深學(xué)習(xí)綜述討論簡介deepLearning課件

OutlineConceptionofdeeplearningDevelopmenthistoryDeeplearningframeworksDeepneuralnetworkarchitecturesConvolutionalneuralnetworks

IntroductionNetworkstructureTrainingtricksApplicationinAestheticImageEvaluationIdea

OutlineConceptionofdeeplear1DeepLearning(Hinton,2006)Deeplearningisabranchofmachinelearningbasedonasetofalgorithmsthatattempttomodelhighlevelabstractionsindata.Theadvantageofdeeplearningistoextractingfeaturesautomatically

insteadofextractingfeaturesmanually.ComputervisionSpeechrecognitionNaturallanguageprocessingDeepLearning(Hinton,2006)Deep2DevelopmentHistory194319401950196019701980199020002010MPmodel1958Single-layerPerceptron1969XORproblem1986BPalgorithm1989CNN-LeNet19951997SVMLSTMGradientdisappearanceproblem19912006DBNReLU201120122015DropoutAlexNetBNFasterR-CNNResidualNetGeoffreyHintonW.PittsRosenblattMarvinMinskyYannLeCunHintonHintonHintonLeCunBengioDevelopmentHistory194319403DeepLearningFrameworksDeepLearningFrameworks4DeepneuralnetworkarchitecturesDeepBeliefNetworks(DBN)RecurrentNeuralNetworks(RNN)GenerativeAdversarialNetworks(GANs)ConvolutionalNeuralNetworks(CNN)LongShort-TermMemory(LSTM)Deepneuralnetworkarchitectu5DBN(DeepBeliefNetwork,2006)Hiddenunitsandvisibleunits

Eachunitisbinary(0or1).

Everyvisibleunitconnectstoallthehiddenunits.

Everyhiddenunitconnectstoallthevisibleunits.

Therearenoconnectionsbetweenv-vandh-h.HintonGE.Deepbeliefnetworks[J].Scholarpedia,2009,4(6):5947.Fig1.RBM(restrictedBoltzmannmachine)structure.Fig2.DBN(deepbeliefnetwork)structure.Idea?ComposedofmultiplelayersofRBM.Howtowetraintheseadditionallayers?

UnsupervisedgreedyapproachDBN(DeepBeliefNetwork,2006)H6RNN(RecurrentNeuralNetwork,2013)What?RNNaimstoprocessthesequencedata.RNNwillrememberthepreviousinformationandapplyittothecalculationofthecurrentoutput.Thatis,thenodesofthehiddenlayerareconnected,andtheinputofthehiddenlayerincludesnotonlytheoutputoftheinputlayerbutalsotheoutputofthehiddenlayer.MarhonSA,CameronCJF,KremerSC.RecurrentNeuralNetworks[M]//HandbookonNeuralInformationProcessing.SpringerBerlinHeidelberg,2013:29-65.Applications?MachineTranslationGeneratingImageDescriptionsSpeechRecognitionHowtotrain?

BPTT(Backpropagationthroughtime)RNN(RecurrentNeuralNetwork,27Testingstage:Wiley-IEEEPress,2009.OverFeat:IntegratedRecognition,LocalizationandDetectionusingConvolutionalNetworks[J].Classify:TrainingalinearSVMclassifierforeachclass.LuX,LinZ,JinH,etal.DropoutLayer),ConvolutionlayerShortcutconnectionslayers_['conv2d1'])ArchitectureofMSDLM:SqueezeNet:AlexNet-levelaccuracywith50xfewerparametersand<0.layershavealargereceptivefieldCanceledthefullyconnnectedlayerextracttheartificialfeatures),wecandirectlyinputtheoriginalimage.arXivpreprintarXiv:1502.GANsInspiredbyzero-sumGameinGameTheory,whichconsistsofapairofnetworks-ageneratornetworkandadiscriminatornetwork.cm=confusion_matrix(y_test,preds)CNNStructureEvolutionResidualNetX_train,y_train=data[0]GANs(GenerativeAdversarialNetworks,2014)GANsInspiredbyzero-sumGameinGameTheory,whichconsistsofapairofnetworks-ageneratornetworkandadiscriminatornetwork.Thegeneratornetworkgeneratesasamplefromtherandomvector,thediscriminatornetworkdiscriminateswhetheragivensampleisnaturalorcounterfeit.Bothnetworkstraintogethertoimprovetheirperformanceuntiltheyreachapointwherecounterfeitandrealsamplescannotbedistinguished.GoodfellowI,Pouget-AbadieJ,MirzaM,etal.Generativeadversarialnets[C]//Advancesinneuralinformationprocessingsystems.2014:2672-2680.Applacations:ImageeditingImagetoimagetranslationGeneratetextGenerateimagesbasedontextCombinedwithreinforcementlearningAndmore…Testingstage:GANs(Generative8LongShort-TermMemory(LSTM,1997)LongShort-TermMemory(LSTM,199NeuralNetworksNeuronNeuralnetworkNeuralNetworksNeuronNeuralne10ConvolutionalNeuralNetworks(CNN)Convolutionneuralnetworkisakindoffeedforwardneuralnetwork,whichhasthecharacteristicsofsimplestructure,lesstrainingparametersandstrongadaptability.CNN

