2.2 Understanding and Visualizing Convolutional Neural Networks

2.2 Understanding and Visualizing Convolutional Neural Networks

ID：40703532

大?。?90.14 KB

頁數：9頁

時間：2019-08-06

2.2 Understanding and Visualizing Convolutional Neural Networks_第1頁
2.2 Understanding and Visualizing Convolutional Neural Networks_第2頁
2.2 Understanding and Visualizing Convolutional Neural Networks_第3頁
2.2 Understanding and Visualizing Convolutional Neural Networks_第4頁
2.2 Understanding and Visualizing Convolutional Neural Networks_第5頁
資源描述：

《2.2 Understanding and Visualizing 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、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

當前文檔最多預覽五頁，下載文檔查看全文

此文檔下載收益歸作者所有

當前文檔最多預覽五頁，下載文檔查看全文
溫馨提示：
1. 部分包含數學公式或PPT動畫的文件，查看預覽時可能會顯示錯亂或異常，文件下載后無此問題，請放心下載。
2. 本文檔由用戶上傳，版權歸屬用戶，天天文庫負責整理代發(fā)布。如果您對本文檔版權有爭議請及時聯系客服。
3. 下載前請仔細閱讀文檔內容，確認文檔內容符合您的需求后進行下載，若出現內容與標題不符可向本站投訴處理。
4. 下載文檔時可能由于網絡波動等原因無法下載或下載錯誤，付費完成后未能成功下載的用戶請聯系客服處理。