053566,
-0.0160045363008976,
-0.0504542775452137,
-0.03322579711675644,
0.10524512827396393,
0.05658810958266258,
-0.12469211220741272,
0.0682673305273056,
-0.00611979141831398,
-0.032425735145807266,
-0.07217933982610703,
0.053966570645570755,
0.10202545672655106,
-0.0485832616686821,
0.08101147413253784,
-0.015875088050961494,
0.015848640352487564,
0.04136842116713524,
0.03553062304854393,
-0.03870691731572151,
-0.04543443024158478,
0.019086601212620735,
-0.010269267484545708,
0.024536609649658203,
0.0662674754858017,
0.003765395376831293,
-0.00900898315012455,
0.047594521194696426,
0.03590353578329086,
0.03941728174686432,
0.00025171085144393146,
-0.008559186942875385,
0.04794827848672867,
0.011409880593419075,
0.0011205486953258514,
0.038824599236249924,
-0.06197592616081238,
0.03329426050186157,
0.05853200703859329,
-0.07363153994083405,
0.036774929612874985,
0.042228829115629196,
-0.007793643977493048,
0.022201301530003548,
-0.08935800939798355,
0.0016368081560358405,
-0.051115334033966064,
-0.024095479398965836,
0.07222183793783188,
0.033103495836257935,
0.07528369128704071,
0.04713006317615509,
0.027449902147054672,
0.07407911121845245,
-0.016006940975785255,
0.06427262723445892,
0.05626646429300308,
0.030637690797448158,
0.00939467828720808,
-0.0818796455860138,
-0.017021212726831436
] |
Breaking Neural Network Scaling Laws with Modularity | https://openreview.net/forum?id=5Qxx5KpFms | [
"Akhilan Boopathy",
"Sunshine Jiang",
"William Yue",
"Jaedong Hwang",
"Abhiram Iyer",
"Ila R Fiete"
] | Poster | Modular neural networks outperform nonmodular neural networks on tasks ranging from visual question answering to robotics. These performance improvements are thought to be due to modular networks' superior ability to model the compositional and combinatorial structure of real-world problems. However, a theoretical explanation of how modularity improves generalizability, and how to leverage task modularity while training networks remains elusive. Using recent theoretical progress in explaining neural network generalization, we investigate how the amount of training data required to generalize on a task varies with the intrinsic dimensionality of a task's input. We show theoretically that when applied to modularly structured tasks, while nonmodular networks require an exponential number of samples with task dimensionality, modular networks' sample complexity is independent of task dimensionality: modular networks can generalize in high dimensions. We then develop a novel learning rule for modular networks to exploit this advantage and empirically show the improved generalization of the rule, both in- and out-of-distribution, on high-dimensional, modular tasks. | scaling laws, modularity, neural network, generalization, compositionality, combinatorial generalization | We show theoretically that modular neural networks trained on modular tasks can generalize to high-dimensional tasks with a fixed number of training points; we propose a learning rule to exploit this advantage empirically. | 565 | 2409.05780 | [
-0.06358923763036728,
-0.11381597071886063,
0.06478290259838104,
-0.005625706631690264,
0.111325703561306,
-0.0471825897693634,
0.033507946878671646,
0.0052994005382061005,
0.029044711962342262,
-0.05064988136291504,
0.02896111086010933,
-0.05358445644378662,
0.017482884228229523,
0.09030044078826904,
0.012214972637593746,
-0.05462897568941116,
0.05728096142411232,
0.10734658688306808,
-0.17094869911670685,
0.0040059746243059635,
0.05728895589709282,
-0.007397443521767855,
0.05223964899778366,
0.03910306841135025,
0.015823079273104668,
0.07326381653547287,
0.007734367158263922,
-0.04556114226579666,
-0.011024897918105125,
-0.008257027715444565,
0.02888982556760311,
0.03985802084207535,
-0.0064620887860655785,
-0.004406004212796688,
-0.020993946120142937,
0.08188941329717636,
-0.054329998791217804,
0.0507856123149395,
-0.00007350228406721726,
-0.08601021021604538,
0.026302171871066093,
0.04554535448551178,
0.061204247176647186,
0.036478444933891296,
0.055611830204725266,
0.016504745930433273,
0.00016127129492815584,
-0.10375076532363892,
-0.00033026887103915215,
0.006617730483412743,
-0.06330734491348267,
-0.010239220224320889,
-0.058277156203985214,
0.056608766317367554,
0.05227823927998543,
-0.008254156447947025,
0.03189007565379143,
0.0018144322093576193,
-0.024549588561058044,
-0.05547038093209267,
0.0031964615918695927,
-0.11214950680732727,
-0.03801679611206055,
-0.07791567593812943,
0.052653662860393524,
0.03141539916396141,
0.0538017712533474,
-0.041890304535627365,
-0.014389868825674057,
0.011301142163574696,
0.03509058058261871,
0.08404731750488281,
-0.061142873018980026,
0.06128346920013428,
0.06275883316993713,
-0.006986438296735287,
0.01364076230674982,
-0.006179960910230875,
0.003924763295799494,
-0.005715713370591402,
-0.032380033284425735,
0.03830653801560402,
0.05076315626502037,
0.0914749875664711,
0.10201691836118698,
-0.012497900985181332,
-0.018065715208649635,
0.022386236116290092,
-0.0021268005948513746,
0.006036806385964155,
-0.06454625725746155,
-0.024415185675024986,
0.010137369856238365,
-0.07576095312833786,
-0.021602336317300797,
-0.02563810721039772,
-0.017538007348775864,
-0.07447289675474167,
-0.007572445087134838,
0.07027358561754227,
-0.026069577783346176,
0.031472887843847275,
0.04631434753537178,
0.01625577174127102,
0.09012480080127716,
0.039267752319574356,
0.06584586948156357,
0.033059656620025635,
0.07659947127103806,
-0.11030174046754837,
-0.019526982679963112,
0.006902647204697132,
0.06394456326961517,
0.000999769545160234,
0.0076845805160701275,
-0.054613493382930756,
0.046833138912916183,
0.0330439992249012,
-0.026626035571098328,
0.11807961761951447,
-0.006852572318166494,
0.016007203608751297,
-0.03326933830976486,
-0.010725032538175583,
0.08154579252004623,
0.027277415618300438,
-0.07099062204360962,
3.31947629323515e-33,
0.03227504715323448,
0.005324068479239941,
0.08103226870298386,
0.12511248886585236,
0.1122761070728302,
-0.03386897221207619,
-0.011481355875730515,
0.014622686430811882,
0.09931343048810959,
0.02328457497060299,
-0.07098830491304398,
0.03789898753166199,
-0.03057152032852173,
0.06997678428888321,
0.10350515693426132,
0.03321041166782379,
-0.0275528933852911,
0.0067122094333171844,
0.031756557524204254,
-0.012949955649673939,
0.06073985993862152,
-0.05935802683234215,
0.041904374957084656,
-0.021084321662783623,
0.040428467094898224,
-0.07371895760297775,
0.017366087064146996,
-0.02815885655581951,
-0.06354101747274399,
-0.012509195134043694,
-0.04461543262004852,
0.0012206468963995576,
-0.04279695823788643,
0.028078630566596985,
