Whats-the-insight-of-self-paced-learning-孟德宇

DeyuMengXi’[e-mail]’stheInsightofSelf-pacedLearningProblemMachinelearninghasmadegoodprogressindatasetLFWwell-annotatedWhatwethinkwehave:Butwhatwereallyhaveisalways:ØAsimpleidea:pseudo-labelthedataandthenfeedthembackintotrainingModelTrainingPseudo-labelunsupervisedsamplesFeedpseudo-annotatedsamplesbackintotrainingØFavoringhigh-confidencesamples:Onlyfeedbacksomehigh-confidencesamplesØFromlesstomore:Withtheiteration,moresamplescanbeaddedSimilarstrategyhasbeenwidelyutilizedinSemi-supervisedlearningTransferlearningDomainadaptationBabylearningNeverendinglearningProblemModelLearningsamples:ØFromhigh-confidencetolowconfidenceØFromlesstomoreØHowhuman/animallearns:FirstinputeasysamplesandgraduallyinvolvemoreintotrainingfromeasytocomplexY.Bengio,J.Louradour,R.Collobert,andJ.Weston.Curriculumlearning.InICML,pages41–48,2009.CurriculumLearningCurriculumLearningØInsightfromcognitivescienceØMachinelearningalgorithmscanbenefitfromasimilartrainingstrategyØLearningfromeasieraspectsofthetask,andgraduallyincreasethedifficultylevelØExpectedtwoadvantages:•Helpfindabetterlocalminima(asaregularizer)•Speedtheconvergenceoftrainingtowardstheglobalminimum(forconvexproblem)ØBasicsteps:•Sortsamplesaccordingtocertain“easiness”measure•GraduallyaddsamplesintotrainingfromeasytocomplexSelf-pacedLearningM.P.Kumar,B.Packer,andD.Koller.Self-pacedlearningforlatentvariablemodels.InNIPS,pages1189–1197,2010.ØAlgorithm:AlternativesearchpFixw:pFixv:Astandardclassificationproblem.Self-pacedLearningØModel:SPLRegularizerØKoller’sSPLmodel:Øv’svalueisdeterminedbyaSPLregularizer:SPL:PrincipleProblemModelLearningsamples:ØFromhigh-confidencetolowconfidenceØFromlesstomoreØThreebasicprinciplesfordefiningaself-pacedregularizer:SPLRegularizer(LuJiang,DeyuMengetal. ACMMM,2014;QianZhao,DeyuMeng,etal.AAAI,2015)FavorsEasy(high-confidence)SamplesSPLRegularizer(LuJiang,DeyuMengetal. ACMMM,2014;QianZhao,DeyuMeng,etal.AAAI,2015)ØThreebasicprinciplesfordefiningaself-pacedregularizer:SPLRegularizer(LuJiang,DeyuMengetal. ACMMM,2014;QianZhao,DeyuMeng,etal.AAAI,2015)Whenthemodelisyoung,uselesssamples;whenthemodelismature,usemore.ØThreebasicprinciplesfordefiningaself-pacedregularizer:ConvexSPLRegularizer(LuJiang,DeyuMengetal. ACMMM,2014;QianZhao,DeyuMeng,etal.AAAI,2015)ØThreebasicprinciplesfordefiningaself-pacedregularizer:ApplicationPrincipleSPL:ProblemModelLearningsamples:ØFromhigh-confidencetolowconfidenceØFromlesstomoreThreeprinciplesforconstructingSPregularier:ØConvexØFromeasytohardØFromlesstomoreØSomesoftextensionsforself-pacedregularizer:SPLRegularizer(LuJiang,DeyuMengetal. ACMMM,2014;QianZhao,DeyuMeng,etal.AAAI,2015)LinearSoftWeighting:MixtureWeighting:LogarithmicSoftWeighting:ØSomesoftextensionsforself-pacedregularizerSPLRegularizer(LuJiang,DeyuMengetal. ACMMM,2014;QianZhao,DeyuMeng,etal.AAAI,2015)Moreextensions•SPaR:LuJiang,DeyuMeng,QianZhaoetal.ACMMM,2014.–SoftextensiononMEDEx0problem•SPMF:QianZhao,DeyuMeng,LuJiangetal.AAAI,2015.–Mixtureextensiononmatrixfactorization•SPLD:LuJiang,DeyuMeng,Shoou-IYuetal.NIPS,2014.–Diversityextensiononactionrecognition•SPCL:LuJiang,DeyuMeng,TerukoMitamuraetal.AAAI.2015.DingwenZhang,DeyuMeng,JunweiHan,IJCAI,2016LuJiang,DeyuMeng,AlexHauptmann,IJCAI2016–CurriculumextensiononMEDandmatrixfactorization•SP-MIL:DingwenZhang,DeyuMeng,JunweiHan.ICCV.2015.–Weaklysupervisedextensiononco-saliencydetection•MOSPL:AAAI2015–Multi-objectiveextensiononactionrecognition•ASPL:Inprocess(CooperatedwithLiangLin,WangmengZuo)–Activecurriculumextensiononfaceidentification•State-of-the-artperformanceon–WebQuerydataset–Hollywood2dataset–OlympicSportsdataset–iCosegdataset–MSRCdataset–YFCC100M–FCVID–TrecvidMEDEx0test2013–TrecvidMEDEx0test2014ApplicationPrincipleSPL:ProblemModelLearningsamples:ØFromhigh-confidencetolowconfidenceØFromlesstomoreThreeprinciplesforconstructingSPregularier:ØConvexØFromeasytohardØFromlesstomoreZero-ExampleSearch•Zero-ExampleSearch(alsoknownasEx0)representsamultimediasearchconditionwherezerorelevantexamplesareprovided–Content-basedsearch•Anexample:TRECVIDMultimediaEventDetection(MED)competition.Thetaskisverychallenging–Detectevery-dayeventinInternetvideos•Birthdayparty•Weddingceremony•ChangingavehicletireInformedia@CMU2013PipelineforEx0AninitialrankinglistofdataAnnotatepsudo-labelsfortop-rankeddataPickupthesetop-ranked(high-confidence)samplesandaddthemintotrainingsetRetraintheclassifierGraduallylowerthehigh-confidencethresholdSPaR:Self-PAcedReranking•Model(LuJiang,DeyuMengetal. ACMMM,2014)AlternativeoptimizationforsolvingSPaRInitializeaclassifierFixwandv,updateyFixwandy,updatevFixyandv,updatewGraduallyincreasetheageparameterØOnTRECVIDMED2013Ex0datasetØOnWebQuerydataset(Lu Jiang, Deyu Meng et al. ACM MM, 2014)TheoreticalinsightofSPLisstillentirelyunknown•Whyit’seffectiveinoutlier/heavynoisecases•Whereitconvergesto•What’sthetheoreticalinsightofSPLworkingmechanismApplicationPrincipleSPL:ProblemModelLearningsamples:ØFromhigh-confidencetolowconfidenceØFromlesstomoreThreeprin