机器学习概论机器学习概论 (1).pdf

《机器学习概论机器学习概论 (1).pdf》由会员分享，可在线阅读，更多相关《机器学习概论机器学习概论 (1).pdf（55页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、Welcome to Introduction to Machine Learning!2010.3.51*Images come from InternetCourse InstructorsIntroduction to Machine Learning:Introduction2Teacher:ZHANG Min?Associate Professor Information Retrieval Group DCST HP:http:/ little bit about MinExperiencesAchieved Bachelor and PhD in 1999 and 2003 re

2、spectively in DCST,Tsinghua Uni.Now Associate Professor in THUIR group,CST Dept.Visiting scholar/researcher in NUS,DFKI(Germany),Kyoto University,City University of HongKong,MSRA,and NUSResearch Interests:Information Retrieval&Recommendation,user behavior analysis,data mining,machine learning Achiev

3、ements and AwardsPublished over 100 papers on important international conferences and journals on related fieldsSIGIR,IJCAI,WWW,CIKM,WSDM,JIR,JASIST,3500 citations,H-index 32(Google Scholar,Feb.2019)Multiple Top Performances in TREC and NTCIR since 2002.One of the top rankedWorld-wide IR researchers

4、Beijing Science and Technology Progress Award,1stprize,2015Excellent Young Faculty Teaching Award,Tsinghua UniversityInformation Retrieval:Introduction3A little bit about MinAcademic Activities and Industry ConnectionsVice Director of the AI institute of CS Dept.Vice Director of MOE MSRA Key Lab on

5、Web and multimedia,Tsinghua UniversityAssociate Editor of Transaction on Information Systems(TOIS)(CCF A journal)and reviewers for multiple top journalsProgram Chair of SIGIR19Tutorials,WSDM19Workshop,SIGIR18 Short Papers,EVIA19,WSDM17,AIRS16Area Chair,Senior PC or PC forIJCAI,SIGIR,WWW,KDD,WSDM,CIK

6、M,ACLCommittee members of China AssociationsChinese Information Processing Committee(CCF)Information Retrieval Committee(CIPS)Machine Learning Committee(CAAI)The first Machine Learning course in CS Dept.,Tsinghua Univ.(since 2003)Information Retrieval:Introduction4Course InformationIntroduction to M

7、achine Learning:Introduction5Key pointsBasic conceptsFundamental learning theoryClassical/Important algorithms Problem definitionBasic ideaAlgorithm design and analysisFuture workAnalyses(problem,features,results)Ability to solve practical problemsStudy materialsIntroduction to Machine Learning:Intr

8、oduction6Slides and your notesRecommended reference(partially as textbook):Tom Mitchell,Machine Learning.McGraw-Hill(China machine press,English version)?(?Other good referencesIntroduction to machine learning?EthemAlpaydin?(?Pattern classification(Richard O.Duda,etc)?(?(?TopicsIntroduction to Machi

9、ne Learning:Introduction7 1.Introduction 2.Decision Tree learning 3.Bayesian learning(Nave Bayes)4.Hidden Markov Model 5.SVM and kernel-based learning 8.Overview6.Instance-based learning:KNN,distance-weighted NN 7.Unsupervised learning,clusteringNotNotinvolve(due to courses redundancy,time or/and di

10、fficulty issues)Introduction to Machine Learning:Introduction8Neural Network Basics(NN)Have been introduced in AI courseAnother courses for undergraduates and graduate students in DCSTBut the advanced NN will be introduced in brief overview of Deep Learning section of this courseEvolutionary Computa

11、tionGenetic Algorithms in AI courses(undergraduate or graduate)Insights of Probabilistic Graphical ModelsLater for graduate studentsBut the overview and basic concepts will be introduced in this courseActive learning,Semi-supervised learning,reinforcement learningLater for graduate studentsIntroduct

12、ion to Machine Learning:Introduction9A short but important part of this courseApplication topicsProblem understanding and designFeatures selectionAnalysis and discussionAnd Good books/Big idea/GurusFancy new techniquesCoffee TimeGradingIntroduction to Machine Learning:Introduction10(Subject to modif

13、ications)Around 23 Homework assignments(20%)2 Experiments(40%)Project and Examination(choose one of the following)(40%)Final exam Or Project(no predefined topic)Course activities (bonus)Topic 1.Introduction11Introduction to Machine Learning:IntroductionApplication background of machine learningIntro

14、duction to Machine Learning:Introduction12App.background for machine learningIntroduction to Machine Learning:Introduction13Data miningUsing historical data to improve decisionsApp.background?Data Mining?1?Introduction to Machine Learning:Introduction14Business IntelligenceExamples:location(P&G prod

15、ucts,Walmart,)Coupons and repeat consumptionsApp.Background:Data Mining?2?Introduction to Machine Learning:Introduction15Credit risk analysisCustomer 103?time=t0?Years of credit:9Loan balance:$2,400Income:$52KOwn House:YesDelinquent accts:2Max billing cycles late:3Profitable customer?:?Customer 103?