0.03446098044514656,
0.015635129064321518,
0.014237922616302967,
-0.04134882614016533,
0.053762029856443405,
-0.007249154150485992,
0.00020074666826985776,
-0.023559218272566795,
-0.01851627789437771,
0.019001569598913193,
-0.07877746224403381,
-0.015424935147166252,
0.01823979988694191,
0.07109151780605316,
-0.0642966777086258,
-0.0710408091545105,
0.030628737062215805,
0.013958924449980259,
0.047728028148412704,
-0.10106933861970901,
0.03107043355703354,
0.010315001010894775,
-0.007010371889919043,
0.03648559749126434,
0.03682729974389076,
0.02317175827920437,
0.03913987800478935,
0.013295534998178482,
-0.02425324358046055,
0.14060543477535248,
-0.012410158291459084,
0.029713749885559082,
-0.08055344969034195,
0.03191457688808441,
0.13828974962234497,
-0.038705989718437195,
-0.006670005619525909,
0.041917528957128525,
-0.06873100250959396,
-0.014972916804254055,
0.03325352817773819,
0.026850901544094086,
-0.002776361769065261,
-0.1459568291902542,
-0.018730241805315018,
0.10204073041677475,
-0.04363500699400902,
0.041248854249715805,
0.002352094976231456,
-0.0631793737411499,
-0.04422527551651001,
-0.07485783845186234,
0.10503318905830383,
-0.05193657428026199,
-0.01666397973895073,
-0.08376253396272659,
-0.09934480488300323,
-0.06567340344190598,
-0.034562792629003525,
-0.07048597186803818,
-0.007838431745767593,
-2.719768279133397e-33,
-0.07113807648420334,
-0.03297727555036545,
-0.04934977367520332,
0.0054145990870893,
-0.00013849239621777087,
-0.0215592160820961,
-0.034448496997356415,
0.011602387763559818,
-0.011860406957566738,
-0.010609974153339863,
-0.037355490028858185,
-0.01378328911960125,
0.02306019514799118,
0.009613756090402603,
0.007989829406142235,
0.032322973012924194,
-0.09157363325357437,
0.05215812101960182,
0.06661837548017502,
-0.05022754520177841,
-0.020573314279317856,
0.06531799584627151,
-0.07114262133836746,
0.011597289703786373,
-0.028331412002444267,
0.017609255388379097,
-0.1640821248292923,
0.049820490181446075,
-0.009812095202505589,
0.03222634270787239,
-0.0838528648018837,
-0.05900530517101288,
0.018397588282823563,
0.00398052716627717,
-0.01890668459236622,
0.05562656372785568,
-0.014815853908658028,
-0.001361063332296908,
-0.018190186470746994,
0.0571243092417717,
-0.02877396158874035,
-0.0989585891366005,
-0.030137039721012115,
0.009520182386040688,
0.03836195170879364,
-0.10078348219394684,
0.027815531939268112,
-0.06512559205293655,
-0.01884545013308525,
0.061632853001356125,
0.01449127122759819,
0.040629055351018906,
-0.06727030128240585,
-0.005074276123195887,
-0.052065327763557434,
-0.01699521765112877,
0.011625077575445175,
-0.0016834221314638853,
0.13619601726531982,
0.0007158114458434284,
-0.032829709351062775,
-0.04925238713622093,
0.01259654201567173,
0.05639808252453804,
-0.0047869556583464146,
-0.035481467843055725,
-0.026621833443641663,
-0.015361291356384754,
0.06525282561779022,
-0.031067196279764175,
-0.05195888504385948,
-0.05779322609305382,
0.05572464317083359,
0.01663900353014469,
-0.048910558223724365,
-0.021673014387488365,
0.04791956767439842,
0.030529810115695,
-0.014261356554925442,
-0.06472384929656982,
0.000511625490617007,
-0.03534281253814697,
0.026774195954203606,
0.06337273120880127,
0.037144582718610764,
0.07548418641090393,
0.10937431454658508,
0.04338730871677399,
0.05375870317220688,
-0.0011587621411308646,
0.017814606428146362,
-0.005678045097738504,
-0.00040870788507163525,
0.02547398954629898,
-0.026136815547943115,
-3.986352226093004e-8,
-0.08116912096738815,
0.07124266028404236,
-0.04030511900782585,
0.006860488094389439,
0.12152402848005295,
0.0273475032299757,
-0.016259510070085526,
0.05866798013448715,
-0.04957104101777077,
0.03466784209012985,
0.11238864809274673,
-0.008027084171772003,
-0.04110328480601311,
-0.036822736263275146,
0.04270917922258377,
0.007203580811619759,
-0.022791706025600433,
0.011995772831141949,
-0.031369853764772415,
-0.03254726156592369,
0.004836069419980049,
-0.002574135782197118,
0.001083704992197454,
0.010960791260004044,
-0.00585897034034133,
-0.1017870306968689,
-0.06491849571466446,
0.0050050136633217335,
-0.00551440566778183,
0.03724032640457153,
0.00227167084813118,
0.02021961100399494,
-0.0274986382573843,
-0.02029014192521572,
0.001653384417295456,
0.049355652183294296,
-0.0009112819680012763,
-0.022639276459813118,
-0.019517986103892326,
-0.04783358797430992,
-0.03475325182080269,
0.003883110359311104,
-0.011108333244919777,
0.034628577530384064,
0.06502914428710938,
-0.003430906217545271,
-0.02241724729537964,
-0.06954268366098404,
-0.007134960498660803,
-0.013273670338094234,
0.04241875559091568,
0.01674039475619793,
-0.013174489140510559,
0.04447288438677788,
0.05761769786477089,
0.005914120934903622,
0.03205561265349388,
-0.06485908478498459,
0.0012544604251161218,
0.047064099460840225,
0.028483184054493904,
0.10358467698097229,
-0.0914069265127182,
-0.092034250497818
] |
Automated Design of Agentic Systems | https://openreview.net/forum?id=t9U3LW7JVX | [
"Shengran Hu",
"Cong Lu",
"Jeff Clune"
] | Poster | Researchers are investing substantial effort in developing powerful general-purpose agents, wherein Foundation Models are used as modules within agentic systems (e.g. Chain-of-Thought, Self-Reflection, Toolformer). However, the history of machine learning teaches us that hand-designed solutions are eventually replaced by learned solutions. We describe a newly forming research area, Automated Design of Agentic Systems (ADAS), which aims to automatically create powerful agentic system designs, including inventing novel building blocks and/or combining them in new ways. We further demonstrate that there is an unexplored yet promising approach within ADAS where agents can be defined in code and new agents can be automatically discovered by a meta agent programming ever better ones in code. Given that programming languages are Turing Complete, this approach theoretically enables the learning of any possible agentic system: including novel prompts, tool use, workflows, and combinations thereof. We present a simple yet effective algorithm named Meta Agent Search to demonstrate this idea, where a meta agent iteratively programs interesting new agents based on an ever-growing archive of previous discoveries. Through extensive experiments across multiple domains including coding, science, and math, we show that our algorithm can progressively invent agents with novel designs that greatly outperform state-of-the-art hand-designed agents. Importantly, we consistently observe the surprising result that agents invented by Meta Agent Search maintain superior performance even when transferred across domains and models, demonstrating their robustness and generality. Provided we develop it safely, our work illustrates the potential of an exciting new research direction toward automatically designing ever-more powerful agentic systems to benefit humanity. | LLMs, Language Model Agents, Agents, Agentic Systems, Reasoning, Meta Learning, Open-endedness | We describe a newly forming research area which aims to automatically create powerful agentic system designs and present an algorithm to demonstrate that agents can automatically designing themselves by programming in code. | 563 | 2408.08435 | [
