第1題(Singlechoicequestion)Whatdoesaneuroncompute?()AAneuroncomputesanactivationfunctionfollowedbyalinearfunction(z=Wx+b)BAneuroncomputesalinearfunction(z=Wx+b)followedbyanactivationfunctionCAneuroncomputesafunctiongthatscalestheinputxlinearly(Wx+b)DAneuroncomputesthemeanofallfeaturesbeforeapplyingtheoutputtoanactivationfunction第2題(Singlechoicequestion)WhichoftheseisLogisticloss?()AMSElossB

NCElossC

FocallossDCross-Entropyloss第3題(Singlechoicequestion)WhichoftheseisthebasicunitinNN?()AneuronBlayerC

parameterDModule第4題(Singlechoicequestion)WhichoftheseisNOTactivationfunction?()AReLuBSoftmaxCTanhDSigmoid第5題(Singlechoicequestion)Considerthetwofollowingrandomarrays"a"and"b",

a=np.random.randn(4,3);b=np.random.randn(3,2);c=a*bwhatwillbetheshapeof“c:().A(4,3)B(4,4)C(4,2)D(3,3)第6題(Singlechoicequestion)Whichoftheseistheabbreviationofmulti-layerperceptron?()AMLPBRNNCNLPDBTP第7題(Singlechoicequestion)What’sthedifferencebetweenthesupervisedlearningandtheunsupervisedlearning?()ASupervisedlearninghaslabels,unsupervisedlearningdoesn’thavelabels.BSupervisedlearningneedtobetrained,unsupervisedlearningdoesn’tneedtobetrained.CSupervisedLearning:adesignedoutputisknownandusedtocomputeerrorsignal.UnsupervisedLearning:nosuchoutputisknown.第8題(Singlechoicequestion)Whichoftheseisthebasicsupervisedalgorithmofneuralnetwork?()AforwardcomputingBbackpropagationCgradientcomputing第9題(Singlechoicequestion)WhichoftheseistheformulaofSquareloss?()ABCD第10題(Singlechoicequestion)WhichoftheseistheformulaofAbsolutevalueloss?()ABCD第11題(Singlechoicequestion)WhichofthesearenotthereasonforDeepLearningrecentlytakingoff?()AWehaveaccesstoalotmorecomputationalpower.BNeuralNetworksareabrandnewfield.CWehaveaccesstoalotmoredata.DDeeplearninghasresultedinsignificantimprovementsinimportantapplicationssuchasonlineadvertising,speechrecognition,andimagerecognition.第12題(Singlechoicequestion)Whenanexperienceddeeplearningengineerworksonanewproblem,theycanusuallyuseinsightfrompreviousproblemstotrainagoodmodelonthefirsttry,withoutneedingtoiteratemultipletimesthroughdifferentmodels.()ATrueBFalse第13題(Singlechoicequestion)WhichoftheseistheformulaofCross-Entropyloss?()ABCD第14題(Singlechoicequestion)Youarebuildingabinaryclassifierforrecognizingcucumbers(y=1)vs.watermelons(y=0).Whichoneoftheseactivationfunctionswouldyourecommendusingfortheoutputlayer?()AReLUB

LeakyReLUC

sigmoidDTanh第15題(Singlechoicequestion)WhichoftheseistheformulaofHingeloss?()ABCD第16題(Singlechoicequestion)Supposeyouhavebuiltaneuralnetwork.Youdecidetoinitializetheweightsandbiasestobezero.WhichofthefollowingstatementsareTrue?()AEachneuroninthefirsthiddenlayerwillperformthesamecomputation.Soevenaftermultipleiterationsofgradientdescenteachneuroninthelayerwillbecomputingthesamethingasotherneurons.BEachneuroninthefirsthiddenlayerwillperformthesamecomputationinthefirstiteration.Butafteroneiterationofgradientdescenttheywilllearntocomputedifferentthingsbecausewehave“brokensymmetry”.CEachneuroninthefirsthiddenlayerwillcomputethesamething,butneuronsindifferentlayerswillcomputedifferentthings,thuswehaveaccomplished“symmetrybreaking”asdescribedinlecture.第17題(Singlechoicequestion)Amongthefollowing,whichoneis"hyperparameters"?