avoids

thecomplexpre-processingofimage(etc.extracttheartificialfeatures),wecandirectlyinput

theoriginalimage.

Basiccomponents:ConvolutionLayers,PoolingLayers,FullyconnectedLayersConvolutionalNeuralNetworks(11ConvolutionlayerTheconvolutionkerneltranslates

ona2-dimensionalplane,andeachelementoftheconvolutionkernelismultiplied

bytheelementatthecorrespondingpositionoftheconvolutionimageandthensumalltheproduct.Bymovingtheconvolutionkernel,wehaveanewimage,whichconsistsofthesumoftheproductoftheconvolutionkernelateachposition.localreceptivefieldweightsharingReduced

thenumberofparametersConvolutionlayerTheconvoluti12PoolinglayerPoolinglayeraimstocompresstheinputfeaturemap,whichcanreducethenumberofparameters

intrainingprocessandthedegreeof

over-fitting

ofthemodel.Max-pooling:Selectingthemaximumvalueinthepoolingwindow.Mean-pooling:Calculatingtheaverageofallvaluesinthepoolingwindow.PoolinglayerPoolinglayeraim13FullyconnectedlayerandSoftmaxlayerEachnodeofthefullyconnectedlayerisconnectedtoallthenodesofthelastlayer,whichisusedtocombinethefeaturesextractedfromthefrontlayers.Fig1.Fullyconnectedlayer.Fig2.CompleteCNNstructure.Fig3.Softmaxlayer.FullyconnectedlayerandSoft14TrainingandTestingForwardpropagation-Takingasample(X,Yp)fromthesamplesetandputtheXintothenetwork;-CalculatingthecorrespondingactualoutputOp.Backpropagation-CalculatingthedifferencebetweentheactualoutputOpandthecorrespondingidealoutputYp;-Adjustingtheweightmatrixbyminimizingtheerror.Trainingstage:Testingstage:Puttingdifferentimagesandlabelsintothetrainedconvolutionneuralnetworkandcomparingtheoutputandtheactualvalueofthesample.Beforethetrainingstage,weshouldusesomedifferentsmallrandomnumberstoinitializeweights.TrainingandTestingForwardpr15CNNStructureEvolutionHintonBPNeocognitionLeCunLeNetAlexNetHistoricalbreakthroughReLUDropoutGPU+BigDataVGG16VGG19MSRA-NetDeepernetworkNINGoogLeNetInceptionV3InceptionV4R-CNNSPP-NetFastR-CNNFasterR-CNNInceptionV2(BN)FCNFCN+CRFSTNetCNN+RNN/LSTMResNetEnhancedthefunctionalityoftheconvolutionmoduleClassificationtaskDetectiontaskAdd

newfunctionalunitintegration19801998198920142015ImageNetILSVRC(ImageNetLargeScaleVisualRecognitionChallenge)20132014201520152014,2015201520122015BN(BatchNormalization)RPNCNNStructureEvolutionHinton16LeNet(LeCun,1998)LeNet

isaconvolutionalneuralnetworkdesignedbyYannLeCunforhandwrittennumeralrecognitionin1998.Itisoneofthemostrepresentativeexperimentalsystemsinearlyconvolutionalneuralnetworks.LeNetincludestheconvolutionlayer,poolinglayer

