2026年AWS认证机器学习工程师考试真题题库_第1页
2026年AWS认证机器学习工程师考试真题题库_第2页
2026年AWS认证机器学习工程师考试真题题库_第3页
2026年AWS认证机器学习工程师考试真题题库_第4页
2026年AWS认证机器学习工程师考试真题题库_第5页
已阅读5页,还剩62页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

2026年AWS认证机器学习工程师考试真题题库Question1AMachineLearningEngineerisbuildingacomputervisionmodeltodetectdefectsinmanufacturingpartsonanassemblyline.Thetrainingdatasetconsistsof10millionhigh-resolutionimagesstoredinAmazonS3.Themodeltrainingjobtakesseveraldaystocompleteonasingle`ml.p3.2xlarge`instance.TheEngineerneedstoreducethetrainingtimesignificantlywithoutincurringamassiveincreaseincost,whilealsoensuringthatthemodelconvergestoahighaccuracy.WhichcombinationofstepswillmeettheserequirementsMOSTeffectively?A.EnableSageMakerDistributedDataParallelism(DDP)andusemultiple`ml.p3.16xlarge`instances.Enablemanagedspottrainingandconfigureacheckpointingstrategy.B.UseSageMakerAutomaticModelTuningtofindtheoptimalhyperparametersfaster.Switchtousing`ml.p3.8xlarge`instanceswithoutdistributedtraining.C.EnableSageMakerModelParallelism(SMMP)anduseasingle`ml.p3.16xlarge`instance.StorethedatasetinAmazonEFSinsteadofS3forfasterI/O.D.ConverttheimagestotheTFRecordformat,loadthemintoanAmazonDynamoDBtable,anduseasingle`ml.p3.8xlarge`instancewithincreasedreadthroughput.CorrectAnswer:ADetailedExplanation:Theprimarygoalistoreducetrainingtimesignificantlyforalargedataset(10millionimages)withoutamassivecostincrease.OptionAiscorrect.SageMakerDistributedDataParallelism(DDP)isdesignedspecificallytoscaledeeplearningtrainingacrossmultipleGPUsandinstances.ByusingDDPonmultiple`ml.p3.16xlarge`instances,thetrainingworkloadispartitionedandprocessedinparallel,drasticallyreducingwall-clocktime.ManagedSpotTrainingallowstheuseofspareEC2capacityatuptoa90%discount,mitigatingthecostincreaseofusingmultiplelargeinstances.CheckpointingisessentialwithSpotInstancestosavestatesotrainingcanresumeiftheinstanceisinterrupted.OptionBisincorrect.WhileAutomaticModelTuninghelpsfindbetterhyperparameters,ittypicallyrunsmultipletrainingjobssequentiallyorinparallel,whichusuallyincreasestotalcomputationtimeandcosttofindthebestmodel,ratherthanspeedingupasingletrainingjob'sconvergencetime.Scalingupinstancesizewithoutdistributedlibraries(likeDDPorHorovod)doesnotautomaticallyparallelizethetraininglogicunlesstheframeworksupportsitnativelyandisconfiguredcorrectly,butDDPistheexplicitAWSsolutionforthis.OptionCisincorrect.SageMakerModelParallelism(SMMP)isusedwhenthemodelistoolargetofitintoasingleGPU'smemory(tensorparallelismorpipelineparallelism).Theproblemstatementimpliesthedatasetislarge,notnecessarilythemodelarchitecture.Furthermore,storingdatainEFScanbeslowerthanS3forhigh-throughputrandomaccessrequiredbydistributedtrainingunlesstunedspecifically,andS3isthestandardoptimizedstorageforSageMaker.OptionDisincorrect.ConvertingimagestoaformatlikeTFRecordorShardedRecordIOisgoodpractice,butstoringtheminDynamoDBisananti-patternforthisusecase.DynamoDBisaNoSQLdatabasenotdesignedforstreaminghigh-throughputbinaryblobs(images)toatrainingjob.ItwouldbeprohibitivelyexpensiveandslowcomparedtostreamingdirectlyfromS3.Question2AfinancialservicescompanyisusingAmazonSageMakertotrainafrauddetectionmodelusingtheXGBoostalgorithm.Thedataishighlyimbalanced,witharatioof99:1(non-fraudulent:fraudulent).Themodelachieveshighaccuracybutfailstodetectmostofthefraudulenttransactions.WhichconfigurationchangesshouldtheMLEngineermaketoimprovethemodel'sabilitytodetectfraudulenttransactions?