资源描述:
《2012-ImageNet Classification with Deep Convolutional Neural Networks》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、ImageNetClassificationwithDeepConvolutionalNeuralNetworksAlexKrizhevskyIlyaSutskeverGeoffreyE.HintonUniversityofTorontoUniversityofTorontoUniversityofTorontokriz@cs.utoronto.cailya@cs.utoronto.cahinton@cs.utoronto.caAbstractWetrainedalarge,deepconvolutionalneuralnetworktocla
2、ssifythe1.2millionhigh-resolutionimagesintheImageNetLSVRC-2010contestintothe1000dif-ferentclasses.Onthetestdata,weachievedtop-1andtop-5errorratesof37.5%and17.0%whichisconsiderablybetterthanthepreviousstate-of-the-art.Theneuralnetwork,whichhas60millionparametersand650,000neu
3、rons,consistsoffiveconvolutionallayers,someofwhicharefollowedbymax-poolinglayers,andthreefully-connectedlayerswithafinal1000-waysoftmax.Tomaketrain-ingfaster,weusednon-saturatingneuronsandaveryefficientGPUimplemen-tationoftheconvolutionoperation.Toreduceoverfittinginthefully-co
4、nnectedlayersweemployedarecently-developedregularizationmethodcalled“dropout”thatprovedtobeveryeffective.WealsoenteredavariantofthismodelintheILSVRC-2012competitionandachievedawinningtop-5testerrorrateof15.3%,comparedto26.2%achievedbythesecond-bestentry.1IntroductionCurrent
5、approachestoobjectrecognitionmakeessentialuseofmachinelearningmethods.Toim-provetheirperformance,wecancollectlargerdatasets,learnmorepowerfulmodels,andusebet-tertechniquesforpreventingoverfitting.Untilrecently,datasetsoflabeledimageswererelativelysmall—ontheorderoftensofthou
6、sandsofimages(e.g.,NORB[16],Caltech-101/256[8,9],andCIFAR-10/100[12]).Simplerecognitiontaskscanbesolvedquitewellwithdatasetsofthissize,especiallyiftheyareaugmentedwithlabel-preservingtransformations.Forexample,thecurrent-besterrorrateontheMNISTdigit-recognitiontask(<0.3%)ap
7、proacheshumanperformance[4].Butobjectsinrealisticsettingsexhibitconsiderablevariability,sotolearntorecognizethemitisnecessarytousemuchlargertrainingsets.Andindeed,theshortcomingsofsmallimagedatasetshavebeenwidelyrecognized(e.g.,Pintoetal.[21]),butithasonlyrecentlybecomeposs
8、ibletocol-lectlabeleddatasetswithmillionsofimages.ThenewlargerdatasetsincludeLabel