andfull-connectedlayer,whicharethebasiccomponentsofmodernCNNnetwork.LeNetisconsideredtobethebeginningoftheCNN.networkstructure:3convolutionlayers+2poolinglayers+1fullyconnectedlayer+1outputlayerHaykinS,KoskoB.GradientBasedLearningAppliedtoDocumentRecognition[D].Wiley-IEEEPress,2009.LeNet(LeCun,1998)LeNetisaco17AlexNet(Alex,2012)Networkstructure:5convolutionlayers+3fullyconnectedlayersThenonlinearactivationfunction:ReLU(Rectifiedlinearunit)Methodstopreventoverfitting:Dropout,DataAugmentationBigDataTraining:ImageNet--imagedatabaseofmillionordersofmagnitudeOthers:GPU,LRN(localresponsenormalization)layerKrizhevskyA,SutskeverI,HintonGE.ImageNetclassificationwithdeepconvolutionalneuralnetworks[C]//InternationalConferenceonNeuralInformationProcessingSystems.CurranAssociatesInc.2012:1097-1105.AlexNet(Alex,2012)Networkstru18X_train,y_train=data[0]filename=Thearchitectureofthemulti-scenedeeplearningmodel(MSDLM).Max-pooling:2*2pixelwindow,withstride2Why3*3filters?X_train,y_train,X_val,y_val,X_test,y_test=load_dataset()SemanticSegmentationRNNwillrememberthepreviousinformationandapplyittothecalculationofthecurrentoutput.FasterR-CNN(2015)ImageNetclassificationwithdeepconvolutionalneuralnetworks[C]//InternationalConferenceonNeuralInformationProcessingSystems.[7]R,DonahueJ,DarrellT,etal.2012:1097-1105.InceptionV2(2015)IntroductionarXivpreprintarXiv:1502.Theconvolutionallayerparametersaredenotedas“conv<receptivefieldsize>-<numberofchannels>”LongShort-TermMemory(LSTM)ReLU(RectifiedLinearUnit)segmentedimagesIEEETransactionsonMultimedia,2015,17(11):2021-2034.4Mlpconvlayers+GlobalaveragepoolinglayerOverfeat(2013)SermanetP,EigenD,ZhangX,etal.OverFeat:IntegratedRecognition,LocalizationandDetectionusingConvolutionalNetworks[J].EprintArxiv,2013.X_train,y_train=data[0]Over19VGG-Net(OxfordUniversity,2014)input:afixed-size224*224RGBimagefilters:averysmallreceptivefield--3*3,withstride1Max-pooling:2*2pixelwindow,withstride2Fig1.ArchitectureofVGG16Table1:ConvNetconfigurations(shownincolumns).Theconvolutionallayerparametersaredenotedas“conv<receptivefieldsize>-<numberofchannels>”

SimonyanK,ZissermanA.VeryDeepConvolutionalNetworksforLarge-ScaleImageRecognition[J].ComputerScience,2014.Why3*3filters?Stackedconv.layershavealargereceptivefieldMorenon-linearityLessparameterstolearnVGG-Net(OxfordUniversity,201420Network-in-Network(NIN,ShuichengYan,2013)Networkstructure:4Mlpconvlayers+GlobalaveragepoolinglayerFig1.linearconvolution

MLPconvolutionFig2.fullyconnectedlayer

globalaveragepoolinglayerMinLinetal,NetworkinNetwork,Arxiv2013.Fig3.NINstructureLinearcombinationofmultiplefeaturemaps.Informationintegrationofcross-channel.ReducedtheparametersReducedthenetworkAvoidedover-fittingNetwork-in-Network(NIN,Shuich21GoogLeNet(InceptionV1,2014)Fig1.Inceptionmodule,na?veversionProposedinceptionarchitectureandoptimizeditCanceled

