2.1 Convolutional Neural Networks

《2.1 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、CS231nConvolutionalNeuralNetworksforVisualRecognitionTableofContents:•ArchitectureOverview•ConvNetLayers◦ConvolutionalLayer◦PoolingLayer◦NormalizationLayer◦Fully-ConnectedLayer◦ConvertingFully-ConnectedLayerstoConvolutionalLayers•ConvNetArchitectures◦LayerPatterns◦LayerSizin

2、gPatterns◦CaseStudies(LeNet/AlexNet/ZFNet/GoogLeNet/VGGNet)◦ComputationalConsiderations•AdditionalReferencesConvolutionalNeuralNetworks(CNNs/ConvNets)ConvolutionalNeuralNetworksareverysimilartoordinaryNeuralNetworksfromthepreviouschapter:Theyaremadeupofneuronsthathavelearnab

3、leweightsandbiases.Eachneuronreceivessomeinputs,performsadotproductandoptionallyfollowsitwithanon-linearity.Thewholenetworkstillexpressasingledifferentiablescorefunction:Fromtherawimagepixelsononeendtoclassscoresattheother.Andtheystillhavealossfunction(e.g.SVM/Softmax)onthel

4、ast(fully-connected)layerandallthetips/trickswedevelopedforlearningregularNeuralNetworksstillapply.Sowhatdoeschange?ConvNetarchitecturesmaketheexplicitassumptionthattheinputsareimages,whichallowsustoencodecertainpropertiesintothearchitecture.Thesethenmaketheforwardfunctionmo

5、reefficienttoimplementandvastlyreducestheamountofparametersinthenetwork.ArchitectureOverviewRecall:RegularNeuralNets.Aswesawinthepreviouschapter,NeuralNetworksreceiveaninput(asinglevector),andtransformitthroughaseriesofhiddenlayers.Eachhiddenlayerismadeupofasetofneurons,wher

6、eeachneuronisfullyconnectedtoallneuronsinthepreviouslayer,andwhereneuronsinasinglelayerfunctioncompletelyindependentlyanddonotshareanyconnections.Thelastfully-connectedlayeriscalledthe"outputlayer"andinclassificationsettingsitrepresentstheclassscores.RegularNeuralNetsdon'tsc

7、alewelltofullimages.InCIFAR-10,imagesareonlyofsize32x32x3(32wide,32high,3colorchannels),soasinglefully-connectedneuroninafirsthiddenlayerofaregularNeuralNetworkwouldhave32*32*3=3072weights.Thisamountstillseemsmanageable,butclearlythisfully-connectedstructuredoesnotscaletolar

8、gerimages.Forexample,animageofmorerespectiblesize,e.g.200x200x3,wouldleadto