DIY Deep Learning for Vision- a Hands-On Tutorial with Caffe.pdf

《DIY Deep Learning for Vision- a Hands-On Tutorial with Caffe.pdf》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、DIYDeepLearningforVision:aHands-OnTutorialwithCaffecaffe.berkeleyvision.orggithub.com/BVLC/caffeEvanShelhamer,JeffDonahue,YangqingJia,RossGirshickLookforfurtherdetailsintheoutlinenotesWhyDeepLearning?TheUnreasonableEffectivenessofDeepFeaturesClassesseparateinthedeeprepresentationsandtransfer

2、tomanytasks.[DeCAF][Zeiler-Fergus]WhyDeepLearning?TheUnreasonableEffectivenessofDeepFeaturesMaximalactivationsofpoolunits[R-CNN]5convDeConvvisualizationRichvisualstructureoffeaturesdeepinhierarchy.5[Zeiler-Fergus]WhatisDeepLearning?CompositionalModelsLearnedEnd-to-EndWhatisDeepLearning?Vasts

3、paceofmodels!Caffemodelsareloss-driven:-supervised-unsupervisedslidecreditMarc’aurelioRanzato,CVPR‘14tutorial.ConvolutionalNeuralNets(CNNs):1989LeNet:alayeredmodelcomposedofconvolutionandsubsamplingoperationsfollowedbyaholisticrepresentationandultimatelyaclassifierforhandwrittendigits.[LeNet

4、]ConvolutionalNets:2012AlexNet:alayeredmodelcomposedofconvolution,+datasubsampling,andfurtheroperationsfollowedbyaholistic+gpurepresentationandall-in-allalandmarkclassifieron+non-saturatingnonlinearityILSVRC12.[AlexNet]+regularizationConvolutionalNets:2012AlexNet:alayeredmodelcomposedofconvo

5、lution,pooling,andfurtheroperationsfollowedbyaholisticrepresentationandall-in-allalandmarkclassifieronILSVRC12.[AlexNet]Thefully-connected“FULL”layersarelinearclassifiers/matrixmultiplications.ReLUarerectified-linearnon-linearitiesontheoutputoflayers.parametersFLOPsfigurecreditY.LeCunandM.A.

6、Ranzato,ICML‘13tutorialConvolutionalNets:2014ILSVRC14Winners:~6.6%Top-5error-GoogLeNet:compositionofmulti-scaledimension-+depthreducedmodules(pictured)+data-VGG:16layersof3x3convolutioninterleavedwith+dimensionalityreductionmaxpooling+3fully-connectedlayersLearningaboutDeepLearningRefertothe

7、TutorialonDeepLearningforVisionfromCVPR‘14.●FundamentalsonsupervisedandunsuperviseddeeplearningbyMarc’AurelioRanzato●Listofreferencestoexplore●AdvancedideasandcurrentresearchdirectionsPairswellwiththistutorial!Frameworks●Torch7○NYU○scientificcomput