thefullyconnnectedlayerUsedauxiliaryclassifierstoacceleratenetworkconvergenceSzegedyC,LiuW,JiaY,etal.Goingdeeperwithconvolutions[C]//ProceedingsoftheIEEEConferenceonComputerVisionandPatternRecognition.2015:1-9.Fig2.InceptionmodulewithdimensionreductionsFig3.GoogLeNetnetwork(22layers)GoogLeNet(InceptionV1,2014)Fi22InceptionV2(2015)IoffeS,SzegedyC.Batchnormalization:Acceleratingdeepnetworktrainingbyreducinginternalcovariateshift[J].arXivpreprintarXiv:1502.03167,2015.InceptionV2(2015)IoffeS,Sze23InceptionV3(2015)SzegedyC,VanhouckeV,IoffeS,etal.Rethinkingtheinceptionarchitectureforcomputervision[C]//ProceedingsoftheIEEEConferenceonComputerVisionandPatternRecognition.2016:2818-2826.InceptionV3(2015)SzegedyC,V24ResNet(KaiwenHe,2015)Asimpleandcleanframeworkoftraining“very”deepnetworks.State-of-the-artperformanceforImageclassificationObjectdetectionSemanticSegmentationandmoreHeK,ZhangX,RenS,etal.DeepResidualLearningforImageRecognition[J].2015:770-778.Fig1.ShortcutconnectionsFig2.ResNetstructure(152layers)ResNet(KaiwenHe,2015)Asimpl25FractalNetFractalNet26InceptionV4(2015)SzegedyC,IoffeS,VanhouckeV,etal.Inception-v4,inception-resnetandtheimpactofresidualconnectionsonlearning[J].arXivpreprintarXiv:1602.07261,2016.InceptionV4(2015)SzegedyC,I27Inception-ResNetHeK,ZhangX,RenS,etal.DeepResidualLearningforImageRecognition[J].2015:770-778.Inception-ResNetHeK,ZhangX,28Canceledthefullyconnnectedlayer('maxpool1',layers.RecurrentNeuralNetworks(RNN)X_val,y_val=data[1]19401950196019701980199020002010breakthroughTheadvantageofdeeplearningistoextractingfeaturesautomaticallyinsteadofextractingfeaturesmanually.RegionProposalNetwork(RPN).SpringerInternationalPublishing,2015:524-535.ClassificationtaskSqueezeNet:AlexNet-levelaccuracywith50xfewerparametersand<0.Advantages:plot_conv_weights(net1.CNNavoidsthecomplexpre-processingofimage(etc.CNNStructureEvolutionBPTT(Backpropagationthroughtime)[5]SimonyanK,ZissermanA.Avoidedover-fittingX_train,y_train,X_val,y_val,X_test,y_test=load_dataset()extracttheartificialfeatures),wecandirectlyinputtheoriginalimage.Inceptionmodule,na?veversion-Takingasample(X,Yp)fromthesamplesetandputtheXintothenetwork;RNN(RecurrentNeuralNetwork,2013)MarvinMinskyImagetoimagetranslation2015:1440-1448.Conv2DLayer),DeepLearningFrameworksRenS,HeK,GirshickR,etal.RegionProposalNetwork(RPN).y_train=y_train.X_test=X_test.RNNwillrememberthepreviousinformationandapplyittothecalculationofthecurrentoutput.4Mlpconvlayers+GlobalaveragepoolinglayerOverFeat:IntegratedRecognition,LocalizationandDetectionusingConvolutionalNetworks[J].DeepneuralnetworkarchitecturesDongZ,ShenX,LiH,etal.AllparametersinDCNNarejointlytrained.DeeplearningframeworksOutputafixedlengthfeaturevectorwithinputsofarbitrarysizes.SqueezeNet:AlexNet-levelaccuracywith50xfewerparametersand<0.Informationintegrationofcross-channel.ComparisonCanceledthefullyconnnected29SqueezeNet

SqueezeNet:AlexNet-levelaccuracywith50xfewerparametersand<0.5MBmodelsizeSqueezeNet

SqueezeNet:AlexNet30XceptionXception31R-CNN(2014)Regionproposals:SelectiveSearch

Resizetheregionproposal:Warpallregionproposalstotherequiredsize(227*227,

AlexNetInput)

ComputeCNNfeature:Extracta4096-dimensionalfeaturevectorfromeachregionproposalusingAlexNet.

Classify:TrainingalinearSVMclassifierforeachclass.[1]UijlingsJRR,SandeKEAVD,GeversT,etal.SelectiveSearchforObjectRecognition[J].InternationalJournalofComputerVision,2013,104(2):154-171.[2]GirshickR,DonahueJ,DarrellT,etal.RichFeatureHierarchiesforAccurateObjectDetectionandSemanticSegmentation[J].2014:580-587.R-CNN:Regionproposals+CNNR-CNN(2014)Regionproposals:32SPP-Net(Spatialpyramidpoolingnetwork,2015)HeK,ZhangX,RenS,etal.SpatialPyramidPoolinginDeepConvolutionalNetworksforVisualRecognition[J].IEEETransactionsonPatternAnalysis&MachineIntelligence,2015,37(9):1904-1916.Fig2.Anetworkstructurewithaspatialpyramidpoolinglayer.Fig1.Top:AconventionalCNN.Bottom:Spatialpyramidpoolingnetworkstructure.Advantages:Getthefeaturemapoftheentireimagetosavemuchtime.Outputafixedlengthfeaturevectorwithinputsofarbitrarysizes.Extractthefeatureofdifferentscale,andcanexpressmorespatialinformation.TheSPP-Netmethodcomputesaconvolutionalfeaturemapfortheentireinputimageandthenclassifieseachobjectproposalusingafeaturevectorextractedfromthesharedfeaturemap.SPP-Net(Spatialpyramidpoolin33FastR-CNN(2015)AFastR-CNNnetworktakesanentireimageandasetofobjectproposalsasinput.Thenetworkprocessestheentireimagewithseveralconvolutional(conv)andmaxpoolinglayerstoproduceaconvfeaturemap.Foreachobjectproposal,aregionofinterest(RoI)poolinglayerextractsafixed-lengthfeaturevectorfromthefeaturemap.Eachfeaturevectorisfedintoasequenceoffullyconnectedlayersthatfinallybranchintotwosiblingoutputlayers.

GirshickR.Fastr-cnn[C]//ProceedingsoftheIEEEInternationalConferenceonComputerVision.2015:1440-1448.FastR-CNN(2015)AFastR-CNNn34FasterR-CNN(2015)FasterR-CNN=RPN+FastR-CNN

ARegionProposalNetwork(RPN)takesanimage(ofanysize)asinputandoutputsasetofrectangularobjectproposals,eachwithanobjectnessscore.

RenS,HeK,GirshickR,etal.Fasterr-cnn:Towardsreal-timeobjectdetectionwithregionproposalnetworks[C]//Advancesinneuralinformationprocessingsystems.2015:91-99.Figure1.FasterR-CNNisasingle,unifiednetworkforobjectdetection.Figure2.RegionProposalNetwork(RPN).FasterR-CNN(2015)FasterR-CNN35TrainingtricksDataAugmentationDropoutReLUBatchNormalizationTrainingtricksDataAugmentati36DataAugmentation-rotation-flip-zoom-shift-scale-contrast-noisedisturbance-color-...DataAugmentation-rotation37Dropout(2012)Dropoutconsistsofsettingtozerotheoutputofeachhiddenneuronwithprobabilityp.Theneuronswhichare“droppedout”inthiswaydonotcontributetotheforwardbackpropagationanddonotparticipateinbackpropagation.Dropout(2012)Dropoutconsists38ReLU(RectifiedLinearUnit)

advantagesrectifiedSimplifiedcalculationAvoidedgradientdisappearedReLU(RectifiedLinearUnit)

ad39BatchNormalization(2015)Intheinputofeachlayerofthenetwork,insertanormalizedlayer.Foralayerwithd-dimensionalinputx=(x(1)...x(d)),wewillnormalizeeachdimension：IoffeS,SzegedyC.Batchnormalization:Acceleratingdeepnetworktrainingbyreducinginternalcovariateshift[J].arXivpreprintarXiv:1502.03167,2015.Internal

Covariate

Shift

BatchNormalization(2015)Inth40ApplicationinAestheticImageEvaluationDongZ,ShenX,LiH,etal.PhotoQualityAssessmentwithDCNNthatUnderstandsImageWell[M]//MultiMediaModeling.SpringerInternationalPublishing,2015:524-535.LuX,LinZ,JinH,etal.Ratingimageaestheticsusingdeeplearning[J].IEEETransactionsonMultimedia,2015,17(11):2021-2034.WangW,ZhaoM,WangL,etal.Amulti-scenedeeplearningmodelforimageaestheticevaluation[J].SignalProcessingImageCommunication,2016,47:511-518.ApplicationinAestheticImage41PhotoQualityAssessmentwithDCNNthatUnderstandsImageWellDCNN_Aesthtrainedwellnetworkatwo-classSVMclassifierDCNN_Aesth_SPoriginalimagessegmentedimagesspatialpyramidImageNetCUHKAVADongZ,ShenX,LiH,etal.PhotoQualityAssessmentwithDCNNthatUnderstandsImageWell[M]//MultiMediaModeling.SpringerInternationalPublishing,2015:524-535.PhotoQualityAssessmentwith42RatingimageaestheticsusingdeeplearningSupportheterogeneousinputs,i.e.,globaland

localviews.AllparametersinDCNNarejointlytrained.Fig1.GlobalviewsandlocalviewsofanimageFig3.DCNNarchitectureFig2.SCNNarchitecture

SCNNDCNN

Enablesthenetworktojudgeimageaestheticswhilesimultaneouslyconsideringboththeglobalandlocalviewsofanimage.LuX,LinZ,JinH,etal.Ratingimageaestheticsusingdeeplearning[J].IEEETransactionsonMultimedia,2015,17(11):2021-2034.Ratingimageaestheticsusing43Generativeadversarialnets[C]//Advancesinneuralinformationprocessingsystems.SermanetP,EigenD,ZhangX,etal.DeepLearningFrameworksMarhonSA,CameronCJF,KremerSC.withgzip.Enhancedthefunctionalityoftheconvolutionmodule5MBmodelsizeextracttheartificialfeatures),wecandirectlyinputtheoriginalimage.VeryDeepConvolutionalNetworksforLarge-ScaleImageRecognition[J].[6]SzegedyC,LiuW,JiaY,etal.RNN(RecurrentNeuralNetwork,2013)HaykinS,KoskoB.VeryDeepConvolutionalNetworksforLarge-ScaleImageRecognition[J].[2]GoodfellowI,Pouget-AbadieJ,MirzaM,etal.BN(BatchNormalization)InceptionmodulewithdimensionreductionsPhotoQualityAssessmentwithDCNNthatUnderstandsImageWell[M]//MultiMediaModeling.MarvinMinskyprint("DownloadingMNISTdataset.Dropoutconsistsofsettingtozerotheoutputofeachhiddenneuronwithprobabilityp.Thenonlinearactivationfunction:ReLU(Rectifiedlinearunit)RecurrentNeuralNetworks[M]//HandbookonNeuralInformationProcessing.Amulti-scenedeeplearningmodelforimageaestheticevaluationDesignasceneconvolutionallayerconsistofmulti-groupdescriptorsinthenetwork.Designapre-trainingproceduretoinitializeourmodel.Fig1.Thearchitectureofthemulti-scenedeeplearningmodel(MSDLM).Fig2.TheoverviewofproposedMSDLM.ArchitectureofMSDLM:4

convolutionallayers+1sceneconvolutionallayer+3fullyconnectedlayersWangW,ZhaoM,WangL,etal.Amulti-scenedeeplearningmodelforimageaestheticevaluation[J].SignalProcessingImageCommunication,2016,47:511-518.Generativeadversarialnets[C]44Example-Loadthedatasetdefload_dataset():url=filename=

if

print("DownloadingMNISTdataset...")

urlretrieve(url,filename)

withgzip.open(filename,'rb')asf:data=pickle.load(f)X_train,y_train=data[0]X_val,y_val=data[1]X_test,y_test=data[2]X_train=X_train.reshape((-1,1,28,28))X_val=X_val.reshape((-1,1,28,28))X_test=X_test.reshape((-1,1,28,28))y_train=y_train.astype(np.uint8)y_val=y_val.astype(np.uint8)y_test=y_test.astype(np.uint8)

returnX_train,y_train,X_val,y_val,X_test,y_test

X_train,y_train,X_val,y_val,X_test,y_test=load_dataset()plt.imshow(X_train[0][0],cmap=cm.binary)Example-Loadthedatasetdefl45Example–Modelnet1=NeuralNet(layers=[('input',layers.InputLayer),

