Recent Advances in Convolutional Neural Networks

《Recent Advances in Convolutional Neural Networks》由會員上傳分享，免費在線閱讀，更多相關內容在學術論文-天天文庫。

1、1RecentAdvancesinConvolutionalNeuralNetworksJiuxiangGu?,ZhenhuaWang?,JasonKuen,LianyangMa,AmirShahroudy,BingShuai,TingLiu,XingxingWang,andGangWang,Member,IEEEAbstract—Inthelastfewyears,deeplearninghasledtoverythem,fourrepresentativeworksareZFNet[7],VGGNet[8],goodperformanceonavarietyofproble

2、ms,suchasvisualGoogleNet[9]andResNet[10].Fromtheevolutionoftherecognition,speechrecognitionandnaturallanguageprocessing.architectures,atypicaltrendisthatthenetworksaregettingAmongdifferenttypesofdeepneuralnetworks,convolutionaldeeper,e.g.,ResNet,whichwonthechampionofILSVRCneuralnetworkshaveb

3、eenmostextensivelystudied.Duetothelackoftrainingdataandcomputingpowerinearlydays,itis2015,isabout20timesdeeperthanAlexNetand8timeshardtotrainalargehigh-capacityconvolutionalneuralnetworkdeeperthanVGGNet.Byincreasingdepth,thenetworkcanwithoutover?tting.Aftertherapidgrowthintheamountofthebette

4、rapproximatethetargetfunctionwithincreasednon-annotateddataandtherecentimprovementsinthestrengthsoflinearityandgetbetterfeaturerepresentations.However,itgraphicsprocessorunits(GPUs),theresearchonconvolutionalalsoincreasesthecomplexityofthenetwork,whichmakesneuralnetworkshasbeenemergedswiftly

5、andachievedstate-of-the-artresultsonvarioustasks.Inthispaper,weprovideabroadthenetworkbemoredif?culttooptimizeandeasiertogetsurveyoftherecentadvancesinconvolutionalneuralnetworks.over?tting.Alongthisway,variousmethodsareproposedtoBesides,wealsointroducesomeapplicationsofconvolutionaldealwith

6、theseproblemsinvariousaspects.Inthispaper,weneuralnetworksincomputervision.trytogiveacomprehensivereviewofrecentadvancesandIndexTerms—ConvolutionalNeuralNetwork,Deeplearning.givesomethoroughdiscussions.Inthefollowingsections,weidentifybroadcategoriesofworksrelatedtoCNN.We?rstgiveanoverviewof

7、thebasicI.INTRODUCTIONcomponentsofCNNinSectionII.Then,weintroducesomeONVOLUTIONALNeuralNetwork(CNN)isawell-recentimprovementsondifferentaspectsofCNNincludingCknowndeeplearningarchitectureinspiredbythenaturalconvolutionallayer,poolinglayer,activatio