()AlearningrateBmodelweightsCmodelbias第18題(Singlechoicequestion)Whichofthefollowingstatementsistrue?()AThedeeperlayersofaneuralnetworkaretypicallycomputingmorecomplexfeaturesoftheinputthantheearlierlayers.BTheearlierlayersofaneuralnetworkaretypicallycomputingmorecomplexfeaturesoftheinputthanthedeeperlayers.第19題(TrueorFalseQuestions)Duringforwardpropagation,intheforwardfunctionforalayerlyouneedtoknowwhattheactivationfunctionisinalayer(Sigmoid,tanh,ReLU,etc.).Duringbackpropagation,thecorrespondingbackwardfunctionalsoneedstoknowwhattheactivationfunctionisforlayerl,sincethegradientdependsonit.()ATrueBFalse第20題(TrueorFalseQuestions)VectorizationallowsyoutocomputeforwardpropagationinanL-layerneuralnetworkwithoutanexplicitfor-loop(oranyotherexplicititerativeloop)overthelayersl=1,2,…,L.()ATrueBFalse第1題(TrueorFalseQuestions)SVMisabinaryclassifier.()ATrueBFalse第2題(Singlechoicequestion)WhatisthetaskofSVM?()AAbinaryclassificationtaskwithy=+1/-1.BAbinaryclassificationtaskwithy=1/0.第3題(TrueorFalseQuestions)SVMisamethodtryingtofindthelargestmarginbetweentwofeatures.()ATrueBFalse第4題(Singlechoicequestion)What’sthemeaningofthemargin?()AThemarginisdefinedintermsofthedistancefromtheboundarytoexamples.BThemarginisbasedonthevalueofthelinearfunction.第5題(Singlechoicequestion)WhichoftheseisthehyperplaneoftheSVM?()ABC第6題(Singlechoicequestion)WhichoftheseisthedistancebetweenH1andorigin?()A

B第7題(TrueorFalseQuestions)InSVM,wecantrytofindauniquesolutionbyrequiringthatthetrainingexamplesareclassifiedcorrectlywithanon-zero“margin”,themarginisdefinedintermsofthedistancefromtheboundarytotheexamplesratherthanbasedonthevalueofthelinearfunction.()ATrueBFalse第8題(Singlechoicequestion)Whichoneisthemethodofachievingthelargestmargin?

()AB第9題(TrueorFalseQuestions)SVMoptimizationisarelaxedquadraticoptimizationproblem.()ATrueBFalse第10題ATrueBFalse第11題(Singlechoicequestion)WhatwillhappenifusinglargeC?()AItwillcausefewviolations.BItwillcausemanyviolations.第12題(TrueorFalseQuestions)LinearSVMcanstillhandleproblemwhentheinputspaceismappedtoahigh-dimensionalfeaturespace.()ATrueBFalse第13題(TrueorFalseQuestions)Gaussiankernelisonetypeofkernelinnon-linearSVM.()ATrueBFalse第14題ATrueBFalse第15題

(Singlechoicequestion)WhatistheconsequenceofusingsmallC?()AItwillcausefewviolations.BItwillcausemanyviolations.第16題(TrueorFalseQuestions)Many(lowdimensional)problemsaresolvedwellbyalinearclassifierwithslack.()ATrueBFalse第17題(Singlechoicequestion)Howcanwegetnon-linearmargincurvesintheoriginalspace?()AMappingexamplestofeaturevectorsandmaximizingalinearmargininthefeaturespace.BJustusethelinearclassifierwithslack.第18題(Singlechoicequestion)What’sthefunctionofparameterC?

()AaccuracyBefficiencyCRegularization第19題(TrueorFalseQuestions)ASVMwithnokernelfunctioncanalsobeseenasusingalinearkernel.()ATrueBFalse第20題(Singlechoicequestion)IfweuseGaussiankernelinSVM,whenwelargethesigma:().Ahighvariance,lowbias.Blowvariance,highbias.Exercise第1題(Singlechoicequestion)AMarkovchainmodelisNOTdefinedby:().AAsetofstates.BAsetoftransitionswithassociatedprobabilities.CAsetofpredictions.第2題(TrueorFalseQuestions)Asetoftransitionswithassociatedprobabilitiesmeansthatthetransitionsemanatingfromagivenstatedefineadistributionoverthepossiblenextstates.()ATrueBFalse第3題ATrueBFalse第4題ATrueBFalse第5題(TrueorFalseQuestions)TheMarkovpropertyspecifiesthattheprobabilityofastatedependsonlyontheprobabilityofthepreviousstate.()ATrueBFalse第6題ATrueBFalse第7題(Singlechoicequestion)What’sthemeaningofNinHMM

