VIsualizing and understanding convolutional networks(ECCV 2010)

《VIsualizing and understanding convolutional networks(ECCV 2010)》由會員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。

1、VisualizingandUnderstandingConvolutionalNetworksMatthewD.ZeilerandRobFergusDept.ofComputerScience,NewYorkUniversity,USA{zeiler,fergus}@cs.nyu.eduAbstract.LargeConvolutionalNetworkmodelshaverecentlydemon-stratedimpressiveclassi?cationperformanceontheImageNetbench-markKrizhevskyeta

2、l.[18].Howeverthereisnoclearunderstandingofwhytheyperformsowell,orhowtheymightbeimproved.Inthispaperweexplorebothissues.Weintroduceanovelvisualizationtechniquethatgivesinsightintothefunctionofintermediatefeaturelayersandtheoper-ationoftheclassi?er.Usedinadiagnosticrole,thesevisua

3、lizationsallowusto?ndmodelarchitecturesthatoutperformKrizhevskyetal.ontheImageNetclassi?cationbenchmark.Wealsoperformanablationstudytodiscovertheperformancecontributionfromdi?erentmodellayers.WeshowourImageNetmodelgeneralizeswelltootherdatasets:whenthesoftmaxclassi?erisretrained,

4、itconvincinglybeatsthecurrentstate-of-the-artresultsonCaltech-101andCaltech-256datasets.1IntroductionSincetheirintroductionbyLeCunetal.[20]intheearly1990’s,ConvolutionalNetworks(convnets)havedemonstratedexcellentperformanceattaskssuchashand-writtendigitclassi?cationandfacedetecti

5、on.Inthelast18months,sev-eralpapershaveshownthattheycanalsodeliveroutstandingperformanceonmorechallengingvisualclassi?cationtasks.Ciresanetal.[4]demonstratestate-of-the-artperformanceonNORBandCIFAR-10datasets.Mostnotably,Krizhevskyetal.[18]showrecordbeatingperformanceontheImageNe

6、t2012classi?cationbenchmark,withtheirconvnetmodelachievinganerrorrateof16.4%,comparedtothe2ndplaceresultof26.1%.Followingonfromthiswork,Girshicketal.[10]haveshownleadingdetectionperformanceonthePASCALVOCdataset.Sev-eralfactorsareresponsibleforthisdramaticimprovementinperformance:

7、(i)theavailabilityofmuchlargertrainingsets,withmillionsoflabeledexamples;(ii)powerfulGPUimplementations,makingthetrainingofverylargemodelspracti-caland(iii)bettermodelregularizationstrategies,suchasDropout[14].Despitethisencouragingprogress,thereisstilllittleinsightintotheinterna

8、loperationandbehaviorofthesecomplexmodel