您好,欢迎访问三七文档
当前位置:首页 > IT计算机/网络 > AI人工智能 > 机器学习当前研究进展
1、AFewUsefulThingstoKnowaboutMachineLearning11/10/2012Presentedby:YangSONG11/10/2012AFewUsefulThingstoKnowaboutMachineLearning2/71Author’sIntroductionPedroDomingosIntroductionTwelvekeylessonsConclusions11/10/2012AFewUsefulThingstoKnowaboutMachineLearning4/71OutlineIntroductionTwelvekeylessonsConclusions11/10/2012AFewUsefulThingstoKnowaboutMachineLearning5/71IntroductionMachineLearningAfewquotes“AbreakthroughinmachinelearningwouldbeworthtenMicrosofts”(BillGates,Chairman,Microsoft)Machi。
2、nelearningisthehotnewthing”(JohnHennessy,President,Stanford)“Machinelearningistoday’sdiscontinuity”(JerryYang,Founder,Yahoo)“Machinelearningtodayisoneofthehottestaspectsofcomputerscience”(SteveBallmer,CEO,Microsoft)11/10/2012AFewUsefulThingstoKnowaboutMachineLearning6/71IntroductionMachineLearningTraditionalProgrammingMachineLearning11/10/2012AFewUsefulThingstoKnowaboutMachineLearning7/71IntroductionTypesofLearningSupervised(inductive)learningTrainingdataincludesdesiredoutputsUnsupervis。
3、edlearningTrainingdatadoesnotincludedesiredoutputsSemi-supervisedlearningTrainingdataincludesafewdesiredoutputsReinforcementlearningRewardsfromsequenceofactions11/10/2012AFewUsefulThingstoKnowaboutMachineLearning8/71IntroductionArecentreportfromtheMcKinseyGlobalInstitute,2011“machinelearning(a.k.a.dataminingorpredictiveanalytics)willbethedriverofthenextbigwaveofinnovation.”11/10/2012AFewUsefulThingstoKnowaboutMachineLearning9/71IntroductionPurposeofthisarticle“Severalfinetextbooksareava。
4、ilabletointerestedpractitionersandresearchers.However,muchofthe‘folkknowledge’thatisneededtosuccessfullydevelopmachinelearningapplicationsisnotreadilyavailableinthem.”“Asaresult,manymachinelearningprojectstakemuchlongerthannecessaryorwindupproducinglessthanidealresults.Yetmuchofthisfolkknowledgeisfairlyeasytocommunicate.”“民科”Thisarticlewillfocusonclassificationproblem11/10/2012AFewUsefulThingstoKnowaboutMachineLearning10/71OutlineIntroductionTwelvekeylessonsConclusions11/10/2012AFewUsefulThi。
5、ngstoKnowaboutMachineLearning11/71OutlineIntroductionTwelvekeylessons1.Learning=representation+evaluation+optimization2.It’sgeneralizationthatcounts(泛化)3.Dataaloneisnotenough(先验知识)4.Overfittinghasmanyfaces(过拟合)5.Intuitionfailsinhighdimensions(高维)6.Theoreticalguaranteesarenotwhattheyseem(理论)Conclusions11/10/2012AFewUsefulThingstoKnowaboutMachineLearning12/71OutlineIntroductionTwelvekeylessons7.Featureengineeringisthekey(特征)8.Moredatabeatsaclevereralgorithm(数据)9.Learnmanymodels,notju。
6、stone(集成学习)10.Simplicitydoesnotimplyaccuracy(简单与精确)11.Representabledoesnotimplylearnable(可表示与可学习)12.Correlationdoesnotimplycausation(关联与因果)Conclusions11/10/2012AFewUsefulThingstoKnowaboutMachineLearning13/711.Learning=representation+evaluation+optimization11/10/2012AFewUsefulThingstoKnowaboutMachineLearning14/711.Learning=representation+evaluation+optimizationLearningalgorithmshavethreecomponents:Representation:Aclassifiermustberepresentedinsomeformallanguagethatthecomputercanhandle.Choos。
7、ingarepresentationistantamounttochoosingthesetofclassifiersitcanpossiblylearn.Thissetiscalledthehypothesisspaceofthelearner.Evaluation:Anevaluationfunction(alsocalledobjectivefunctionorscoringfunction)isneededtodistinguishdifferentclassifiers.Optimization:Weneedamethodtosearchamongtheclassifiersforthehighest-scoringone.Thechoiceofoptimizationtechniqueiskeytotheefficiencyofthelearner.HypothesisspaceScoringfunctionSearchmethods11/10/2012AFewUsefulThingstoKnowaboutMachineLearning15/711.Learning。
8、=representation+evaluation+optimizationRepresentationInstancesKNN,SVMHyperplanesNaïveBayes,LogisticregressionDecisiontreesSetsofrulesPropositionalrules,LogicprogramsNeuralnetworksGraphicalmodelsBayesiannetworks,Conditionalrandomfields11/10/2012AFewUsefulThingstoKnowaboutMachineLearning16/711.Learning=representation+evaluation+optimizationEvaluationAccuracy/ErrorratePrecisionandrecallSquarederrorLikelihoodPosteriorprobabilityInformationgainK-LdivergenceCost/UtilityMargin11/10。
9、/2012AFewUsefulThingstoKnowaboutMachineLearning17/711.Learning=representation+evaluation+optimizationOptimizationCombinatorialoptimizationGreedysearch,Beamsearch,Branch-and-boundContinuousoptimization(convexoptimization)UnconstrainedGradientdescent,Conjugategradient,Quasi-NewtonmethodsConstrainedLinearprogramming,Quadraticprogramming11/10/2012AFewUsefulThingstoKnowaboutMachineLearning18/711.Learning=representation+evaluation+optimizationNotallcombinationsofonecomponentfromeachcolumninth。
10、etablemakeequalsenseE.g.discreterepresentationsnaturallygowithcombinatorialoptimization,continuouswithcontinuousThereisnosimplerecipeforchoosingeachcomponent,thenextsectionswilltouchonsomeofthekeyissues.11/10/2012AFewUsefulThingstoKnowaboutMachineLearning19/712.It’sgeneralizationthatcounts11/10/2012AFewUsefulThingstoKnowaboutMachineLearning20/712.It’sgeneralizationthatcountsThefundamentalg。
本文标题:机器学习当前研究进展
链接地址:https://www.777doc.com/doc-1489385 .html