-0.06624733656644821,
-0.03940233588218689,
-0.07349137216806412,
0.04550819844007492,
0.012624269351363182,
-0.09925033152103424,
0.005453215446323156,
-0.010863377712666988,
-0.06080832704901695,
0.024250559508800507,
-0.10292468965053558,
-0.050063151866197586,
0.006958425976336002,
0.055866654962301254,
0.04174401983618736,
0.02481982670724392,
0.04083671420812607,
0.04634856432676315,
-0.02865760587155819,
-0.03125093877315521,
0.006177043076604605,
0.043748993426561356,
-0.006231073755770922,
0.01328064501285553,
-0.06055363640189171,
0.061893612146377563,
0.01468359399586916,
-0.11046798527240753,
0.009847437031567097,
0.02301229164004326,
0.07230659574270248,
0.014786594547331333,
0.016621163114905357,
0.017936859279870987,
0.024372940883040428,
0.05816120281815529,
-0.026184767484664917,
0.09095947444438934,
0.010930068790912628,
-0.05673593655228615,
-0.1006450206041336,
-0.010320131666958332,
-0.01765054650604725,
-0.00632271496579051,
0.016336791217327118,
-0.026332123205065727,
-0.019444404169917107,
-0.03537045046687126,
0.02775442972779274,
0.00858672708272934,
-0.1371605098247528,
-0.07463067770004272,
-0.02773343212902546,
-0.017778217792510986,
0.009049881249666214,
-0.013217804953455925,
0.02236611768603325,
-0.038035355508327484,
-0.03578107804059982,
-0.0696089118719101,
0.00298651703633368,
-0.03890013322234154,
-0.02430878020823002,
-0.023325152695178986,
-0.058245494961738586,
0.027147574350237846,
-0.048362378031015396,
0.02051093988120556,
0.045048732310533524,
-0.030729731544852257,
-0.03709089756011963,
0.04448550567030907,
-0.0008727493695914745,
0.042919471859931946,
0.07366009801626205,
0.05513668432831764,
0.00584515742957592,
0.021298730745911598,
0.015809843316674232,
-0.05622163414955139,
-0.03783590346574783,
0.02003008872270584,
-0.006466432008892298,
0.13627007603645325,
0.035373229533433914,
0.01428403053432703,
0.059000976383686066,
0.051604460924863815,
0.057021185755729675,
0.011492458172142506,
-0.008629954420030117,
-0.03495758771896362,
0.06346103549003601,
-0.05056581273674965,
0.049029748886823654,
0.05732239782810211,
0.0008103790460154414,
-0.1592794805765152,
-0.02074078842997551,
0.10782384872436523,
-0.10222573578357697,
0.04006759077310562,
-0.022863255813717842,
-0.029138825833797455,
-0.04412100464105606,
0.031000200659036636,
0.02673800103366375,
0.03551451116800308,
0.064084991812706,
-0.06917594373226166,
-0.0006158838514238596,
0.03203500434756279,
0.034174926578998566,
-0.006790034007281065,
0.06730519235134125,
-0.003647229168564081,
-0.0136327613145113,
0.039595238864421844,
0.032122738659381866,
0.08416012674570084,
0.055773667991161346,
0.04169797897338867,
-0.04182763025164604,
0.057059045881032944,
0.08780563622713089,
-0.006088294088840485,
-0.023521456867456436,
7.067829679986896e-34,
0.03113825060427189,
0.0794273316860199,
0.04722658544778824,
0.09575923532247543,
0.06998192518949509,
-0.08607200533151627,
0.04533375799655914,
0.007845498621463776,
0.015425607562065125,
-0.002800186863169074,
0.075588159263134,
0.04089757800102234,
-0.03135433793067932,
0.06049301102757454,
-0.0064758602529764175,
-0.0039529139176011086,
-0.06215912848711014,
0.025649022310972214,
-0.02225256897509098,
-0.09199122339487076,
0.024823928251862526,
0.03620840981602669,
0.05502258241176605,
-0.002791492035612464,
0.03998008742928505,
0.028951792046427727,
-0.0083285728469491,
-0.03016630746424198,
0.06882636994123459,
0.011754506267607212,
-0.06124892830848694,
0.005528964102268219,
0.012125889770686626,
0.13175643980503082,
-0.08178725838661194,
0.015639301389455795,
-0.05277137830853462,
-0.07595781981945038,
0.042511820793151855,
0.05112069472670555,
0.008291044272482395,
0.029194984585046768,
0.04154047742486,
-0.014261893928050995,
-0.02057250589132309,
-0.08439023047685623,
-0.044334348291158676,
0.03398026525974274,
-0.025630049407482147,
-0.030266935005784035,
0.03822455182671547,
0.02201191708445549,
0.0020459031220525503,
0.02513805404305458,
-0.011095285415649414,
0.023215824738144875,
-0.04613042622804642,
0.00811684224754572,
0.052903082221746445,
0.027520479634404182,
0.021517232060432434,
-0.007940576411783695,
-0.0022559284698218107,
0.08646942675113678,
0.0014907321892678738,
0.0002751926949713379,
0.06004014238715172,
0.06608272343873978,
0.07693670690059662,
0.029506627470254898,
-0.051960404962301254,
0.049014411866664886,
-0.03062276542186737,
0.012668111361563206,
0.001768900896422565,
-0.05532624572515488,
0.11245426535606384,
-0.12357911467552185,
0.021293198689818382,
-0.061006952077150345,
-0.04442024603486061,
0.04920171946287155,
-0.01579580269753933,
-0.0035691920202225447,
-0.0005579171702265739,
-0.09255234152078629,
-0.05093487352132797,
-0.07137549668550491,
-0.07940240204334259,
-0.05493294447660446,
-0.14373911917209625,
-0.04998776689171791,
-0.08265295624732971,
0.05171862617135048,
0.02479853667318821,
-2.5743482403351565e-33,
-0.02283402532339096,
-0.11060781031847,
-0.08654502034187317,
0.017738351598381996,
-0.0854773297905922,
-0.0767928957939148,
-0.08049751818180084,
-0.05208144336938858,
0.0002225010539405048,
-0.004660145379602909,
-0.013167771510779858,
0.006396140903234482,
0.04984568431973457,
0.0541759692132473,
-0.01988823339343071,
-0.030978618189692497,
0.02254648692905903,
-0.04694581776857376,
0.015344444662332535,
0.05054923892021179,
-0.02264622040092945,
0.15912331640720367,
-0.07517705112695694,
-0.016369175165891647,
0.0070397439412772655,
-0.02968892827630043,
0.002744879573583603,
0.06440601497888565,
0.01556478813290596,
0.055414482951164246,
0.013029823079705238,
-0.0046797748655080795,
-0.017315125092864037,
-0.036372482776641846,
-0.015988193452358246,
0.09146365523338318,
0.03350169211626053,
-0.047248344868421555,
-0.0829259529709816,
0.015246540307998657,
0.050534702837467194,
-0.08211003988981247,
-0.02138660103082657,
-0.037673112004995346,
0.005534449592232704,
-0.02461731992661953,
-0.0565163716673851,
0.07881741225719452,
0.06076446548104286,
-0.06251589953899384,
-0.011790064163506031,
