WhenAllWeNeedisaPieceofthePie:AGenericFrameworkforOptimizingTwo-wayPartialAUCZhiyongYang12QianqianXu3ShilongBao12YuanHe4XiaochunCao12QingmingHuang3567Abstractframework.TheAreaUndertheROCCurve(AUC)i...
IsSpace-TimeAttentionAllYouNeedforVideoUnderstanding?GedasBertasius1HengWang1LorenzoTorresani12AbstractVideounderstandingsharesseveralhigh-levelsimilaritieswithNLP.Firstofall,videosandsentencesareb...
DoWeActuallyNeedDenseOver-Parameterization?In-TimeOver-ParameterizationinSparseTrainingShiweiLiu1LuYin1DecebalConstantinMocanu12MykolaPechenizkiy1AbstractFigure1.Asthefigureproceeds,weperformanOver...
AttentionisnotallyouNeed:pureattentionlosesrankdoublyexponentiallywithdepthYiheDong1Jean-BaptisteCordonnier2AndreasLoukas3Abstractattentionlayers.Surprisingly,wefindthatpureself-attentionnetworks(S...
DoWeReallyNeedtoAccesstheSourceData?SourceHypothesisTransferforUnsupervisedDomainAdaptationJianLiang1DapengHu1JiashiFeng1Abstractsegmentation(Zhangetal.,2017;Hoffmanetal.,2018),etc.Overthelastdecad...
DoWeNeedZeroTrainingLossAfterAchievingZeroTrainingError?TakashiIshida12IkkoYamane1TomoyaSakai3GangNiu2MasashiSugiyama21Abstract(a)w/oFlooding(b)w/FloodingOverparameterizeddeepnetworkshavethecapac-[...
ToUnderstandDeepLearningWeNeedtoUnderstandKernelLearning†MikhailBelkin1SiyuanMa1SoumikMandal1Abstractdeeplearningwillbedifficultuntilmoretractable“shallow”kernelmethodsarebetterunderstood.Genera...