[IJCAI 2011 Ciresan] Flexible, high performance convolutional neural networks for image classification

《[IJCAI 2011 Ciresan] Flexible, high performance convolutional neural networks for image classification》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、ProceedingsoftheTwenty-SecondInternationalJointConferenceonArtificialIntelligenceFlexible,HighPerformanceConvolutionalNeuralNetworksforImageClassificationDanC.Cires¸an,UeliMeier,JonathanMasci,LucaM.Gambardella,JurgenSchmidhuber¨IDSIA,USIandSUPSIGalleria2,6928Manno-Lu

2、gano,Switzerland{dan,ueli,jonathan,luca,juergen}@idsia.chAbstract(CNNs)[LeCunetal.,1998;Behnke,2003;Simardetal.,2003],whoseweights(filters)arerandomlyinitializedandWepresentafast,fullyparameterizableGPUim-changedinasupervisedwayusingback-propagation(BP).plementationo

3、fConvolutionalNeuralNetworkDespitethehardwareprogressofthepastdecades,compu-variants.Ourfeatureextractorsareneithercare-tationalspeedisstillalimitingfactorforCNNarchitecturesfullydesignednorpre-wired,butratherlearnedincharacterizedbymanybuildingblockstypicallysetbyt

4、rialasupervisedway.Ourdeephierarchicalarchitec-anderror.Tosystematicallytesttheimpactofvariousarchi-turesachievethebestpublishedresultsonbench-tecturesonclassificationperformance,wepresentafastCNNmarksforobjectclassification(NORB,CIFAR10)implementationonGraphicsProces

5、singUnits(GPUs).Previ-andhandwrittendigitrecognition(MNIST),withousGPUimplementationsofCNNs[Chellapillaetal.,2006;errorratesof2.53%,19.51%,0.35%,respectively.UetzandBehnke,2009;Strigletal.,2010]werehard-codedDeepnetstrainedbysimpleback-propagationper-tosatisfyGPUhar

6、dwareconstraintsorusegeneralpurposeformbetterthanmoreshallowones.Learningislibraries,whereasourimplementationisflexibleandfullyon-surprisinglyrapid.NORBiscompletelytrainedline(i.e.,weightupdatesaftereachimage).Anotableexcep-withinfiveepochs.TesterrorratesonMNISTtionis

7、[Jarrettetal.,2009]whoperformedathoroughanaly-dropto2.42%,0.97%and0.48%after1,3and17sisoftheinfluenceofallbuildingblocksofamultistagear-epochs,respectively.chitectureonrecognitionperformance.OurimplementationallowsfortraininglargeCNNswithindaysinsteadofmonths,1Introd

8、uctionsuchthatwecaninvestigatetheinfluenceofvariousstructuralThehumanvisualsystemefficientlyrecognizesandlocal-parametersbyexploringlargepar