-0.03569910302758217,
-0.03909982368350029,
-0.09364420175552368,
-0.034329935908317566,
0.00010930436837952584,
0.07470966130495071,
0.025552254170179367,
0.04951666668057442,
0.04347805678844452,
-0.06555280089378357,
-0.033687155693769455,
0.06388085335493088,
-0.04016600176692009,
-0.0289448294788599,
-0.018195740878582,
-0.06281726062297821,
-0.03216763958334923,
-0.013644388876855373,
-0.007252814248204231,
0.06935582309961319,
-0.03793903812766075,
-0.09174729883670807,
0.02598385512828827,
0.03333462402224541,
0.009735414758324623,
-0.006654465105384588,
-0.03340791538357735,
-0.056757647544145584,
-0.04018881544470787,
-0.02211805060505867,
0.029308566823601723,
0.04830371215939522,
0.024364979937672615,
-0.006870926823467016,
0.04008055850863457,
-0.023826882243156433,
0.10015526413917542,
0.024050738662481308,
0.05442719906568527,
-0.026510225608944893,
0.02158580720424652,
0.018858440220355988,
0.08169630169868469,
-0.10537136346101761,
-3.843748430654159e-8,
-0.0324319563806057,
0.05523042753338814,
0.05250666290521622,
0.018397709354758263,
0.017613079398870468,
0.059270016849040985,
-0.042113009840250015,
0.0048891338519752026,
-0.0757453665137291,
-0.10494208335876465,
0.09571275115013123,
0.005512358155101538,
0.03134843334555626,
0.038936883211135864,
-0.0033305492252111435,
0.0768255665898323,
0.007654291577637196,
-0.05266429856419563,
-0.04337107390165329,
-0.07579648494720459,
0.02925860695540905,
-0.003912255633622408,
0.05786404386162758,
-0.046231709420681,
0.00369125884026289,
-0.07923011481761932,
-0.04514716938138008,
-0.013950148597359657,
0.019368942826986313,
0.11090819537639618,
0.040809668600559235,
-0.0057864137925207615,
0.06884636729955673,
-0.01976509392261505,
0.0804508775472641,
0.062271665781736374,
0.0013901813654229045,
-0.06594516336917877,
0.012623196467757225,
-0.12746532261371613,
0.014356995932757854,
0.08197808265686035,
-0.012702377513051033,
-0.0237959586083889,
0.03434060886502266,
-0.018628429621458054,
0.001768965390510857,
-0.05571925640106201,
0.04002010077238083,
-0.04244674742221832,
-0.03796093910932541,
-0.016145385801792145,
-0.05293794721364975,
0.040437113493680954,
0.046514127403497696,
0.048720382153987885,
-0.033668991178274155,
-0.052400656044483185,
0.08660139888525009,
0.02776213176548481,
-0.03690522164106369,
0.04687142372131348,
0.0413125678896904,
0.017478272318840027
] |
STORM: Spatio-TempOral Reconstruction Model For Large-Scale Outdoor Scenes | https://openreview.net/forum?id=M2NFWRPMUd | [
"Jiawei Yang",
"Jiahui Huang",
"Boris Ivanovic",
"Yuxiao Chen",
"Yan Wang",
"Boyi Li",
"Yurong You",
"Apoorva Sharma",
"Maximilian Igl",
"Peter Karkus",
"Danfei Xu",
"Yue Wang",
"Marco Pavone"
] | Poster | We present STORM, a spatio-temporal reconstruction model designed for reconstructing dynamic outdoor scenes from sparse observations. Existing dynamic reconstruction methods often rely on per-scene optimization, dense observations across space and time, and strong motion supervision, resulting in lengthy optimization times, limited generalization to novel views or scenes, and degenerated quality caused by noisy pseudo-labels for dynamics. To address these challenges, STORM leverages a data-driven Transformer architecture that directly infers dynamic 3D scene representations—parameterized by 3D Gaussians and their velocities—in a single forward pass. Our key design is to aggregate 3D Gaussians from all frames using self-supervised scene flows, transforming them to the target timestep to enable complete (i.e., "amodal") reconstructions from arbitrary viewpoints at any moment in time. As an emergent property, STORM automatically captures dynamic instances and generates high-quality masks using only reconstruction losses. Extensive experiments on public datasets show that STORM achieves precise dynamic scene reconstruction, surpassing state-of-the-art per-scene optimization methods (+4.3 to 6.6 PSNR) and existing feed-forward approaches (+2.1 to 4.7 PSNR) in dynamic regions. STORM reconstructs large-scale outdoor scenes in 200ms, supports real-time rendering, and outperforms competitors in scene flow estimation, improving 3D EPE by 0.422m and Acc5 by 28.02%. Beyond reconstruction, we showcase four additional applications of our model, illustrating the potential of self-supervised learning for broader dynamic scene understanding. For more details, please visit our project at https://jiawei-yang.github.io/STORM/. | autonomous driving; reconstruction model; spatiotemporal | We present STORM, a spatio-temporal reconstruction model designed to reconstruct space-time scenes and motions from sparse observations for outdoor scenes. | 561 | 2501.00602 | [
-0.031218411400914192,
-0.06585463136434555,
0.07617960125207901,
0.0433441698551178,
0.09581838548183441,
0.005374102853238583,
-0.019056566059589386,
-0.0839669480919838,
0.05517704039812088,
-0.09873733669519424,
-0.08326726406812668,
-0.05603530630469322,
-0.01195491198450327,
0.08109119534492493,
-0.045134883373975754,
0.0029736903961747885,
0.10958104580640793,
0.06246711313724518,
-0.06088956817984581,
0.004793807398527861,
-0.023989247158169746,
0.054405536502599716,
-0.0345463789999485,
0.04795820266008377,
0.10250819474458694,
0.017603859305381775,
0.01701909862458706,
-0.06863155961036682,
0.0038195557426661253,
0.007133473642170429,
0.06636619567871094,
-0.04066621512174606,
-0.039371635764837265,
0.012736791744828224,
-0.020338134840130806,
0.016044024378061295,
-0.002653353149071336,
0.05225441977381706,
-0.06366594135761261,
0.03342096880078316,
0.049551770091056824,
0.052724745124578476,
-0.001489014714024961,
-0.0015999311581254005,
0.05027322098612785,
0.1003870889544487,
0.08784905076026917,
-0.0779992863535881,
0.053560350090265274,
0.014277119189500809,
-0.04614628851413727,
0.011794629506766796,
-0.03465507552027702,
0.03387304022908211,
-0.07338724285364151,
0.0644061490893364,
0.005482176784425974,
-0.11730244010686874,
0.05102892965078354,
0.022605905309319496,
0.015964221209287643,
-0.009944786317646503,
0.03188139572739601,
0.016218123957514763,
-0.05590873211622238,
-0.003244438674300909,
0.008565586060285568,
0.01811360754072666,
0.011330882087349892,
0.044954635202884674,
-0.04215928167104721,
0.06244603171944618,
0.049931857734918594,
-0.037909235805273056,
-0.015367362648248672,
-0.013077257201075554,
0.048745930194854736,
-0.07667476683855057,
0.025351937860250473,
-0.053605590015649796,
0.08027703315019608,
-0.02263871766626835,
0.002349408343434334,
0.04833312705159187,
0.024961277842521667,
0.02725859172642231,
-0.02413252927362919,
0.1387697011232376,
