15）20161028 Convolutional Neural Network

《15）20161028 Convolutional Neural Network》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、ConvolutionalNeuralNetworkWhyCNNforImage?[Zeiler,M.D.,ECCV2014]x……1x2…………………………Representedaspixelsx……NThemostbasicUse1stlayerasmoduleUse2ndlayerasclassifierstobuildclassifiersmodule……Canthenetworkbesimplifiedbyconsideringthepropertiesofimages?WhyCNNforIma

2、ge•SomepatternsaremuchsmallerthanthewholeimageAneurondoesnothavetoseethewholeimagetodiscoverthepattern.Connectingtosmallregionwithlessparameters“beak”detectorWhyCNNforImage•Thesamepatternsappearindifferentregions.“upper-leftbeak”detectorDoalmostthesamethi

3、ngTheycanusethesamesetofparameters.“middlebeak”detectorWhyCNNforImage•SubsamplingthepixelswillnotchangetheobjectbirdbirdsubsamplingWecansubsamplethepixelstomakeimagesmallerLessparametersforthenetworktoprocesstheimageThewholeCNNcatdog……ConvolutionMaxPoolin

4、gCanrepeatFullyConnectedmanytimesFeedforwardnetworkConvolutionMaxPoolingFlattenThewholeCNNProperty1SomepatternsaremuchConvolutionsmallerthanthewholeimageProperty2MaxPoolingThesamepatternsappearinCanrepeatdifferentregions.manytimesProperty3ConvolutionSu

5、bsamplingthepixelswillnotchangetheobjectMaxPoolingFlattenThewholeCNNcatdog……ConvolutionMaxPoolingCanrepeatFullyConnectedmanytimesFeedforwardnetworkConvolutionMaxPoolingFlattenCNN–ConvolutionThosearethenetworkparameterstobelearned.1-1-1100001Filter1-11-101

6、0010Matrix-1-11001100100010-11-1Filter2010010-11-1Matrix001010-11-1……6x6imageEachfilterdetectsasmallProperty1pattern(3x3).1-1-1CNN–Convolution-11-1Filter1-1-11stride=11000010100103-10011001000100100100010106x6image1-1-1CNN–Convolution-11-1Filter1-1-11Ifst

7、ride=21000010100103-3001100100010010010Wesetstride=1below0010106x6image1-1-1CNN–Convolution-11-1Filter1-1-11stride=11000010100103-1-3-1001100100010-310-3010010001010-3-3016x6image3-2-2-1Property2-11-1CNN–Convolution-11-1Filter2-11-1stride=1Dothesameproces

8、sforeveryfilter1000010100103-1-3-1-1-1-1-1001100100010-310-3-1-1-21Feature010010001010-3-1-3-1Map0-2116x6image3-2-2-1-10-434x4imageCNN–Colorfulimage11-1-1-1-1-1-111-1-11-1-1-11-1-1-111-1-1-1-111-1-1-11-1Filter1-11-1