16-卷积神经网络基础上.pptxVIP

DeepLearninganditsApplicationConvolutionalneuralnetworks

Convolutionalneuralnetworks19SJTUDeepLearningLecture.

Dataaugmentation19SJTUDeepLearningLecture.ImageflippingColortransformationImagecropping

ImageNetcompetition19SJTUDeepLearningLecture.

ConvolutionalNeuralNetworks30721FullyConnectedLayer32x32x3image-stretchto3072x110x3072weightsactivationinput1101number:theresultoftakingadotproductbetweenarowofWandtheinput(a3072-dimensionaldotproduct)19SJTUDeepLearningLecture.

32203ConvolutionLayer32x32x3image-preservespatialstructurewidthheight32depthSJTUDeepLearningLecture.

ConvolutionLayer32x32x3image5x5x3filter32Convolvethefilterwiththeimagei.e.“slideovertheimagespatially,computingdotproducts”32321SJTUDeepLearningLecture.

ConvolutionLayer32x32x3image5x5x3filter32Convolvethefilterwiththeimagei.e.“slideovertheimagespatially,computingdotproducts”323Filtersalwaysextendthefulldepthoftheinputvolume22SJTUDeepLearningLecture.

32332x32x3image5x5x3filter321number:theresultoftakingadotproductbetweenthefilterandasmall5x5x3chunkoftheimage(i.e.5*5*3=75-dimensionaldotproduct+bias)ConvolutionLayer23SJTUDeepLearningLecture.

32332x32x3image5x5x3filter32convolve(slide)overallspatiallocationsactivationmap2412828ConvolutionLayerSJTUDeepLearningLecture.

3232332x32x3image5x5x3filterconvolve(slide)overallspatiallocationsactivationmaps12828considerasecond,greenfilter25ConvolutionLayerSJTUDeepLearningLecture.

323628Forexample,ifwehad65x5filters,we’llget6separateactivationmaps:activationmaps3228ConvolutionLayerWestacktheseuptogeta“newimage”ofsize28x28x6!26ConvolutionLayerSJTUDeepLearningLecture.

ConvolutionLayerPreview:ConvNetisasequenceofConvolutionLayers,interspersedwithactivationfunctions32323CONV,ReLUe.g.65x5x3filters28286CONV,ReLUe.g.105x5x6filtersCONV,