0.0130746616050601,
0.038639236241579056,
-0.0947970300912857,
-0.06979928910732269,
-0.031243929639458656,
-0.03674565628170967,
0.022839199751615524,
-0.04449978470802307,
0.0331079438328743,
-0.10803066939115524,
0.07735906541347504,
0.0694742351770401,
-0.022022049874067307,
-0.06610970944166183,
0.07891932129859924,
0.0017261650646105409,
0.023606711998581886,
-0.021992893889546394,
0.0246568713337183,
0.012359806336462498,
0.004673062823712826,
0.02611694298684597,
0.07702629268169403,
0.0021013026125729084,
0.03795408084988594,
-0.002102241851389408,
0.09381558746099472,
0.021869488060474396,
-0.038192570209503174,
-0.037841569632291794,
-0.06802287697792053,
0.021606629714369774,
-0.008624075911939144,
-0.08999848365783691,
-0.0016414293786510825,
0.062282610684633255,
0.04465346783399582,
0.04226073622703552,
0.03533153608441353,
5.5746290526735955e-34,
-0.0013944581151008606,
0.06712301075458527,
0.02187037281692028,
0.0440918505191803,
0.04499660059809685,
0.05458866432309151,
0.043856825679540634,
0.007177973166108131,
-0.017275681719183922,
0.039561718702316284,
-0.05276631936430931,
-0.046647366136312485,
0.027552805840969086,
0.08362426608800888,
-0.011173802427947521,
-0.004851628560572863,
-0.04715641960501671,
0.03874961659312248,
-0.0146509213373065,
0.04553718492388725,
0.007535511162132025,
-0.015170156955718994,
-0.026265010237693787,
-0.039179831743240356,
-0.04653756320476532,
-0.047756120562553406,
0.06867069005966187,
0.014893039129674435,
0.029019426554441452,
-0.02022293023765087,
-0.05427245423197746,
0.014962265267968178,
0.0249489676207304,
0.027008790522813797,
0.048700690269470215,
-0.031060300767421722,
-0.048465706408023834,
-0.049627311527729034,
0.03099675290286541,
-0.028046108782291412,
-0.055039554834365845,
0.08515709638595581,
-0.11167225986719131,
-0.046219076961278915,
-0.08611887693405151,
-0.055881794542074203,
0.0794823169708252,
0.0956776812672615,
-0.0841575562953949,
0.012776483781635761,
0.11802634596824646,
0.020621268078684807,
-0.0519237257540226,
-0.05238950252532959,
-0.045838650315999985,
-0.026858896017074585,
0.0663355216383934,
0.000889866438228637,
0.05263395234942436,
0.04190373048186302,
0.003985097166150808,
0.05563829094171524,
-0.00041057795169763267,
-0.02838502638041973,
0.0383211113512516,
0.021740879863500595,
0.10278408229351044,
0.11843272298574448,
0.06618056446313858,
-0.03302806615829468,
-0.00959088932722807,
-0.05850409343838692,
-0.007671831175684929,
-0.04611857607960701,
0.062364544719457626,
-0.012040427885949612,
0.06900370866060257,
0.03421740233898163,
0.0117532629519701,
0.09152797609567642,
-0.12832750380039215,
-0.0442962720990181,
0.04505552351474762,
0.009695427492260933,
0.014189277775585651,
-0.0020970336627215147,
0.017093246802687645,
-0.013933684676885605,
-0.059150226414203644,
-0.061786238104104996,
0.0033928293269127607,
-0.03754908964037895,
-0.023935185745358467,
-0.014883078634738922,
-0.049630776047706604,
9.974541601177362e-34,
0.04105277359485626,
0.024644583463668823,
-0.09589593857526779,
0.020436540246009827,
-0.02667827159166336,
-0.09108664840459824,
-0.08693258464336395,
0.027980094775557518,
0.05094397813081741,
-0.09399203211069107,
-0.08778634667396545,
0.005076251458376646,
0.04827244207262993,
0.012699104845523834,
0.03875051438808441,
-0.03814762830734253,
-0.007395808584988117,
-0.032366085797548294,
-0.10247235745191574,
-0.050840917974710464,
0.038095299154520035,
0.04242624342441559,
-0.051896244287490845,
-0.034200556576251984,
-0.092562235891819,
0.11496702581644058,
-0.00638410821557045,
0.052643608301877975,
0.039763692766427994,
-0.004306439775973558,
-0.03452710434794426,
-0.013136778026819229,
0.005615100730210543,
-0.08151546120643616,
-0.0817921981215477,
0.0617758147418499,
0.048076286911964417,
-0.027185071259737015,
-0.020138490945100784,
-0.008903159759938717,
-0.00667321914806962,
-0.0046281940303742886,
0.00908257719129324,
0.07839822769165039,
-0.040561776608228683,
0.024602878838777542,
-0.00656958157196641,
0.03195459395647049,
0.011769005097448826,
0.04558827728033066,
-0.11207400262355804,
0.051920875906944275,
-0.11060068756341934,
0.05588021129369736,
-0.0006710104062221944,
-0.07224232703447342,
-0.05926856771111488,
-0.07334083318710327,
0.007121846079826355,
-0.04891533777117729,
-0.046783819794654846,
-0.04700908809900284,
-0.026781057938933372,
-0.02473551779985428,
0.004900043830275536,
-0.04879393056035042,
-0.009293624199926853,
-0.04866691678762436,
-0.06397254765033722,
0.09296239167451859,
0.014792826026678085,
-0.03206869214773178,
-0.028910791501402855,
0.0219045951962471,
0.0004388747620396316,
-0.023777101188898087,
0.00801099743694067,
0.03409244865179062,
0.026543904095888138,
0.010721339844167233,
0.04267225041985512,
0.001328327925875783,
0.024445487186312675,
0.05536514148116112,
0.05953449010848999,
0.04487406462430954,
0.012363525107502937,
-0.031733740121126175,
0.048043277114629745,
0.04022347554564476,
-0.01890408992767334,
0.0004059633065480739,
-0.029092533513903618,
0.044519487768411636,
-0.006888630799949169,
-3.9322468836644475e-8,
-0.07394558191299438,
0.09933370351791382,
0.0344139002263546,
-0.028464430943131447,
-0.047615740448236465,
-0.060419242829084396,
0.06071528047323227,
0.10347244143486023,
0.025435352697968483,
-0.050713036209344864,
0.06279928237199783,
-0.08153339475393295,
0.020252544432878494,
-0.010389043018221855,
-0.026132194325327873,
0.02308024652302265,
0.012762416154146194,
-0.0653340071439743,
-0.04214229807257652,
0.035219136625528336,
-0.0575350783765316,
0.020140018314123154,
-0.038980696350336075,
-0.010932013392448425,
0.04112200438976288,
0.03449316322803497,
0.02951633557677269,
0.014568856917321682,
0.08892814815044403,
0.000936203810852021,
-0.03360810503363609,
0.007157200016081333,
-0.016573337838053703,
-0.025207526981830597,
-0.08834436535835266,
0.023074829950928688,
0.01241922378540039,
0.06517591327428818,
0.018202507868409157,
-0.09042101353406906,
-0.05612177029252052,
0.04654468223452568,
-0.03788687661290169,
0.015479187481105328,
-0.03187764436006546,
0.12937518954277039,
-0.034315261989831924,
-0.07959825545549393,
0.019184233620762825,
0.005286579020321369,
0.015696387737989426,
-0.0359821543097496,
-0.0026139169931411743,
0.08750741928815842,
0.07658706605434418,
0.052849773317575455,
0.03759133443236351,
-0.06610839068889618,
0.08041483908891678,
0.008790637366473675,
-0.0824449211359024,
0.018141554668545723,
-0.09099540114402771,
0.01476824190467596
] |
Training Language Models on Synthetic Edit Sequences Improves Code Synthesis | https://openreview.net/forum?id=AqfUa08PCH | [
"Ulyana Piterbarg",
"Lerrel Pinto",
"Rob Fergus"
] | Poster | Software engineers mainly write code by editing existing programs. In contrast, language models (LMs) autoregressively synthesize programs in a single pass. One explanation for this is the scarcity of sequential edit data. While high-quality instruction data for code synthesis is scarce, edit data for synthesis is even scarcer. To fill this gap, we develop a synthetic data generation algorithm called LintSeq. This algorithm refactors programs into sequences of synthetic edits by using a linter to procedurally sample across interdependent lines of source code. Synthetic edits sampled with LintSeq reflect the syntax and semantics of their programming language. To test the algorithm, we use it to refactor a dataset of instruction + program pairs into instruction + program-diff-sequence tuples. Then, we fine-tune a series of smaller LMs ranging from 2.6B to 14B parameters on both the re-factored and original versions of this dataset. We perform comprehensive evaluations comparing edit sequence code LMs against baselines on HumanEval, MBPP(+), CodeContests, DS-1000, and BigCodeBench. We show that models fine-tuned to iteratively synthesize code match or outperform baselines on pass@1, and exhibit better scaling across higher pass@k as a function of total test-time FLOPs. Finally, we also pretrain our own tiny LMs for code understanding. We show that fine-tuning these models to synthesize code edit-by-edit results in strong performance on HumanEval and MBPP(+) compared to existing code language models of similar scale such as CodeT5+, AlphaCode, and Codex. | language model, code synthesis, reasoning, synthetic data | Training LMs to synthesize code edit-by-edit with SFT improves the slope of test-time scaling laws | 560 | 2410.02749 | [
-0.09676683694124222,
-0.04470881447196007,
0.0076913596130907536,
0.028106747195124626,
0.014149582013487816,
-0.02526119537651539,
-0.08537191152572632,
-0.003938205074518919,
-0.046599652618169785,
0.0037190988659858704,
-0.05506938695907593,
-0.03541239723563194,
0.0020176060497760773,
-0.09112156927585602,
-0.008627967908978462,
0.026504695415496826,
-0.01174868643283844,
0.028122678399086,
-0.027645505964756012,
-0.059647127985954285,
0.021417822688817978,
0.0351821593940258,
-0.01911194436252117,
0.031658757477998734,
0.0836520716547966,
0.0187031589448452,
0.031089480966329575,
-0.03241460397839546,
0.06352609395980835,
-0.02760162390768528,
0.03393489867448807,
0.08372093737125397,
0.03622424975037575,
0.0619560144841671,
-0.003547902684658766,
0.06227625906467438,
-0.007139622699469328,
-0.015397603623569012,
-0.004150293301790953,
0.006768560968339443,
-0.10023021697998047,
0.04132644832134247,
0.04190424829721451,
-0.03179636970162392,
0.07093275338411331,
-0.041659388691186905,
-0.026347298175096512,
-0.07991715520620346,
-0.058177243918180466,
0.028790105134248734,
-0.08492461591959,
-0.028163937851786613,
-0.07595554739236832,
0.03500816226005554,
-0.0029547002632170916,
-0.048156578093767166,
0.09092546999454498,
-0.016851967200636864,
0.039519816637039185,
-0.03356649726629257,
-0.04582265764474869,
-0.0028614304028451443,
-0.022870054468512535,
-0.008265355601906776,
-0.07764758914709091,
-0.02581878751516342,
0.0162747111171484,
0.054113876074552536,
0.02695477567613125,
-0.030243676155805588,
-0.10395734757184982,
0.0738598182797432,
-0.06504227966070175,
0.07702894508838654,
-0.04450579732656479,
0.030180858448147774,
-0.012859586626291275,
0.031596045941114426,
0.01348430197685957,
-0.13012632727622986,
0.020103394985198975,
-0.004154975060373545,
-0.020478786900639534,
0.028141435235738754,
0.03345606476068497,
0.011692643165588379,
0.11135152727365494,
0.07560750842094421,
0.0758533850312233,
0.10334650427103043,
0.009110772982239723,
0.010074302554130554,
0.04497236758470535,
0.018964817747473717,
0.031974442303180695,
0.05269894748926163,
0.028830304741859436,
-0.018520578742027283,
0.05068483576178551,
0.07387363165616989,
0.008619801141321659,
0.0996849536895752,
-0.020689405500888824,
-0.054888345301151276,
0.007440629415214062,
-0.029811561107635498,
0.13238206505775452,
-0.027011768892407417,
0.005950396880507469,
-0.06510999798774719,
0.044039662927389145,
0.04890168458223343,
0.02192479558289051,
-0.026335181668400764,
0.010923927649855614,
-0.023125478997826576,
0.02354256995022297,
-0.038334134966135025,
-0.05196763202548027,
0.04715140908956528,
0.04600485786795616,
0.04841563478112221,
-0.0007031421409919858,
0.04721163585782051,
-0.0378832072019577,
-0.03474479541182518,
-0.0527939535677433,
2.569453592060928e-33,
0.09803951531648636,
0.031499963253736496,
0.011379565112292767,
0.02172606624662876,
0.03335368260741234,
-0.026025809347629547,
0.046030811965465546,
0.04565298184752464,
-0.07150810211896896,
0.008330196142196655,
0.04110144078731537,
-0.010581137612462044,
-0.06684594601392746,
0.11391643434762955,
0.07996130734682083,
0.05781504511833191,
-0.026725472882390022,
0.003021303564310074,
-0.07634112983942032,
-0.002480476861819625,
-0.01628696545958519,
-0.013497128151357174,
-0.04231519252061844,
-0.059958744794130325,
0.004783444106578827,
0.06050502508878708,
0.011066796258091927,
-0.02125585451722145,
-0.006000711116939783,
0.0030788483563810587,
-0.11160552501678467,
-0.04699255898594856,
0.033034637570381165,
0.08105122298002243,
0.027777979150414467,
-0.040377553552389145,
0.021111737936735153,
-0.08489858359098434,
0.01948099583387375,
0.031875111162662506,
0.049534451216459274,
0.025878820568323135,
0.013013866730034351,
0.035784900188446045,
-0.007458879146724939,
-0.06647947430610657,
-0.03703759238123894,
0.026888292282819748,
0.02206110768020153,
-0.036981236189603806,
0.044794581830501556,
0.1193360984325409,
-0.020044509321451187,
-0.013202972710132599,
-0.05796751007437706,
0.06617243587970734,
0.026178769767284393,
-0.029644185677170753,
0.008379466831684113,
0.1364549994468689,
-0.021839037537574768,
0.09283820539712906,
0.012641247361898422,
0.04007987305521965,
-0.010111766867339611,
0.03664551302790642,
-0.025323018431663513,
0.03635513782501221,
0.06864942610263824,
-0.0009409633930772543,
-0.06759457290172577,
-0.03684959560632706,
-0.044399529695510864,
0.050301942974328995,
0.03904381021857262,
-0.07748141139745712,
0.01759662851691246,
-0.04644761607050896,
-0.025802597403526306,
-0.07432763278484344,
0.032300956547260284,
0.06035960093140602,
0.014064713381230831,
-0.03739446401596069,
-0.03938627243041992,
-0.13246792554855347,
-0.0011526565067470074,
0.015302465297281742,
-0.02758186310529709,
-0.01030130498111248,
-0.06439750641584396,
