资源描述：
《Understanding Neural Networks through deep Visua....pdf》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、UnderstandingNeuralNetworksThroughDeepVisualizationJasonYosinskiYOSINSKI@CS.CORNELL.EDUCornellUniversityJeffCluneJEFFCLUNE@UWYO.EDUAnhNguyenANGUYEN8@UWYO.EDUUniversityofWyomingThomasFuchsFUCHS@CALTECH.EDUJetPropulsionLaboratory,CaliforniaInstituteofTechn
2、ologyHodLipsonHOD.LIPSON@CORNELL.EDUCornellUniversityAbstract1.IntroductionThelastseveralyearshaveproducedtremendousprogressRecentyearshaveproducedgreatadvancesinintrainingpowerful,deepneuralnetworkmodelsthataretraininglarge,deepneuralnetworks(DNNs),in-a
3、pproachingandevensurpassinghumanabilitiesonavari-cludingnotablesuccessesintrainingconvolu-etyofchallengingmachinelearningtasks(Taigmanetal.,tionalneuralnetworks(convnets)torecognize2014;Schroffetal.,2015;Hannunetal.,2014).Aflagshipnaturalimages.However,ou
4、runderstandingofexampleistrainingdeep,convolutionalneuralnetworkshowthesemodelswork,especiallywhatcompu-(CNNs)withsupervisedlearningtoclassifynaturalimagestationstheyperformatintermediatelayers,has(Krizhevskyetal.,2012).Thatareahasbenefittedfromthelaggedb
5、ehind.Progressinthefieldwillbecombinedeffectsoffastercomputing(e.g.GPUs),betterfurtheracceleratedbythedevelopmentofbet-trainingtechniques(e.g.dropout(Hintonetal.,2012)),bet-tertoolsforvisualizingandinterpretingneuralteractivationunits(e.g.rectifiedlinearun
6、its(Glorotetal.,nets.Weintroducetwosuchtoolshere.The2011)),andlargerlabeleddatasets(Dengetal.,2009;Linfirstisatoolthatvisualizestheactivationspro-etal.,2014).ducedoneachlayerofatrainedconvnetasitprocessesanimageorvideo(e.g.aliveweb-Whiletherehasthusbeenco
7、nsiderableimprovementsincamstream).Wehavefoundthatlookingatliveourknowledgeofhowtocreatehigh-performingarchitec-activationsthatchangeinresponsetouserinputturesandlearningalgorithms,ourunderstandingofhowhelpsbuildvaluableintuitionsabouthowcon-theselargene
8、uralmodelsoperatehaslaggedbehind.Neu-vnetswork.Thesecondtoolenablesvisualizingralnetworkshavelongbeenknownas“blackboxes”be-arXiv:1506.06579v1[cs.CV]22Jun2015featuresateachlayerofaDNNviaregularizedcauseitisdifficulttounderst