16、time=t1?Years of credit:9Loan balance:$3,250Income:?Own House:YesDelinquent accts:2Max billing cycles late:4Profitable customer?:?Customer 103?time=tn?Years of credit:9Loan balance:$4,500Income:?Own House:YesDelinquent accts:3Max billing cycles late:6Profitable customer?:NO Telephone confirmation of

17、 CreditTransactionsApp.background for machine learningIntroduction to Machine Learning:Introduction16Data miningUsing historical data to improve decisionsSelf customizing programsEmail filtering(and organization)Applications that learns user interests(e.g.news feeding,forum,social community)Recommen

18、dation in e-commerceApp.background for machine learningIntroduction to Machine Learning:Introduction17Data miningUsing historical data to improve decisionsSelf customizing programsEmail filtering(and organization)Applications that learns user interests(e.g.news feeding,forum,social community)Recomme

19、ndation in e-commerceApp.Background:Learning users interestsIntroduction to Machine Learning:Introduction18MilitaryMovieMovieClubApp.background for machine learningIntroduction to Machine Learning:Introduction19Data miningUsing historical data to improve decisionsSelf customizing programsEmail filte

20、ring(and organization)Applications that learns user interests(e.g.news readers,forum,social community)Recommendation in e-commerceApp.Background:Recommendation in Social NetworksIntroduction to Machine Learning:Introduction20App Background:e-commerce recommendationIntroduction to Machine Learning:In

21、troduction21Introduction to Machine Learning:Introduction22Data miningUsing historical data to improve decisionsSelf customizing programsEmail filtering(and organization)Applications that learns user interests(e.g.news readers,forum,social community)Recommendation in e-commerceSoftware applications

22、we cant/dont want to program by handFace/Speech/Handwritten recognition(Pattern Recognition)Autonomous drivingRanking in information retrieval(Search Engine)App.background for machine learningFace RecognitionSpeech recognitionHandwritten recognitionIntroduction to Machine Learning:Introduction24http

23、:/ to Machine Learning:Introduction25Data miningUsing historical data to improve decisionsSelf customizing programsEmail filtering(and organization)Applications that learns user interests(e.g.news readers,forum,social community)Recommendation in e-commerceSoftware applications we cant/dont want to p

24、rogram by handFace/Speech/Handwritten recognition(Pattern Recognition)Autonomous drivingRanking in information retrieval(Search Engine)App.background for machine learningExample:Autonomous driving(CMU)26ALVINN(19891996,CMU)70mph,highwayIntroduction to Machine Learning:IntroductionExample:Autonomous

25、driving(THU)Introduction to Machine Learning:Introduction27Example:Autonomous driving:Google Driverless CarIntroduction to Machine Learning:Introduction282009 Googles self-driving cars projectDec.2016Google transitioned the project into a new company called WaymoPlans to make self-driving cars avail

26、able to the public in 2020Introduction to Machine Learning:Introduction29Data miningUsing historical data to improve decisionsSelf customizing programsNewsreader that learns user interestsEmail filteringRecommendation in e-commerceSoftware applications we cant/dont want to program by handSpeech reco

27、gnition/OCR(Pattern Recognition)Autonomous drivingRanking in information retrieval(Search Engine)App.background for machine learningExample:information retrievalIntroduction to Machine Learning:Introduction30Ranking:1000+parametersLearning to rank(machine learning+information retrieval)Example:infor

28、mation retrieval(cont.)Relevance feedbacke.g.Image retrieval,video retrieval,Query:a horse picRetrieval results before feedback x xx xx xx xx xx xx xx xx xx xx xx xx xRetrieval results after 1 relevance feedbackx xx xx x31Introduction to Machine Learning:IntroductionIntroduction to Machine Learning:

29、Introduction32Data miningUsing historical data to improve decisionsSelf customizing programsNewsreader that learns user interestsEmail filteringRecommendation in e-commerceSoftware applications we cant program by handSpeech recognition/OCR(Pattern Recognition)Autonomous drivingRanking in information

30、 retrieval(Search Engine)App.background for machine learningNow you should have had ideas on the differences between machine learning and:-Data mining/Pattern recognition/-Neural networks/Deep Learning/-Whats machine learning?Introduction to Machine Learning:Introduction33Whats machine learning?1?In

31、troduction to Machine Learning:Introduction34Learning denotes changes in a system that.enable a system to do the same task moreefficiently the next time.”-Herbert SimonHerbert Simon(1916 2001)Won theA.M.Turing Award for his work on computer science in 1975 (One of the founders of AI)Won the Nobel Pr

32、ize in economics in 1978 Won the National Medal of Science in 1986In 1993,he was awarded the American Psychological Associations award for outstanding lifetime contributions to psychology.In 1994,he became one of foreign scientists inducted into the Chinese Academy of Sciences Whats machine learning

33、?2?Introduction to Machine Learning:Introduction35Learning is constructing or modifyingrepresentations of what is being experienced.-Ryszard S.MichalskiRyszard S.Michalski(1937-2007)http:/www.mli.gmu.edu/michalski/Michalski,Ryszard S.and Kodratoff,Y.Machine Learning,anAI approach 1990Co-founder:Mach