-0.07755865156650543,
-0.003788466565310955,
-0.06658726930618286,
0.03972891718149185,
-3.203969279587475e-33,
-0.000595874444115907,
-0.00938314851373434,
-0.021944977343082428,
0.1086219996213913,
-0.11808498203754425,
-0.08913154900074005,
-0.01135470811277628,
0.03142553195357323,
0.0408974252641201,
-0.023829469457268715,
-0.015262841247022152,
-0.06151992827653885,
0.09850118309259415,
-0.03531598299741745,
0.05581061542034149,
-0.07094638049602509,
-0.023362720385193825,
-0.00994269922375679,
-0.018426159396767616,
0.04470585659146309,
0.046466391533613205,
0.07948245853185654,
0.018747922033071518,
-0.02694641426205635,
-0.050107818096876144,
0.016702959313988686,
-0.052993401885032654,
0.10260723531246185,
0.023270893841981888,
-0.023458167910575867,
0.010676585137844086,
0.04534009471535683,
-0.0386081337928772,
-0.011500840075314045,
-0.03897107020020485,
-0.04474620521068573,
0.10021893680095673,
-0.00836259126663208,
-0.048137638717889786,
0.023981712758541107,
0.08536814153194427,
-0.010546991601586342,
-0.016588369384407997,
0.05786440148949623,
0.014643696136772633,
-0.006625113543123007,
-0.12864674627780914,
-0.020992441102862358,
0.03626853972673416,
-0.0666777715086937,
-0.003901864169165492,
0.027548447251319885,
-0.03687454015016556,
-0.0012858319096267223,
0.037369225174188614,
-0.11786390841007233,
-0.05502910539507866,
-0.059003036469221115,
0.02003195509314537,
-0.07128667086362839,
-0.04842018708586693,
-0.011884213425219059,
0.0378035269677639,
-0.10377246886491776,
0.04461167752742767,
0.005897506605833769,
-0.08334022015333176,
-0.039142608642578125,
-0.001368892495520413,
-0.016675520688295364,
0.078463114798069,
-0.05452955141663551,
-0.01522930059581995,
0.01802954263985157,
-0.02706591971218586,
-0.05994805693626404,
0.03041711449623108,
-0.05255324766039848,
0.0019502362702041864,
-0.007608856074512005,
0.013138556852936745,
-0.021311912685632706,
0.057876959443092346,
0.0034198567736893892,
-0.07513422518968582,
0.11443609744310379,
0.009251548908650875,
0.09095470607280731,
-0.03402816504240036,
0.035417765378952026,
-0.07244065403938293,
0.06222573667764664,
0.005875346250832081,
-0.0020653046667575836,
-0.07123114168643951,
-3.969067208231536e-8,
0.02530273236334324,
-0.009293274953961372,
-0.003542377846315503,
0.0710284411907196,
0.001698397332802415,
-0.036125171929597855,
-0.08140768110752106,
0.06015290319919586,
-0.009369132108986378,
-0.04590765759348869,
0.050375886261463165,
0.047312866896390915,
-0.029082728549838066,
-0.05216502398252487,
-0.008637689985334873,
0.09950190037488937,
-0.025721771642565727,
-0.04211787134408951,
-0.008081438019871712,
0.045836739242076874,
-0.024276187643408775,
0.12437670677900314,
0.0069769336842000484,
0.029896000400185585,
0.04310419410467148,
-0.07035414129495621,
0.06193074211478233,
0.06068646162748337,
0.07527930289506912,
0.05561281740665436,
0.048265159130096436,
-0.004677469376474619,
-0.027486618608236313,
-0.029168877750635147,
0.01795131154358387,
-0.033782027661800385,
0.06927178055047989,
-0.003945028409361839,
0.05889737233519554,
-0.045025717467069626,
0.04792623594403267,
0.016437605023384094,
-0.08029592037200928,
0.006965763866901398,
-0.00013248373579699546,
-0.1272144466638565,
-0.03654364496469498,
-0.11077708750963211,
0.013379370793700218,
-0.0958096832036972,
0.010024401359260082,
-0.01064880471676588,
-0.012812969274818897,
-0.05659864470362663,
0.014329935424029827,
0.06988955289125443,
-0.11153510212898254,
-0.015987098217010498,
-0.003203743137419224,
-0.001972464146092534,
-0.008692569099366665,
-0.013693070039153099,
0.06274142116308212,
0.02697400003671646
] |
Learning under Temporal Label Noise | https://openreview.net/forum?id=5o0phqAhsP | [
"Sujay Nagaraj",
"Walter Gerych",
"Sana Tonekaboni",
"Anna Goldenberg",
"Berk Ustun",
"Thomas Hartvigsen"
] | Poster | Many time series classification tasks, where labels vary over time, are affected by label noise that also varies over time. Such noise can cause label quality to improve, worsen, or periodically change over time. We first propose and formalize temporal label noise, an unstudied problem for sequential classification of time series. In this setting, multiple labels are recorded over time while being corrupted by a time-dependent noise function. We first demonstrate the importance of modeling the temporal nature of the label noise function and how existing methods will consistently underperform. We then propose methods to train noise-tolerant classifiers by estimating the temporal label noise function directly from data. We show that our methods lead to state-of-the-art performance under diverse types of temporal label noise on real-world datasets. | label noise; time series; healthcare; classification | null | 549 | 2402.04398 | [
-0.08904413133859634,
-0.04473380744457245,
0.02251248247921467,
0.009432381018996239,
0.031227441504597664,
0.0003247660642955452,
0.014764332212507725,
-0.10600561648607254,
-0.012266959063708782,
-0.13022209703922272,
-0.09822189807891846,
-0.019206205382943153,
-0.03866586089134216,
0.06278000771999359,
-0.06902282685041428,
-0.02695237100124359,
0.09109346568584442,
0.02714104950428009,
0.030681315809488297,
-0.010891103185713291,
-0.003940229304134846,
0.07274306565523148,
0.023712417110800743,
0.10345886647701263,
0.0885404497385025,
0.013512194156646729,
0.0006937807193025947,
0.005238404497504234,
0.01709039881825447,
0.018416287377476692,
-0.03276664763689041,
-0.008509178645908833,
0.007467813789844513,
-0.04318128526210785,
-0.03621123358607292,
-0.05434693023562431,
-0.03511534258723259,
0.047194577753543854,
0.03253340721130371,
-0.0047542136162519455,
0.059609200805425644,
0.023882871493697166,
0.005345174577087164,
0.03104572929441929,
-0.04012306034564972,
0.02556983008980751,
0.12551957368850708,
-0.11129976063966751,
0.004481129813939333,
0.10414589196443558,
-0.05709954723715782,
0.006551800295710564,
0.006383793894201517,
0.044770248234272,
0.0076353587210178375,
0.053669124841690063,
0.051492802798748016,
0.0484861321747303,
-0.002691195346415043,
0.08057199418544769,
0.056207023561000824,
-0.021801842376589775,
-0.0030474758241325617,
-0.02459133230149746,
-0.034673627465963364,
0.026263410225510597,
-0.006107366643846035,
0.0969817116856575,
-0.033883750438690186,
0.05855654925107956,
-0.016676943749189377,
0.11416667699813843,
-0.07668644189834595,
0.07481826841831207,
-0.0004590855387505144,
0.060978882014751434,
0.0026262220926582813,
-0.03465643525123596,
0.062193457037210464,
-0.039511099457740784,
-0.01223656628280878,
-0.008497127331793308,
0.017565052956342697,
-0.05203316733241081,