(SelectTWO)A.Adjustthe`scale_pos_weight`hyperparameterinXGBoosttoberoughlyequaltotheratioofnegativetopositiveinstances.B.Increasethe`eval_metric`tofocuson`auc`insteadof`accuracy`.C.UseSMOTE(SyntheticMinorityOver-samplingTechnique)inaSageMakerProcessingjobtobalancethetrainingdataset.D.Reducethe`num_round`hyperparametertopreventoverfittingtothemajorityclass.E.Enableearlystoppingandsetthe`tolerance`parametertoahighervalue.CorrectAnswer:A,CDetailedExplanation:Theproblemisclassimbalancecausingthemodeltobebiasedtowardthemajorityclass(highaccuracy,lowrecallfortheminorityclass).OptionAiscorrect.InXGBoost,the`scale_pos_weight`hyperparameterisspecificallydesignedtohandleimbalanceddata.Ithelpsbalancethepositiveandnegativeweights.Acommonheuristicistosetitto`sum(negativeinstances)/sum(positiveinstances)`.Thisincreasesthepenaltyformisclassifyingtheminority(fraudulent)class.OptionCiscorrect.SMOTEisatechniquetogeneratesyntheticsamplesfortheminorityclass.ByusingaSageMakerProcessingjobtoapplySMOTEbeforetraining,themodelseesabalanceddataset,whichallowsittolearnthecharacteristicsoftheminorityclassmoreeffectively.OptionBispartiallycorrectinpractice(changingthemetricisgoodforevaluation),butitdoesnotinherentlyimprovethemodel'sabilitytodetectfraud;itjustchangeshowyoumeasuresuccess.Thequestionasksforconfigurationchangestoimprovedetection.However,betweenA/CandB,AandCaddresstherootcause(imbalance)directly.Inthecontextof"configurationchanges"tothetrainingpipelineoralgorithm,AandCaremoredirectinterventions.Self-correction:Whilechangingthemetriciscrucialformonitoring,itdoesn'tfixthemodel.AandCfixthemodel.OptionDisincorrect.Reducing`num_round`(numberofboostingrounds)mightactuallyunderfitthemodelandpreventsitfromlearningthecomplexpatternsoftheminorityclass.Overfittingisarisk,butsimplyreducingroundsusuallyharmsperformance.OptionEisincorrect.Earlystoppingisgoodforgeneralization,butincreasingthe`tolerance`makesitstopsooner,potentiallybeforethemodelhaslearnedtheminorityclasspatterns.Question3AnMLEngineerisdeployingareal-timerecommendationmodelusingAmazonSageMaker.ThemodelrequiresaccesstoalookuptablestoredinAmazonDynamoDBtofetchuserprofilefeaturesatinferencetime.Theinferencelatencymustbekeptbelow100milliseconds.Thecurrentimplementation,whichqueriesDynamoDBdirectlyfromtheinferencescript,oftenexceedsthelatencyrequirementduringpeaktraffic.HowcantheEngineeroptimizetheinferencepipelinetomeetthelatencyrequirement?A.EnableDynamoDBAccelerator(DAX)forthelookuptableandupdatetheinferenceendpointtousetheDAXclusterendpoint.B.MigratethelookuptabletoAmazonS3andhostitasalocalfilesystemontheSageMakerendpointinstanceusing`MountPoint`.C.IncreasetheinstancecountoftheSageMakerendpointtohandlethepeaktrafficload.D.BatchtheinferencerequestsandprocessthemasynchronouslyusingSageMakerAsynchronousInference.CorrectAnswer:ADetailedExplanation:ThebottleneckisthelatencyofreadingfromDynamoDBduringinference.OptionAiscorrect.DynamoDBAccelerator(DAX)isafullymanaged,highlyavailable,in-memorycachethatcanreducereadlatencyforDynamoDBfrommillisecondstomicroseconds.BypointingtheinferencescripttotheDAXclusterendpointinsteadofthestandardDynamoDBendpoint,theEngineercansignificantlyreducetheI/Owaittime,keepingtotalinferenceunder100ms.OptionBisincorrect.Whilehostingafilelocallyisfast,itisnotfeasiblefora"lookuptable"thatimpliesdynamicupdatesorasizethatmightexceedinstancememory(ifit'salargeuserprofileDB).Also,updatingthelocalfileoneveryendpointinstancewhendatachangesisoperationallycomplexandslowtopropagate.OptionCisincorrect.Increasingtheinstancecountincreasesthroughput(requestspersecond)butdoesnotreducethelatencyofanindividualrequest.Ifasinglerequesttakes200msduetoDBlookup,scalingoutwon'tmakethatsinglerequestfaster.OptionDisincorrect.AsynchronousInferenceisdesignedforpayloadswithlargesizesorlongprocessingtimeswhereanimmediateresponseisnotneeded.Therequirementis"real-time"withastrictlatencySLA(<100ms),whichimpliesSynchronous(Real-time)Inference.Question4Acompanyhastrainedadeeplearningmodelfornaturallanguageprocessing(NLP)usingTensorFlowonAmazonSageMaker.Themodelsizeis10GB.TheMLEngineerwantstodeploythismodeltoaSageMakerAsynchronousInferenceendpointtoprocesslargebatchesofdocuments.TheEngineerneedstoensurethattheendpointscalesdowntozerowhentherearenorequeststominimizecosts.WhichofthefollowingisarequiredconfigurationtoenablescalingtozeroforAsynchronousInference?A.Configuretheinitialinstancecountto0andsetthetargettrackingscalingpolicybasedon`ApproximateBacklogSizePerInstance`.B.Configuretheinitialinstancecountto1andsetascheduledscalingpolicytoreducecountto0duringnon-businesshours.C.Enable"ServerlessInference"whichautomaticallyscalestozero.D.Configurealifecycleconfigurationhookin`on_create`todownloadthemodelfromS3onlywhentheinstanceisspawned.CorrectAnswer:ADetailedExplanation:OptionAiscorrect.AsynchronousInferencesupportsscalingtozero.Toachievethis,youmustconfiguretheAutoScalinggroupfortheendpoint.Specifically,youneedtosettheminimumcapacity(initialinstancecount)to0.Totriggerscaling,youuseatargettrackingpolicymetriclike`ApproximateBacklogSizePerInstance`or`ApproximateBacklogSize`(forthequeue).Whenthebacklogiszero,thepolicycanscaleintozero.OptionBisincorrect.Scheduledscalingallowsforspecifictimes,butitdoesn'tdynamicallyscaletozerobasedondemand.Also,simplysettingcountto0mightnotbesupportedorbehaveexactlylikethenative"scaletozero"featuredesignedforAsynchronousInferencewhichmanagesthequeueprocessinglogic.OptionCisincorrect.ServerlessInferenceisaseparatehostingoption(usingAWSLambda)withapayloadsizelimit(usually6MBforrequest,upto4MBresponse).Themodelis10GB,andwhileitmightcompress,thecontextof"AsynchronousInference"inthepromptsuggestsusingthededicatedAsynchronousInferencefeaturewhichsupportslargerpayloadsandmulti-modelendpoints,whereasServerlesshasstrictmemory/timelimits.OptionDisincorrect.Lifecyclehooksareusefulfordownloadingartifacts,buttheydonotenablethe"scaletozero"capability.ScalingiscontrolledbytheAutoScalingpolicy,notthelifecyclehookitself.Question5YouareanMLEngineerworkingonapredictivemaintenancesystemforIoTdevices.DataisingestedviaAWSIoTCoreandstoredinAmazonS3.Youneedtoperformfeatureengineeringonthistime-seriesdatabeforetrainingamodel.Thefeatureengineeringlogicinvolvescalculatingrollingaveragesandlagfeaturesoverslidingwindows,whichrequiresstatemanagement.WhichserviceistheMOSTefficientandcost-effectiveforthistransformation?A.AWSGlueSparkJobsB.AmazonSageMakerProcessingJobsC.AWSLambdaD.AmazonEMRCorrectAnswer:BDetailedExplanation:ThetaskisfeatureengineeringaspartofanMLpipeline(preprocessingbeforetraining).OptionBiscorrect.AmazonSageMakerProcessingJobsisthepurpose-builtserviceforrunningfeatureengineering,datapreprocessing,andmodelevaluationtasks.Itallowsyoutorunscripts(Python,Spark,etc.)onfullymanagedinfrastructure.ItistightlyintegratedwithSageMakerPipelines.SincethedataisinS3andthegoalisML,ProcessingJobsallowsyoutoeasilyspinuptherequiredcompute(e.g.,`ml.m5.xlarge`or`ml.r5.xlarge`),performtheheavytransformationsusingSparkorSklearn,andsavetheoutputbacktoS3fortraining,allwithoutmanagingclusters.OptionAisincorrect.WhileAWSGlueisexcellentforETLandcanhandleSparkjobs,ifthisisastepstrictlywithinanMLworkflow(traininghappensnext),SageMakerProcessingisoftenpreferredforitsseamlessintegrationandabilitytousethesamecontainerimagesusedfortraining.However,Glueisastrongcontender.ThedifferentiatorhereisoftencostandoperationaloverheadforML-specificworkflows.SageMakerProcessingallowsyoutousetheinputchannels(S3)andoutputchannels(S3)definitionsveryeasilyforMLpipelines.Self-correction:Glueisvalid,butSageMakerProcessingisspecificallydesignedforthe"Preprocessing->Training->Evaluation"loop.OptionCisincorrect.AWSLambdahasstrictexecutiontimelimits(15minutes)andmemorylimits.Processinglargetime-seriesdatawithrollingwindowsoftenrequiresmoretimeandmemorythanLambdaprovides.OptionDisincorrect.AmazonEMRisabigdataplatformformassivescale.Whileitcandothejob,itismoreoperationallyheavyandcostlyforaperiodicfeatureengineeringjobcomparedtothetransientnatureofSageMakerProcessingJobs,whichspinupanddownautomatically.Question6AnMLEngineerisusingAmazonSageMakerGroundTruthtolabeladatasetofimages.Thedatasetissensitive,andthecompanyrequiresthatthelabelingworkforcemustnotdownloadtheimagestotheirlocalmachines.Theyalsowanttoensurethatonlyaspecificgroupofinternalemployeesareallowedtolabelthedata.Whichconfigurationmeetsthesesecurityrequirements?A.UseaprivateworkforcewithAmazonCognitoandenableVPC-onlyaccessfortheGroundTruthlabelingportal.B.Useapublicworkforce(AmazonMechanicalTurk)andapplyadataencryptionkey.C.UseavendorworkforceandsignaNon-DisclosureAgreement(NDA).D.UseSageMakerGroundTruthwithaprivateworkforceandconfigurethelabelingjobtostreamdatadirectlyfromS3withoutlocalcachingoptions.CorrectAnswer:ADetailedExplanation:OptionAiscorrect.AprivateworkforceallowsyoutorestrictaccesstospecificinternalemployeesviaCognitoorOIDC.Toensureimagesarenotdownloadedandtrafficstayswithinthecorporatenetwork,youcanconfigurethelabelingportaltorunwithinaVPC.Thisensuresthatthedataneverleavesthecompany'scontrolledinfrastructure,andworkersaccesstheportalviaprivatelinks.OptionBisincorrect.Apublicworkforce(MTurk)consistsofexternalworkers.Thisviolatestherequirementthatonlyinternalemployeeslabelthedata.OptionCisincorrect.Avendorworkforceconsistsofexternalthird-partycompanies.WhileNDAsareused,itdoesnotstrictlyenforce"internalemployeesonly"orpreventdownloadstechnicallyinthesamewayaVPCsetupdoes.OptionDisincorrect.Whileusingaprivateworkforceiscorrect,simplyconfiguringthejobtostreamfromS3doesnottechnicallypreventadetermineduserfromdownloadingimagesfromthebrowserinterface(e.g.,right-clicksave).VPCintegrationisthespecificsecurityfeaturethatrestrictstheenvironment.Question7AdatascientisthasdevelopedamodelusingtheSageMakerbuilt-inlinearlearneralgorithm.Themodelperformanceissuboptimal.ThedatascientistwantstouseAmazonSageMakerExperimentstotrackmultipletrialswithdifferenthyperparameters.WhichofthefollowingistheBESTwaytoimplementthistrackingefficiently?A.CreateaSageMakerExperimentanduseaSageMakerProcessingJobtoloopthroughhyperparametercombinationsandstarttrainingjobsaschildtrials.B.UseSageMakerAutomaticModelTuning(HPO)whichautomaticallycreatesanExperimentandtracksalltrialsaspartofthetuningjob.C.ManuallycreateaSageMakerTrainingJobforeachhyperparametercombinationandtagthemwiththeExperimentnameusingAWSSDKs.D.UseSageMakerClarifytorunbiasdetectionandautomaticallylogthehyperparameterstoCloudWatch.CorrectAnswer:BDetailedExplanation:OptionBiscorrect.SageMakerAutomaticModelTuning(HyperparameterOptimization-HPO)isthemostefficientwaytotrackmultipletrials.Whenyoustartatuningjob,SageMakerautomaticallycreatesanExperiment(ifnotspecified)andaTrialforeverytrainingjobitlaunches.Itmanagestheorchestration,tracking,andselectionofthebesthyperparametersbasedontheobjectivemetric.OptionAisincorrect.UsingaProcessingJobtoorchestratealoopisamanualapproachtoHPO.Whileitworks,itisre-inventingthewheelandislessefficientthanusingthenativeHPOservice,whichhasbuilt-inBayesianoptimizationstrategies.OptionCisincorrect.Manuallycreatingtrainingjobsisoperationallytediousandpronetoerrors.ItlackstheintelligentsearchstrategyofHPO.OptionDisincorrect.SageMakerClarifyisformodelexplainabilityandbiasdetection,notforhyperparameteroptimizationtracking.Question8AMachineLearningEngineerneedstoserveaPyTorchmodelforreal-timeinference.Themodelisupdatedfrequently(everyfewhours).TheEngineermustensurethatthedeploymentprocessdoesnotresultindowntimefortheend-users.WhichdeploymentstrategyinSageMakermeetsthe`ZeroDowntime`requirement?A.Createanewendpointconfigurationwiththeupdatedmodelandcall`update_endpoint`withthenewconfiguration.Set`DeploymentConfig`to`RollingUpdate`with`MaximumExecutionTimeoutInSeconds`.B.Createanewendpointconfigurationanduse`create_endpoint`tooverwritetheexistingendpoint.C.UseSageMakerServerlessInferenceandaliastheversions.D.Manuallydeletetheoldendpointandcreateanewonewiththesamename.CorrectAnswer:ADetailedExplanation:OptionAiscorrect.The`update_endpoint`APIinSageMakersupportsarollingupdatestrategy.Byspecifyinga`RollingUpdate`policy,SageMakerwilllaunchthenewinstances(withthenewmodel),waitforthemtopasshealthchecks,andthenshiftthetraffictothembeforeterminatingtheoldinstances.Thisensuresthatthereisalwayscapacityavailabletohandlerequests,resultinginzerodowntime.OptionBisincorrect.`create_endpoint`requiresauniqueendpointname.Youcannotoverwriteanexistingendpoint;youmustupdateit.OptionCisincorrect.WhileServerlessInferencehandlesupdateswell,"ZeroDowntime"inthecontextofprovisionedendpointsisspecificallyhandledbytheRollingUpdateconfiguration.Also,Serverlesshascoldstartsandpayloadlimitswhichmightnotfitthegeneral"real-time"requirementiflatencyiscritical.However,strictlyspeaking,updatingaServerlessendpointisatomic,butRollingUpdateisthestandardanswerforprovisionedendpointsavoidingdowntimeduringtheswap.OptionDisincorrect.Deletingandrecreatingtheendpointcausessignificantdowntimeandpotentiallossofserviceavailability.Question9YouaredesigninganMLpipelineonAWS.TherawdataisuploadedtoanS3bucket.Thepipelinemusttriggerautomaticallywhenevernewdataisuploaded.Thepipelineconsistsof:1.DataPreprocessing(Spark),2.ModelTraining(XGBoost),3.ModelEvaluation(Python).WhichserviceorchestratesthesestepsMOSTeffectively?A.AWSStepFunctionsB.AmazonEventBridgeC.AWSLambdaD.AmazonSageMakerPipelinesCorrectAnswer:DDetailedExplanation:OptionDiscorrect.AmazonSageMakerPipelinesisanativeservicespecificallydesignedforbuilding,automating,andmanagingend-to-endMLworkflows.ItintegratesdirectlywithSageMakerProcessing,Training,andProcessingjobs.Itsupportsparameterizedexecution,caching,andconditionalexecution(e.g.,onlyregistermodelifaccuracy>threshold).ItcanbetriggeredbyEventBridge(S3event)easily.OptionAisincorrect.WhileStepFunctionscanorchestrateSageMakerjobs,SageMakerPipelinesoffersahigher-levelabstractiontailoredforML,includingeasyvisualizationofthepipelineDAGandlineagetracking(trackingwhichmodelcamefromwhichdata),whichrequiresmanualimplementationinStepFunctions.OptionBisincorrect.EventBridgeisthetriggermechanism,nottheorchestratorofthemulti-stepworkflowlogic.OptionCisincorrect.Lambdaisnotsuitableforlong-runningorchestrationofmultipledistributedjobsandhasstricttimeoutlimits.Question10AnMLEngineeristrainingamodelonadatasetwith50features.Themodelisoverfittingthetrainingdata.TheEngineerdecidestouseL1regularizationtoencouragesparsity.Whichbuilt-inSageMakeralgorithmsupportsL1regularizationeffectivelyforthisscenario?A.XGBoostB.K-MeansC.FactorizationMachinesD.LinearLearnerCorrectAnswer:DDetailedExplanation:OptionDiscorrect.TheSageMakerLinear`Learner`algorithmisdesignedforregressionandclassificationandexplicitlysupportsL1regularization(viathe`l1`hyperparameter)toproducesparsemodels.Itishighlyoptimizedforthispurpose.OptionAisincorrect.WhileXGBoostsupportsregularization(alphaforL1,lambdaforL2),thequestioncontextoftendistinguishes"LinearLearner"asthego-toforhigh-dimensionalsparselinearmodelswhereL1isaprimaryfeature.However,XGBoostdoessupportL1.Re-evaluating:Thepromptasksforthealgorithmthatsupportsit"effectively"forthisscenario(50features,overfitting,sparsity).Bothsupportit.However,LinearLearnerisspecificallyalinearmodelwhereL1isastandardtechniqueforfeatureselection.XGBoostistree-based.InthecontextofAWSexams,LinearLearnerisstronglyassociatedwithL1/L2regularizationforlinearmodels.Let'slookattheoptionsagain.FactorizationMachinesareforrecommendation.K-Meansisclustering.It'sbetweenXGBoostandLinearLearner.Ifthegoalis"encouragesparsity"inalinearcontext,LinearLearneristhemostdirectanswer.XGBoostusesL1onweights,buttreesdon'tbecome"sparse"inthesamewaylinearcoefficientsdo(theyjustdon'tsplit).LinearLearneristheintendedanswerforlinearregularizationquestions.OptionBisincorrect.K-Meansisforclustering,notsupervisedlearningwithregularization.OptionCisincorrect.FactorizationMachinesareusedforrecommendationandclassification/regressiononsparsedata(likeclick-throughrates),butLinearLearneristhestandardanswerforgenerallinearregressionwithL1.Question11AcompanyhasatrainedmodelstoredinAmazonS3.Theywanttodeploythismodeltoanon-premisesserverthatcannotcommunicatedirectlywiththepublicinternet.Theon-premisesserverhasanAWSIoTGreengrasscoreinstalled.Howcanthemodelbedeployedtotheon-premisesdevice?A.UseSageMakerNeotocompilethemodelforthetargethardware,thendeployitusingAWSIoTGreengrass.B.DownloadthemodelmanuallytoaUSBdriveandtransferittotheon-premisesserver,thenloaditusingTensorFlow.C.UseAWSStorageGatewaytomounttheS3bucketontheon-premisesserverandloadthemodel.D.UseSageMakerBatchTransformtoruninferenceonthedataandsendtheresultstotheon-premisesserver.CorrectAnswer:ADetailedExplanation:OptionAiscorrect.AWSIoTGreengrassallowsyoutorunMLinferenceonedgedevices.SageMakerNeocancompilethemodeltoanoptimizedformatforthespecifichardwarearchitectureoftheon-premisesserver(e.g.,ARM,Intel).Thecomponent(artifact)canbedeployedtotheGreengrasscore,evenifthecoreisinarestrictednetwork,aslongasitcancommunicatewiththeIoTCloudservice(orvialocaldeploymentifcompletelyoffline,butusuallymanagedviaIoT).OptionBisincorrect.Whilephysicallypossible,itisnotanautomatedorscalableAWS-nativesolutionforMLdeployment.OptionCisincorrect.AWSStorageGatewayrequiresconnectivitytotheAWSserviceforcachingandoperations.Ifthedevicecannotcommunicatewiththepublicinternet,thissetupiscomplex(requiresDirectConnectorVPN)anddoesn'tsolvethe"deployment"aspectautomatically.OptionDisincorrect.BatchTransformrunsinAWSCloud,noton-premises.Question12YouareusingAmazonSageMakerCanvastoenablebusinessanalyststobuildmodelswithoutwritingcode.YouneedtoensurethattheanalystscanonlyaccessspecificdatasetsinAmazonS3thathavebeenpre-approvedforanalysis.TheyshouldnothaveaccesstootherS3buckets.Howdoyouconfigurethis?A.CreateanIAMroleforSageMakerCanvaswithapolicythatexplicitlyallows`s3:GetObject`and`s3:ListBucket`onthespecificS3bucketsARNs,anddenyallotherS3actions.B.UseSageMakerCanvasdataconnectorsandconfigureaVPCEndpointpolicytorestrictaccess.C.EnableAWSMacietomonitoraccessandalertonunauthorizedaccess.D.CreateaSageMakerDomainandassigntheanalystsauserprofilewiththedefaultexecutionrolerestrictedbyIAMscope.CorrectAnswer:ADetailedExplanation:OptionAiscorrect.SecurityinSageMakerisgovernedbyIAM.TheexecutionroleattachedtotheSageMakerCanvasuserprofiledictateswhatAWSresourcesthesessioncanaccess.Bycreatingapolicywithexplicit`Allow`forspecificS3ARNsandanexplicit`Deny`for`*`(orrelyingontheimplicitdeny),youenforcetheleastprivilegeprinciplerequired.OptionBisincorrect.WhileVPCendpointsaregoodfornetworkisolation,IAMrolepermissionsaretheprimarymechanismforcontrollingS3objectaccesswithintheapplicationlogic.OptionCisincorrect.Macieisadatasecurityanddataprivacydiscoveryservice,notanaccesscontrolenforcementmechanismforruntimeapplicationlogic.OptionDisincorrect.Simplyassigningauserprofiledoesn'trestrictaccessunlesstheRoleassociatedwiththatprofilehastherestrictivepermissionsdefined(asdescribedinOptionA).Question13AnMLEngineerismonitoringaproductionmodelusingAmazonSageMakerModelMonitor.Thebaselinedatasetusedfordriftdetectionhasafeature`transaction_amount`withameanof\100.WhichconstrainttypeshouldtheEngineerconfiguretodetectthisspecificshift?A.DataQualityB.ModelBiasC.ModelQualityD.FeatureAttributionDriftCorrectAnswer:ADetailedExplanation:OptionAiscorrect.DataQualityconstraintsinModelMonitorareusedtodetectstatisticaldriftintheinputfeatures(like`transaction_amount`).Itcomparesthestatisticsoftheproductiondataagainstthebaselinestatistics(mean,stddev,etc.)andalertsiftheyfalloutsidedefinedthresholds.OptionBisincorrect.ModelBiasmonitorsforbiasinpredictions(e.g.,disparateimpactondifferentgroups),notthestatisticaldistributionofinputfeatures.OptionCisincorrect.ModelQualitymonitorstheperformancemetrics(accuracy,F1,etc.)ofthemodel,typicallycomparinggroundtruthlabelsagainstpredictions.Whileadriftinfeaturesmightcausequalityloss,thespecificdetectionofthefeaturemeanshiftisaDataQualitytask.Op

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

最新文档

评论

0/150

提交评论