34、ine learning research fieldCo-founder:Machine Learning(Journal)Co-founder:ICMLFounder:conceptual clustering,constructive induction,variable-valued logic,variable-precision logicWhats machine learning?3?Introduction to Machine Learning:Introduction36Learning =improving with experience at some taskTom

35、 M.Mitchell(CMU)1973 MIT S.B.;1979 Ph.D.Stanford Uni.Co-founder:Machine Learning(Journal)Co-founder:ICMLIJCAI Computers andThoughtAward,1983T?Task?E?Experience?P?Performance?Learning:change/construct or modify/improveWhats machine learning?examples?Introduction to Machine Learning:Introduction37Lear

36、n to play checkersT:play checkersP?%of games wonE?e.g.play against selflHandwriting recognitionlT:recognition characterslP:classification precisionlE:character sets with known classA general machine learningsystem designIntroduction to Machine Learning:Introduction38Designing a learning systeme.g.Le

37、arning to play checkersIntroduction to Machine Learning:Introduction39What experience?Is training experience representative of performance goal?E.g.Checker GameWill the data(games)seen in training represents those seen in practice?play against self vs.play against a masterUnexpected resultsExample 1

38、:moon or banana?Example 2:digits recognition with NNExample 3:IBM Watsons answer on the question of“an American city”at Jeopardy Toronto?Whats wrong?Due to training data bias:data,training procedure,features System design Training ExperienceIntroduction to Machine Learning:Introduction40Designing a

39、learning systeme.g.Learning to play checkersIntroduction to Machine Learning:Introduction41What experience?Be aware of training data bias:data,training procedure,features What exactly should be learned?What exactly should be learned?Going to learn a Target Function(target concept)But of what type?i.

40、e.what is the concept class?E.g.Checker Game:Possible definition for Target Function V(b)If b is a final board state that is won,then V(b)=100If b is a final board state that is lost,then V(b)=-100If b is a final board state that is drawn,then V(b)=0If b is not a final board state,then V(b)=V(b)Wher

41、e b is the best final board state that can be achieved starting from b and playing optimally until the end of the game.Gives correct results,but is not operational Not efficiently computableTypically limit ourselves to an approximation of V,a.k.a.Hypothesise.g.VIntroduction to Machine Learning:Intro

42、duction42)()(bSuccessorVbVDesigning a learning systeme.g.Learning to play checkersIntroduction to Machine Learning:Introduction43What experience?Be aware of training data bias:data,training procedure,features What exactly should be learned?Correct vs.operational:Approximation,aka.HypothesisHow shall

43、 it be represented?Representing(the Hypothesis Class )E.g.Checker Game:Possible representationTables of all statesCollection of rulesPolynomial function of board featuresNeural networkExpressiveness of type of function must be chosen carefullyGoodness of approximation vs.data requirementsIntroductio

44、n to Machine Learning:Introduction44VRepresenting the Hypothesis Class(Cont.)E.g.Checker Game Learning examples where Vtrain(b)is the label of bwp(b):#of white pieces on board brp(b):#of red pieces on bwk(b):#of white kings on brk(b):#of red kings on bwt(b):#of white pieces threatened by red(i.e.whi

45、ch can be taken on reds next turn)rt(b):#of red pieces threatened by white)()()()()()()(6543210brtwbwtwbrkwbwkwbrpwbwpwwbV+=Introduction to Machine Learning:Introduction45le.g.,+100An example functionDesigning a learning systeme.g.Learning to play checkersIntroduction to Machine Learning:Introductio

46、n46What experience?Be aware of training data bias:data,training procedure,features What exactly should be learned?Correct vs.operational:Approximation,aka.HypothesisHow shall it be represented?Expressiveness:Goodness of approximation vs.data requirementsWhat specific algorithm to learn it?Choose a w

47、eight training rule(Learning Algorithm)An examplePopular alg.:Least Mean Squares,LMSInitialize weightsRepeat:1.Select a training example b at random2.Compute error(b)=Vtrain(b)(b)3.For each board feature fi,fibelong to wp,rp,rt,update weight wiwi wi+c fierror(b)c is some small constant,say 0.1 for e

48、xample,to moderate the rate of learning-settrainingtrainbVbV2)()(VIntroduction to Machine Learning:Introduction47lBest fit of datalCommon goal:minimize squared errorDesigning a learning systeme.g.Learning to play checkersIntroduction to Machine Learning:Introduction48What experience?What exactly sho

49、uld be learned?How shall it be represented?What specific algorithm to learn it?Putting it Together final designPutting it togetherInitialize weights of Use to play games against self,output sequence of board states for each gameLabel each b withstatevalueLearn new weight,yielding newStart new set of

50、 gamesVVIntroduction to Machine Learning:Introduction49)(bSuccessorVV)(2bV)(3bV)(mbV100/100-+b1b2bm-1bmOverview Design choicesIntroduction to Machine Learning:Introduction50Determine type of training experienceGame against expertsAgainst selfTable of correct movesWhat experienceOverview Design choic