0.16860269010066986,
-0.03871899098157883,
0.07123342156410217,
0.03380371630191803,
-0.06121803820133209,
-0.024364888668060303,
-0.044081419706344604,
-0.021370895206928253,
0.08922313898801804,
-0.027668211609125137,
-0.06830241531133652,
0.017966682091355324,
-0.005122902803122997,
0.015871306881308556,
0.08208420872688293,
0.06713329255580902,
-0.1100284680724144,
0.015364512801170349,
-0.027859549969434738,
0.01113960798829794,
0.08021509647369385,
0.002363941865041852,
0.015549459494650364,
-0.04661291092634201,
0.03608573600649834,
-0.03207353129982948,
-0.027950027957558632,
0.006873276550322771,
-0.03537004068493843,
-0.008032516576349735,
0.06890720874071121,
0.0051053110510110855,
-0.0008400174556300044,
-0.04323095455765724,
-0.0070811123587191105,
0.06152001768350601,
-0.012271680869162083,
-0.04510407894849777,
-0.05357476323843002,
-0.0024726158007979393,
-0.00037074708961881697,
0.037615206092596054,
0.04198217764496803,
2.5065246040156106e-33,
0.03568177670240402,
-0.0438612625002861,
0.047820258885622025,
-0.0023616270627826452,
0.053769465535879135,
-0.06713412702083588,
-0.09869474172592163,
-0.10829193890094757,
0.10845845937728882,
0.05315608158707619,
0.046950653195381165,
-0.05108640342950821,
-0.042944591492414474,
-0.011750715784728527,
0.025662921369075775,
-0.01906573958694935,
0.009108755737543106,
0.07979831099510193,
0.04044928029179573,
-0.03582927584648132,
0.016705142334103584,
-0.030194757506251335,
0.02942614257335663,
-0.07688125222921371,
0.08167597651481628,
0.015317041426897049,
0.0307932086288929,
0.02725529670715332,
0.06215670332312584,
-0.0071686040610075,
-0.002689538523554802,
-0.04210931807756424,
0.0873916745185852,
0.034024983644485474,
0.015436585061252117,
-0.034657612442970276,
0.006609499454498291,
0.05920151621103287,
0.004506601952016354,
-0.045583806931972504,
-0.006119881756603718,
0.0037943199276924133,
0.00015161458577495068,
-0.036098506301641464,
0.022802473977208138,
0.013537977822124958,
0.019383007660508156,
-0.03388600796461105,
-0.018182072788476944,
0.004341009072959423,
-0.03536306321620941,
-0.039667341858148575,
0.0011896394426003098,
-0.02163957804441452,
-0.03887192904949188,
0.04380796104669571,
0.027044622227549553,
-0.0320456326007843,
-0.07791563868522644,
-0.0006929935771040618,
-0.023243775591254234,
0.043355587869882584,
0.00934296753257513,
-0.04184052348136902,
-0.013235609978437424,
0.08893375098705292,
0.056323058903217316,
0.06360463798046112,
0.0070109134539961815,
-0.0876661017537117,
0.05877046287059784,
0.03086075559258461,
-0.06192942336201668,
-0.05490785837173462,
0.05579167604446411,
0.02940497174859047,
0.06091761216521263,
-0.09305045753717422,
-0.03675871714949608,
0.06756357848644257,
-0.026677418500185013,
-0.15350745618343353,
0.08482373505830765,
-0.014508678577840328,
-0.0029205887112766504,
-0.016419509425759315,
0.04629914462566376,
-0.022534480318427086,
-0.005241299979388714,
0.027375182136893272,
-0.13025924563407898,
0.00950892735272646,
0.0027034005615860224,
0.012991460040211678,
-0.033999066799879074,
-2.583258303843397e-33,
-0.019053004682064056,
-0.029308034107089043,
0.03450290113687515,
0.04805077612400055,
0.016990089789032936,
-0.010515126399695873,
-0.09943151473999023,
0.0830245092511177,
0.04759383201599121,
0.03631297126412392,
-0.022317929193377495,
-0.10219946503639221,
0.0012839590199291706,
-0.035518694669008255,
-0.053698569536209106,
-0.05174374207854271,
0.011855743825435638,
-0.01106969639658928,
-0.048486609011888504,
0.04017812758684158,
0.07083814591169357,
0.052682265639305115,
-0.07331664115190506,
-0.03558073937892914,
-0.023694775998592377,
0.011068183928728104,
0.010980097576975822,
0.09866024553775787,
0.006678874604403973,
-0.047300953418016434,
-0.08157090097665787,
-0.0432126484811306,
0.060296449810266495,
-0.07872681319713593,
-0.056331805884838104,
-0.01608939841389656,
0.06396713852882385,
-0.06250011175870895,
-0.01512580830603838,
0.05478290840983391,
0.09311827272176743,
0.009542406536638737,
-0.06321857124567032,
-0.0805581733584404,
-0.06562996655702591,
-0.026528343558311462,
-0.09268853813409805,
0.05925307422876358,
-0.04903653264045715,
0.010487426072359085,
0.0014328109100461006,
0.047412361949682236,
-0.08044526726007462,
0.04661237820982933,
0.011162214912474155,
-0.043721772730350494,
-0.0745144933462143,
-0.0290356632322073,
0.030400924384593964,
0.09265519678592682,
-0.08894889801740646,
0.06017910689115524,
-0.013099809177219868,
-0.03266302868723869,
0.01905081793665886,
-0.0014271099353209138,
0.0023615567479282618,
-0.041739147156476974,
0.0292049590498209,
0.010929947718977928,
0.1571640819311142,
0.057335518300533295,
-0.10246707499027252,
0.010272625833749771,
-0.07386747747659683,
0.0017763207433745265,
-0.013194925151765347,
-0.025030674412846565,
0.02762605994939804,
-0.06887835264205933,
0.037941962480545044,
-0.03754947707056999,
0.02272154949605465,
0.10326148569583893,
-0.028672313317656517,
0.07205633074045181,
0.05389009416103363,
0.01537388376891613,
0.048063598573207855,
-0.05080892890691757,
0.006402801256626844,
-0.02075592428445816,
-0.05575314909219742,
0.09476382285356522,
-0.03829754889011383,
-3.329740394519831e-8,
-0.039126332849264145,
0.06208217889070511,
-0.011939975433051586,
-0.031705714762210846,
0.02634863555431366,
-0.01196309458464384,
0.04167016223073006,
0.06722631305456161,
0.019474387168884277,
-0.04227243363857269,
0.12081912159919739,
-0.033964164555072784,
-0.04191729053854942,
-0.04458325356245041,
-0.08607251197099686,
0.00724415248259902,
-0.017472725361585617,
0.023636629804968834,
0.011757373809814453,
-0.022608108818531036,
-0.0022362060844898224,
0.028149936348199844,
0.0666767805814743,
-0.05720800533890724,
0.010875580832362175,
-0.04044056311249733,
0.007821732200682163,
0.08973202854394913,
0.07969960570335388,
0.01579374074935913,
0.025665996596217155,
0.05638093128800392,
-0.03969334065914154,
-0.04353374242782593,
-0.05695556849241257,
-0.00975676067173481,
-0.028400029987096786,
-0.03950027748942375,
-0.049593932926654816,
0.037189532071352005,
-0.06373173743486404,
0.05162271484732628,
-0.07234605401754379,
-0.0018159924075007439,
-0.0148860402405262,
0.0056756590493023396,
-0.001882429700344801,
-0.08051438629627228,
0.030925361439585686,
-0.04204867407679558,
0.005068319849669933,
0.004507944919168949,
0.050109174102544785,
-0.04559773579239845,
0.10128564387559891,
0.04536443576216698,
-0.029123356565833092,
-0.014099149033427238,
-0.03719833865761757,
0.05544140562415123,
0.011442781426012516,
-0.013680454343557358,
-0.035598598420619965,
0.03053293190896511
] |
Subsets and Splits