斯坦福大学机器学习课程讲义第一讲――机器学习介绍

IntroductionWelcomeMachineLearningAndrewNgAndrewNgAndrewNgSPAMAndrewNgMachineLearning-GrewoutofworkinAI-NewcapabilityforcomputersExamples:-DatabaseminingLargedatasetsfromgrowthofautomation/web.E.g.,Webclickdata,medicalrecords,biology,engineering-Applicationscan’tprogrambyhand.E.g.,Autonomoushelicopter,handwritingrecognition,mostofNaturalLanguageProcessing(NLP),ComputerVision.AndrewNgMachineLearning-GrewoutofworkinAI-NewcapabilityforcomputersExamples:-DatabaseminingLargedatasetsfromgrowthofautomation/web.E.g.,Webclickdata,medicalrecords,biology,engineering-Applicationscan’tprogrambyhand.E.g.,Autonomoushelicopter,handwritingrecognition,mostofNaturalLanguageProcessing(NLP),ComputerVision.-Self-customizingprogramsE.g.,Amazon,Netflixproductrecommendations-Understandinghumanlearning(brain,realAI).AndrewNgAndrewNgIntroductionWhatismachinelearningMachineLearningAndrewNg•ArthurSamuel(1959).MachineLearning:Fieldofstudythatgivescomputerstheabilitytolearnwithoutbeingexplicitlyprogrammed.•TomMitchell(1998)Well-posedLearningProblem:AcomputerprogramissaidtolearnfromexperienceEwithrespecttosometaskTandsomeperformancemeasureP,ifitsperformanceonT,asmeasuredbyP,improveswithexperienceE.MachineLearningdefinitionTemplatevertLeftWhite2Classifyingemailsasspamornotspam.Watchingyoulabelemailsasspamornotspam.Thenumber(orfraction)ofemailscorrectlyclassifiedasspam/notspam.Noneoftheabove—thisisnotamachinelearningproblem.Supposeyouremailprogramwatcheswhichemailsyoudoordonotmarkasspam,andbasedonthatlearnshowtobetterfilterspam.WhatisthetaskTinthissetting?“AcomputerprogramissaidtolearnfromexperienceEwithrespecttosometaskTandsomeperformancemeasureP,ifitsperformanceonT,asmeasuredbyP,improveswithexperienceE.”AndrewNgMachinelearningalgorithms:-Supervisedlearning-UnsupervisedlearningOthers:Reinforcementlearning,recommendersystems.Alsotalkabout:Practicaladviceforapplyinglearningalgorithms.AndrewNgAndrewNgIntroductionSupervisedLearningMachineLearningAndrewNg010020030040005001000150020002500Housingpriceprediction.Price($)in1000’sSizeinfeet2Regression:Predictcontinuousvaluedoutput(price)SupervisedLearning“rightanswers”givenAndrewNgBreastcancer(malignant,benign)ClassificationDiscretevaluedoutput(0or1)Malignant?1(Y)0(N)TumorSizeTumorSizeAndrewNgTumorSizeAge-ClumpThickness-UniformityofCellSize-UniformityofCellShape…TemplatevertLeftWhite2Treatbothasclassificationproblems.Treatproblem1asaclassificationproblem,problem2asaregressionproblem.Treatproblem1asaregressionproblem,problem2asaclassificationproblem.Treatbothasregressionproblems.You’rerunningacompany,andyouwanttodeveloplearningalgorithmstoaddresseachoftwoproblems.Problem1:Youhavealargeinventoryofidenticalitems.Youwanttopredicthowmanyoftheseitemswillselloverthenext3months.Problem2:You’dlikesoftwaretoexamineindividualcustomeraccounts,andforeachaccountdecideifithasbeenhacked/compromised.Shouldyoutreattheseasclassificationorasregressionproblems?AndrewNgAndrewNgIntroductionUnsupervisedLearningMachineLearningAndrewNgx1x2SupervisedLearningAndrewNgUnsupervisedLearningx1x2AndrewNgAndrewNgAndrewNgAndrewNgAndrewNg[Source:Su-InLee,DanaPe’er,AimeeDudley,GeorgeChurch,DaphneKoller]GenesIndividualsAndrewNgOrganizecomputingclustersSocialnetworkanalysisImagecredit:NASA/JPL-Caltech/E.Churchwell(Univ.ofWisconsin,Madison)AstronomicaldataanalysisMarketsegmentationAndrewNgCocktailpartyproblemMicrophone#1Microphone#2Speaker#1Speaker#2AndrewNg[AudioclipscourtesyofTe-WonLee.]Microphone#1:Microphone#2:Microphone#1:Microphone#2:Output#1:Output#2:Output#1:Output#2:AndrewNgCocktailpartyproblemalgorithm[W,s,v]=svd((repmat(sum(x.*x,1),size(x,1),1).*x)*x');[Source:SamRoweis,YairWeiss&EeroSimoncelli]Templateblock2x2White1Orderingofbuttonsis:1324Ofthefollowingexamples,whichwouldyouaddressusinganunsupervisedlearningalgorithm?(Checkallthatapply.)Givenadatabaseofcustomerdata,automaticallydiscovermarketsegmentsandgroupcustomersintodifferentmarketsegments.Givenemaillabeledasspam/notspam,learnaspamfilter.Givenasetofnewsarticlesfoundontheweb,groupthemintosetofarticlesaboutthesamestory.Givenadatasetofpatientsdiagnosedaseitherhavingdiabetesornot,learntoclassifynewpatientsashavingdiabetesornot.AndrewNg