=(N,M,A,B,

)?()AThenumberofstatesinthemodel.BThenumberofdistinctobservationsymbolsperstate.CThestatetransitionprobabilitydistribution.DTheobservationsymbolprobabilitydistribution.第8題(Singlechoicequestion)What’sthemeaningofMinHMM?()AThenumberofstatesinthemodel.BThenumberofdistinctobservationsymbolsperstate.CThestatetransitionprobabilitydistribution.DTheobservationsymbolprobabilitydistribution.第9題(Singlechoicequestion)What’sthemeaningofAinHMM?()AThenumberofstatesinthemodel.BThenumberofdistinctobservationsymbolsperstate.CThestatetransitionprobabilitydistribution.DTheobservationsymbolprobabilitydistribution.第10題(Singlechoicequestion)What’sthemeaningofBinHMM?()AThenumberofstatesinthemodel.BThenumberofdistinctobservationsymbolsperstate.CThestatetransitionprobabilitydistribution.DTheobservationsymbolprobabilitydistribution.第11題(Singlechoicequestion)Given:amodel,asetoftrainingsequences.

Do:findmodelparametersthatexplainthetrainingsequenceswithrelativelyhighprobability(goalistofindamodelthatgeneralizeswelltosequenceswehaven’tseenbefore)Whichtaskisright?()ALearningBClassificationCSegmentation第12題

(Singlechoicequestion)Given:asetofmodelsrepresentingdifferentsequenceclasses,atestsequenceDo:determinewhichmodel/classbestexplainsthesequenceWhichtaskisright?()A

LearningBClassificationCSegmentation第13題(Singlechoicequestion)Given:amodelrepresentingdifferentsequenceclasses,atestsequenceDo:segmentthesequenceintosubsequences,predictingtheclassofeachsubsequenceWhichtaskisright?()ALearningBClassificationCSegmentation第14題(Singlechoicequestion)Whatisthemostprobable“path”forgeneratingagivensequence?()AtheForwardalgorithmBtheViterbialgorithmCtheForward-Backward(Baum-Welch)algorithm第15題(Singlechoicequestion)HowlikelyisagivensequenceinHMM?()AtheForwardalgorithmBtheViterbialgorithmCtheForward-Backward(Baum-Welch)algorithm第16題(Singlechoicequestion)HowcanwelearntheHMMparametersgivensasetofsequences?()AtheForwardalgorithmBtheViterbialgorithmCtheForward-Backward(Baum-Welch)algorithm第17題(TrueorFalseQuestions)ExpectationMaximizationalgorithmisafamilyofalgorithmsforlearningprobabilisticmodelsinproblemsthatinvolvehiddenstate.()A

