資源描述：
《2.2 Understanding and Visualizing Convolutional Neural Networks》由會(huì)員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫(kù)。
1、CS231nConvolutionalNeuralNetworksforVisualRecognition(thispageiscurrentlyindraftform)VisualizingwhatConvNetslearnSeveralapproachesforunderstandingandvisualizingConvolutionalNetworkshavebeendevelopedintheliterature,partlyasaresponsethecommoncriticismthatthelearn
2、edfeaturesinaNeuralNetworkarenotinterpretable.Inthissectionwebrieflysurveysomeoftheseapproachesandrelatedwork.Visualizingtheactivationsandfirst-layerweightsLayerActivations.Themoststraight-forwardvisualizationtechniqueistoshowtheactivationsofthenetworkduringthe
3、forwardpass.ForReLUnetworks,theactivationsusuallystartoutlookingrelativelyblobbyanddense,butasthetrainingprogressestheactivationsusuallybecomemoresparseandlocalized.Onedangerouspitfallthatcanbeeasilynoticedwiththisvisualizationisthatsomeactivationmapsmaybeallze
4、roformanydifferentinputs,whichcanindicatedeadfilters,andcanbeasymptomofhighlearningrates.Typical-lookingactivationsonthefirstCONVlayer(left),andthe5thCONVlayer(right)ofatrainedAlexNetlookingatapictureofacat.Everyboxshowsanactivationmapcorrespondingtosomefilter.
5、Noticethattheactivationsaresparse(mostvaluesarezero,inthisvisualizationshowninblack)andmostlylocal.Conv/FCFilters.Thesecondcommonstrategyistovisualizetheweights.TheseareusuallymostinterpretableonthefirstCONVlayerwhichislookingdirectlyattherawpixeldata,butitispo
6、ssibletoalsoshowthefilterweightsdeeperinthenetwork.Theweightsareusefultovisualizebecausewell-trainednetworksusuallydisplayniceandsmoothfilterswithoutanynoisypatterns.Noisypatternscanbeanindicatorofanetworkthathasn'tbeentrainedforlongenough,orpossiblyaverylowreg
7、ularizationstrengththatmayhaveledtooverfitting.Typical-lookingfiltersonthefirstCONVlayer(left),andthe2ndCONVlayer(right)ofatrainedAlexNet.Noticethatthefirst-layerweightsareveryniceandsmooth,indicatingnicelyconvergednetwork.Thecolor/grayscalefeaturesareclustered
8、becausetheAlexNetcontainstwoseparatestreamsofprocessing,andanapparentconsequenceofthisarchitectureisthatonestreamdevelopshigh-frequencygrayscalefeaturesandtheotherlow-freque
