VIsualizing and understanding convolutional networks(ECCV 2010)

《VIsualizing and understanding convolutional networks(ECCV 2010)》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、VisualizingandUnderstandingConvolutionalNetworksMatthewD.ZeilerandRobFergusDept.ofComputerScience,NewYorkUniversity,USA{zeiler,fergus}@cs.nyu.eduAbstract.LargeConvolutionalNetworkmodelshaverecentlydemon-stratedimpressiveclassificationperformanceontheImageNetbench-markKrizhevskyeta

2、l.[18].Howeverthereisnoclearunderstandingofwhytheyperformsowell,orhowtheymightbeimproved.Inthispaperweexplorebothissues.Weintroduceanovelvisualizationtechniquethatgivesinsightintothefunctionofintermediatefeaturelayersandtheoper-ationoftheclassifier.Usedinadiagnosticrole,thesevisua

3、lizationsallowustofindmodelarchitecturesthatoutperformKrizhevskyetal.ontheImageNetclassificationbenchmark.Wealsoperformanablationstudytodiscovertheperformancecontributionfromdifferentmodellayers.WeshowourImageNetmodelgeneralizeswelltootherdatasets:whenthesoftmaxclassifierisretrained,

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

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

6、t2012classificationbenchmark,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