TrueB

False第18題

(Singlechoicequestion)CalculatetheComputationalComplexityofHMMAlgorithms:GivenanHMMwithSstatesandasequenceoflengthL,thecomplexityoftheForward,BackwardandViterbialgorithmsis().AB第19題(Singlechoicequestion)CalculatetheComputationalComplexityofHMMAlgorithms:GivenMsequencesoflengthL,thecomplexityofBaumWelchoneachiterationis().AB第20題(Singlechoicequestion)WhichoneisnottheDP-basedalgorithmsforHMMs?()AForwardBBackwardCViterbiDExpectationMaximizationExercise第1題(Singlechoicequestion)Whichofthefollowingisacharacteristicofunsupervisedlearning?()AEachinputcorrespondstoanoutputlabel.BThepurposeofthetaskistolearntopredictthecorrectlabelbasedontheinputcharacteristics.CTrainingdatahasnocorrespondinglabel.DThenoiseofthelabelwillaffectthequalityofthemodel.第2題(Singlechoicequestion)Whatisthepurposeofunsupervisedlearning?()AExploretherelationshipbetweeninputfeaturesandoutputtags.BEstablishtheinputrepresentationthatcanbeusedfordecision-making,predictthefutureinput,andeffectivelytransfertheinputtoanothermachine.CExploredatastructureswithclassroom/outputguidance.DModelingofknownmodelsbasedonhistoricaldata.第3題(Singlechoicequestion)Whichofthefollowingisnotanunsupervisedlearningmethod?()AClusteringBDimensionReductionCLatentvariablemodelsDRegression第4題(Singlechoicequestion)Whichofthefollowingisnotaclusteringmethod?()APartitionclusteringBHierarchicalclusteringCLocalclusteringDDensity-basedclustering第5題(Singlechoicequestion)①FeatureSelection/Extraction②InterpatternSimilarity③GroupingWhatisthecorrectorderofclustering?()A①②③B①③②C②③①D③①②第6題(Singlechoicequestion)Whichkindofclusteringmethodbuildsaclusteringtreefromdata?()APartitionclusteringBHierarchicalclusteringCGrid-basedclusteringDDensity-basedclustering第7題(Singlechoicequestion)WhichisnotoneofthemethodofPartitionclustering?()AK-meansBK-mediodsCCLARADBRICH第8題(Singlechoicequestion)WhichisnotoneofthemethodofHierarchicalclustering?()ABRICHBROCKCK-meansDChameleon第9題(Singlechoicequestion)WhichisnotoneofthemethodofModel-basedclustering?()AROCKBEMCConceptclusteringDNeuralnetworkbasedclustering第10題(Singlechoicequestion)WhichisthebiggestadvantageofGrid-basedclustering?()ALowmemorycostBHighqualityCFastprocessingspeedDVisualization第11題(Singlechoicequestion)Whichisnotoneoftherecentmethodofunsupervisedlearning?()AVAEBU-netCBERTDSimCLR第12題(Singlechoicequestion)WhatisthetypeofVAE?()AContextualinformationBContrastLearningCReinforcementlearningDReconstruction第13題(Singlechoicequestion)WhichofthefollowingstatementsaboutBERTiswrong?()AThemodelhasmanylayers.BThemodelhasmorethanonekindofembedding.CThemodelhasbothencoderanddecoder.DThemodeloftenperformsbetterindownstreamtasks.第14題(Singlechoicequestion)WhichofthefollowingstatementsaboutTrainingTrickiswrong?()AAllparametersneedtobeinitializedwith0.BWeightsaredrawnfromGaussiandistributionwithfixedmean0andfixedstandarddeviation(0.01),thisisthemostcommoninitializationmethod.CWeightsaredrawnfromproperlyscaleduniformorGaussiandistributionwithzeromeanandaspecialvariance.D第15題(Multiplechoicequestions)WhichisnotoneofthemethodofHierarchicalclustering?()ASTINGBK-mediodsCCLARADWaveCluster正確答案：BC第16題(Singlechoicequestion)Whichkindofclusteringmethodusesamulti-resolutiongriddatastructuretoquantifythespaceintoalimitednumberofunits?()APartitionclusteringBHierarchicalclusteringCGrid-basedclusteringDDensity-basedclustering正確答案：C第17題(Singlechoicequestion)Whichkindofclusteringmethodtriestofindthemodelbehindit,andusesitsprobabilitydistributioncharacteristicsforclustering?()APartitionclusteringBModel-basedclusteringCGrid-basedclusteringDDensity-basedclustering第18題(Singlechoicequestion)Inwhichkindofclusteringmethods,clustersareregardedasdenseobjectregionsseparatedbylow-densityregionsindataspace,andsometimessuchlow-densityregionsareregardedasnoise?()APartitionclusteringBModel-basedclusteringCGrid-basedclusteringDDensity-basedclustering第19題(Multiplechoicequestions)

Whichofthefollowingstatementsistrue?()AThebottom-upstrategygraduallymergessmallcategoriesintolargecategories,whichiscalledcohesion.BThetop-downstrategygraduallymergessmallcategoriesintolargecategories,whichiscalledcohesion.CThebottom-upstrategygraduallysplitslargecategoriesintosmallcategories,whichiscalledsplitting.DThetop-downstrategygraduallysplitslargecategoriesintosmallcategories,whichiscalledsplitting.正確答案：AD第20題(Singlechoicequestion)Whichkindofunsupervisedlearningmethodistoanalyzerelationshipsbetweenasetofdocumentandthetermstheycontainbyproducingasetofconceptsrelatedtothedocumentsandterms?()ALatentvariablemodels

BGraphmodelCAssociationanalysisDLatentsemanticanalysisExercise第1題(Singlechoicequestion)

Whatisthefoundationofartificialintelligence(AI)andsolvesproblemsthatwouldproveimpossibleordifficultbyhumanorstatisticalstandards?()ACNNBANNCGNNDRNN第2題(Singlechoicequestion)

WhichisnotamotivationofCNN?()AExternalstimuliarenottransformedintoelectricalsignalsthroughnerveendings,whicharenottransducedtonervecells(alsocalledneurons).

