深度学习及其应用_ 课件 11-自动化机器学习.pptxVIP

AutoMachineLearningDeepLearninganditsApplication

AutoMLPipelineDataCollectionFeatureEngineeringModelandAlgorithmSelectionHyper-parameterOptimizationQuanmingYao,MengshuoWang,HugoJE,etal.Takinghumanoutoflearningapplications:Asurveyonautomatedmachinelearning[J].arXivpreprintarXiv:1810.13306,2018.2SJTUDeepLearningLecture.

DefinitionofAutoMLDefinitionofMachineLearningAcomputerprogramissaidtolearnfromexperienceEwithrespecttosomeclassesoftaskTandperformancemeasurePifitsperformancecanimprovewithEonTmeasuredbyPDefinitionofAutoMLAutoMLattemptstoconstructmachinelearningprograms(specifiedbyE,TandPintheabovedefinition),withouthumanassistanceandwithinlimitedcomputationalbudgetsTargetQuanmingYao,MengshuoWang,HugoJE,etal.Takinghumanoutoflearningapplications:Asurveyonautomatedmachinelearning[J].arXivpreprintarXiv:1810.13306,2018.3SJTUDeepLearningLecture.

ThreeGoalsGoodPerformancegoodgeneralizationperformanceacrossvariousinputdataandlearningtaskscanbeachievedNoAssistancefromHumanconfigurationscanbeautomaticallydoneformachinelearningtoolsHighComputationalEfficiencytheprogramcanreturnanreasonableoutputwithinalimitedbudget4SJTUDeepLearningLecture.

BasicFrameworkEvaluator:measuretheperformanceofthelearningtoolswithconfigurationsprovidedbytheoptimizerOptimizer:updateorgenerateconfigurationsforlearningtoolsQuanmingYao,MengshuoWang,HugoJE,etal.Takinghumanoutoflearningapplications:Asurveyonautomatedmachinelearning[J].arXivpreprintarXiv:1810.13306,2018.5SJTUDeepLearningLecture.

AutoMLTasksAutoMLforDataFeatureengineering(TraditionalMachineLearning)AutoAugmentation(DeepLearning)AutoMLforModelModelSelection(TraditionalMachineLearning)NeuralArchitectureSearch(DeepLearning)AutoMLforHyper-parametersHyper-parameterOptimization6SJTUDeepLearningLecture.