资源描述:
《Fundamentals of Machine Learning》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、Chapter2FundamentalsofMachineLearning2.1LearningMethodsLearningisafundamentalcapabilityofneuralnetworks.Learningrulesarealgo-rithmsforfindingsuitableweightsWand/orothernetworkparameters.Learningofaneuralnetworkcanbeviewedasanonlinearoptimizationproblemf
2、orfindingasetofnetworkparametersthatminimizethecostfunctionforgivenexamples.Thiskindofparameterestimationisalsocalledalearningortrainingalgorithm.Neuralnetworksareusuallytrainedbyepoch.Anepochisacompleterunwhenallthetrainingexamplesarepresentedtothenetw
3、orkandareprocessedusingthelearningalgorithmonlyonce.Afterlearning,aneuralnetworkrepresentsacom-plexrelationship,andpossessestheabilityforgeneralization.Tocontrolalearningprocess,acriterionisdefinedtodecidethetimeforterminatingtheprocess.Thecomplexityofa
4、nalgorithmisusuallydenotedasO(m),indicatingthattheorderofnumberoffloating-pointoperationsism.Learningmethodsareconventionallydividedintosupervised,unsupervised,andreinforcementlearning;theseschemesareillustratedinFig.2.1.xpandyparetheinputandoutputofthe
5、pthpatterninthetrainingset,ˆypistheneuralnetworkoutputforthepthinput,andEisanerrorfunction.Fromastatisticalviewpoint,unsuper-visedlearninglearnsthepdfofthetrainingset,p(x),whilesupervisedlearninglearnsaboutthepdfofp(y
6、x).Supervisedlearningiswidelyusedi
7、nclassification,approx-imation,control,modelingandidentification,signalprocessing,andoptimization.Unsupervisedlearningschemesaremainlyusedforclustering,vectorquantization,featureextraction,signalcoding,anddataanalysis.Reinforcementlearningisusuallyusedin
8、controlandartificialintelligence.Inlogicandstatisticalinference,transductionisreasoningfromobserved,spe-cific(training)casestospecific(test)cases.Incontrast,inductionisreasoningfromobservedtrainingcasestogeneralrules,whicharethenappliedtothetestcases.Mach
9、inelearningfallsintotwobroadclasses:inductivelearningortransductivelearning.Inductivelearningpursuesthestandardgoalinmachinelearning,whichistoaccuratelyclassifytheentireinputspace.Incontrast,transductivelearningfocusesK.-L.DuandM.N.S.Sw