BNumerousneuronsconstitutethenervecenter.

CThenervecentersynthesizesvarioussignalsandmakesjudgments.

DThehumanbodyrespondstoexternalstimulibasedoninstructionsfromthenervecenter.

第3題(Singlechoicequestion)

WhichisnottheadvantagesofCNNcomparedwithfullconnectionnetwork?()AItneedsfewerparametersBItcostslesstimeforonepassCItiseasiertooverfitDItfocusmoreonlocalfeatures第4題(Singlechoicequestion)

Whichofthefollowingistrue?()AAConvNetiscomprisedofmorethanoneconvolutionallayers(oftenwithapoolingstep)andthenfollowedbyoneormorefullyconnectedlayersasinastandardmultilayerneuralnetwork.BAConvNetiscomprisedofoneormoreconvolutionallayers(oftenwithapoolingstep)andthenfollowedbyoneormorefullyconnectedlayersasinastandardmultilayerneuralnetwork.CAConvNetiscomprisedofoneormoreconvolutionallayers(oftenwithoutanypoolingstep)andthenfollowedbyoneormorefullyconnectedlayersasinastandardmultilayerneuralnetwork.DAConvNetiscomprisedofoneormoreconvolutionallayers(oftenwithapoolingstep)andthenfollowedbymorethanonefullyconnectedlayersasinastandardmultilayerneuralnetwork.

第5題(Singlechoicequestion)

WhichofthefollowingrepresentsCNN?()AB第6題(Multiplechoicequestions)WhatmeasuresdidCNNtaketosolvetheproblemof“Dimensionsaretoolarge,Parametersaretoolarge,difficulttotrain”?()ALocalconnectivityBParametersharingCSubsampleDDropout正確答案：ABC第7題(Singlechoicequestion)

WhatmeasuresdidCNNtaketosolvetheproblemof“ThepositionisnotusedinANN”?()ALocalconnectivityBParametersharingCSubsampledDDropout第8題(Multiplechoicequestions)

WhatmeasuresdidCNNtaketosolvetheproblemof“ThenumberofnetworklayersislimitedanditisdifficulttotrainadeepfullyCNNthroughthegradientdescentmethod.”?()ALocalconnectivityBParametersharingCSubsampledDDropout正確答案：ABC第9題A0B1C-1D-2第10題(Multiplechoicequestions)

WhatadvantagescanparametersharingbringtoCNN?()ASaveparametersBLesstimecostCMoredifficulttooverfitDEasiertodesign正確答案：AC第11題(Singlechoicequestion)

Whatisneededtoavoidthematrixafterconvolutionbecomessmallerandsmaller?()AStrideBFilteringCPaddingDPooling第12題(Singlechoicequestion)

Whichnumberisusuallyusedtobeasthepaddingnumber?()A1B255C-1D0第13題(Singlechoicequestion)

Whichofthefollowingsetsofcorrespondenceiscorrect?()AConvolutionlayer&Poolinglayer-Extractfeatures,Fullconnectedlayer-ClassificationBConvolutionlayer&Poolinglayer-Classification,Fullconnectedlayer-ClassificationCConvolutionlayer&Poolinglayer-Classification,Fullconnectedlayer-ExtractfeaturesDConvolutionlayer&Poolinglayer-Extractfeatures,Fullconnectedlayer-Extractfeatures第14題(Singlechoicequestion)

Poolingwillperformadownsamplingoperationalongwhichdimension?()ATimeBSpatialCChannelDColor第15題(Multiplechoicequestions)

WhatadvantagescanpoolingbringtoCNN?()ASaveparametersBLesstimecostCAddrobustnesstopositionDProvidetranslationinvariance正確答案：ABDC第16題(Singlechoicequestion)

Isitpossibletomimicacomplexmodelwithmorelayersbutnoactivationfunctions?()AYesBNo第17題(Singlechoicequestion)

Whichisnotthepropertyofactivationfunction?()AContinuous.Gradientdescent'srequirements.BTherangeispreferablynotsaturated.Ifthesystemoptimizationentersthesaturationstage,thegradientisapproximately0,andthelearningofthenetworkwillstop.CLinearityDMonotonicity.Whentheactivationfunctionismonotonic,thelossfunctionofthesingle-layerneuralnetworkisconvex,whichisgoodforoptimization.