('conv2d1',

layers.Conv2DLayer),

('maxpool1',

layers.MaxPool2DLayer),

('conv2d2',layers.Conv2DLayer),

('maxpool2',layers.MaxPool2DLayer),

('dropout1',layers.DropoutLayer),

('dense',layers.DenseLayer),

('dropout2',layers.DropoutLayer),

('output',layers.DenseLayer),

],

#inputlayerinput_shape=(None,1,28,28),#layerconv2d1conv2d1_num_filters=32,conv2d1_filter_size=(5,5),,

#layermaxpool1maxpool1_pool_size=(2,2),#layerconv2d2conv2d2_num_filters=32,conv2d2_filter_size=(5,5),,

#layermaxpool2maxpool2_pool_size=(2,2),

#dropout1dropout1_p=0.5,

#densei.e.full-connectedlayerdense_num_units=256,

#dropout2dropout2_p=0.5,

#outputoutput_num_units=10,

#optimizationmethodparamsupdate=nesterov_momentum,update_learning_rate=0.01,update_momentum=0.9,max_epochs=10,verbose=1,)Example–Modelnet1=NeuralNet46Example–TrainandTest#Trainthenetworknn=net1.fit(X_train,y_train)#Usingtheabovetrainingmodeltopredictthetestsetpreds=net1.predict(X_test)cm=confusion_matrix(y_test,preds)plt.matshow(cm)plt.title('Confusionmatrix')plt.colorbar()plt.ylabel('Truelabel')plt.xlabel('Predictedlabel')plt.show()#visualizethefeaturemapofconv2d1visualize.plot_conv_weights(net1.layers_['conv2d1'])Example–TrainandTest#Train47Example–ResultExample–Result48References[1]MarhonSA,CameronCJF,KremerSC.RecurrentNeuralNetworks[M]//HandbookonNeuralInformationProcessing.SpringerBerlinHeidelberg,2013:29-65.[2]GoodfellowI,Pouget-AbadieJ,MirzaM,etal.Generativeadversarialnets[C]//Advancesinneuralinformationprocessingsystems.2014:2672-2680.[3]HaykinS,KoskoB.GradientBasedLearningAppliedtoDocumentRecognition[D].Wiley-IEEEPress,2009.[4]KrizhevskyA,SutskeverI,HintonGE.ImageNetclassificationwithdeepconvolutionalneuralnetworks[C]//InternationalConferenceonNeuralInformationProcessingSystems.CurranAssociatesInc.2012:1097-1105.[5]SimonyanK,ZissermanA.VeryDeepConvolutionalNetworksforLarge-ScaleImageRecognition[J].ComputerScience,2014.[6]SzegedyC,LiuW,JiaY,etal.Goingdeeperwithconvolutions[C]//ProceedingsoftheIEEEConferenceonComputerVisionandPatternRecognition.2015:1-9.[7]R,DonahueJ,DarrellT,etal.RichFeatureHierarchiesforAccurateObjectDetectionandSemanticSegmentation[J].2014:580-587.[8]DongZ,ShenX,LiH,etal.PhotoQualityAssessmentwithDCNNthatUnderstandsImageWell[M]//MultiMediaModeling.SpringerInternationalPublishing,2015:524-535.[9]LuX,LinZ,JinH,etal.Ratingimageaestheticsusingdeeplearning[J].IEEETransactionsonMultimedia,2015,17(11):2021-2034.[10]WangW,ZhaoM,WangL,etal.Amulti-scenedeeplearningmodelforimageaestheticevaluation[J].SignalProcessingImageCommunication,2016,47:511-518.References[1]MarhonSA,Camer49Thanks!深學(xué)習(xí)綜述討論簡介deepLearning課件50DeepLearningFrameworksDeepLearningFrameworks51DeepneuralnetworkarchitecturesDeepBeliefNetworks(DBN)RecurrentNeuralNetworks(RNN)GenerativeAdversarialNetworks(GANs)ConvolutionalNeuralNetworks(CNN)LongShort-TermMemory(LSTM)Deepneuralnetworkarchitectu52ConvolutionalNeuralNetworks(CNN)Convolutionneuralnetworkisakindoffeedforwardneuralnetwork,whichhasthecharacteristicsofsimplestructure,lesstraini