第18題(Singlechoicequestion)

Whichkindofactivationfunctionisgoodatdetectingdifferencesandiscommonlyusedinbinaryclassificationtasks?()ATanhBSigmoidCReluDLeakyRelu第19題(Singlechoicequestion)

WhyTanhfunctionconvergesfasterthanthesigmoidfunction?()ATanhis

zerocentered

sothatgradientsarenotrestrictedtomoveinoneparticulardirection.BTanhneedlesscomputationInbackpropagationCTanhneedlesscomputationInforwardpropagationDTanhisspeciallysupportedbythehardware第20題(Multiplechoicequestions)

Whichofthefollowingsetsofcorrespondenceiscorrect?()AMeansquareerror-oftenusedinregressionproblemsBCrossentropylossfunction-oftenusedinclassificationproblemsCMeansquareerror-oftenusedinclassificationproblemsDCrossentropylossfunction-oftenusedinregressionproblems正確答案：ABExercise第1題(Singlechoicequestion)

Whichofthefollowingisnotsequencedata?()ATime-seriesdataBGeneSequenceCSpeechsoundsDPhotos第2題(Multiplechoicequestions)WhydoweneedRNN?()AVariableneuronsandParameterSharingamongdifferenttimeBHugeamountofdataCLong-termdependencychallengeDDifferentdomainofdata正確答案：AC第3題(Singlechoicequestion)

WhichisnotthesolutionsofLong-termdependencychallenge?()AUsingrecurrentconnectionswithlongdelaysBCNNCLeakyUnitsDLSTM(gatedRNNs)第4題(Singlechoicequestion)

Whatisthetrainingprocessinwhichthefedbackinputsarenotthepredictedoutputsbutthetargetsthemselves?()ALSTMBDropoutCTeachingforcingDPooling第5題(Singlechoicequestion)

Whatcanbeviewedasadirectedgraphicalmodelthatestimatestheconditionaldistributionofasequence?()ACNNBRNNCGNNDANN第6題(Singlechoicequestion)

Inmanyapplicationswewanttooutputattimetapredictionregardinganoutputwhichmaydependonthewholeinputsequence.Tosolvethisproblemwhatmodelwasproposed?()ABidirectionalRNNsBMASSCCNNDGRU第7題(Singlechoicequestion)

Whichisnotthemomentwhenwewanttooutputattimetapredictionregardinganoutputwhichmaydependonthewholeinputsequence?()ASpeechRecognitionBHandwritingRecognitionCFacerecognitionDBioinformatic

第8題(Singlechoicequestion)

TherearethreeblocksofparametersandassociatedtransformationinRNN,whichisnotincluded?()AFromoutputtohiddenstateBFrominputtohiddenstateCFromhiddenstatetohiddenstateDFromhiddenstatetooutput第9題(Singlechoicequestion)

Whichkindofnetworkhasacomputationalgraphthatgeneralizesthatoftherecurrentnetworkfromachaintoatree?()ADeepRecurrentNetworkBRecursiveNeuralNetworkCConvolutionNeuralNetworkDGraphneuralnetwork第10題(Singlechoicequestion)

ArisingfrommultipleproductofthesameweightmatrixW,whenW'sspectralradiusisbiggerthan1,itleadsto①,otherwise②.()A①gradientexploding②gradientexplodingB①gradientexploding

②gradientvanishingC①gradientvanishing

②gradientexplodingD①gradientvanishing

②gradientvanishing第11題(Singlechoicequestion)

WhichsolutionofsolvingLong-termDependenciesproposestosetthoseweightssuchthattherecurrenthiddenunitsdoagoodjobofcapturingthehistoryofpastinputs,andonlylearntheoutputweights?()AEchoStateNetworksBCombiningshortandlongpathsCLeakyUnitDGatedRNNs第12題(Singlechoicequestion)

WhichsolutionofsolvingLong-termDependenciesproposestouserecurrentconnectionswithlongdelays?

()AEchoStateNetworksBLeakyUnitCCombiningshortandlongpathsDGatedRNNs第13題(Singlechoicequestion)

WhichsolutionofsolvingLong-termDependenciesproposestohaveunitswithlinearself-connectionsandaweightnear1ontheseconnections?()AEchoStateNetworksBLeakyUnitCCombiningshortandlongpathsDGatedRNNs第14題(Singlechoicequestion)

WhichsolutionofsolvingLong-termDependenciesproposestoletnetworklearnstodecidewhentoupdateorforgetstate?()AEchoStateNetworksBLeakyUnitCCombiningshortandlongpathsDGatedRNNs第15題(Singlechoicequestion)

Whichofthefollowingstatementsistrue?()ALSTMislesspowerfulthanGRU,andGRUismoreeasytotrainthanLSTMBLSTMismorepowerfulthanGRU,butGRUislesseasytotrainthanLSTMCLSTMislesspowerfulthanGRU,butGRUismoredifficulttotrainthanLSTMDLSTMismorepowerfulthanGRU,butGRUismoreeasytotrainthanLSTM第16題ALSTMBBERTCCNNDGRU第17題ALSTMBBERTCCNNDGRU第18題(Singlechoicequestion)

WhichsolutionofsolvingLong-termDependenciesproposestodividethesmallderivativebyasmallsecondderivative,whilenotscalingupinthedirectionswherethesecondderivativeislarge?

()AEchoStateNetworksBBetteroptimizationCCombiningshortand

longpathsDGatedRNNs第19題(Singlechoicequestion)

WhichsolutionofsolvingLong-termDependenciesproposestochangethesizeofthegradient?

()AEchoStateNetworksBBetteroptimizationCClippingGradientsDGated

RNNs第20題AbreadthfirstsearchBbeamsearchCgreedysearchDCoresearchExercise第1題(Singlechoicequestion)

Adversarialnetworks(GAN)usuallyconsistofhowmanypart(s)()

A1B2C3D4第2題(Singlechoicequestion)

Thefollowingdescriptionofgenerativeadversarialnetworksisincorrect().AGenerativeadversarialnetworkconsistsoftwoparts:generatoranddiscriminatorBWhenthediscriminatorofadversarialnetworkistraining,itsinputistheimagegeneratedbygeneratorandtherealimagefromthetrainingset.CGAN'sGeneratorsgenerateimagesfromrandomnoise(usuallytakenfromauniformorGaussiandistribution)DSincegenerativeadversarialnetworksareunsupervisedmodels,notrainingdataisrequired第3題(Singlechoicequestion)

Thegeneratoranddiscriminatorofadversarialnetworkcanonlyhavethesamelearningfrequency().AcorrectBincorrect第4題(Singlechoicequestion)

Thegoalofadversarialnetworktrainingistomakethelossfunctionofgeneratoraslowaspossible().AcorrectBincorrect第5題(Singlechoicequestion)

Whichofthefollowingstatementsisnottrueaboutthecostfunctionofgeneratingadversarialnetworks(GAN)?

()AThegeneratorshouldminimizetheaccuracyofthediscriminantmodelD,whilethediscriminatorshouldmaximizetheaccuracyoftrueandfalseclassificationBInordertoachievethegamebalancebetweengeneratoranddiscriminator,thecostfunctionofGANneedstoconsidertheperformanceofbothCBytrainingdiscriminatorandgeneratoralternately,theperformanceofdiscriminatorandgeneratorcanbeimprovedtoreachabalancepointDGenerallyspeaking,GANcanalwaysreachtheminimumvalueofcostfunctionthroughtraining第6題(Singlechoicequestion)

Whichofthefollowingstatementsistrueforthetrainingofgenerativeadversarialnetworks(GAN)?

()AThegeneratorisimplementedbyafeedforwardneuralnetworkordeconvolutiondeepnetworkanditsgoalistomakethegeneratedimagelookliketherealsampleBIfthediscriminatorisoverfitted,thegeneratormaygenerateaverystrangesample第7題(Singlechoicequestion)

Whichofthefollowingstatementsisnottrueaboutthecross-entropycostfunctionsofgen