機器學(xué)習(xí)介紹(英文版：備注里有中文翻譯)課件

MachineLearning09八月2023MachineLearning08八月20231Machinelearning,asabranchofartificialintelligence,isgeneraltermsofakindofanalyticalmethod.Itmainlyutilizescomputersimulateorrealizethelearnedbehaviorofhuman.09八月2023Machinelearning,asabranch209八月20231）Machinelearningjustlikeatruechampionwhichgohaughtily;

2）Patternrecognitioninprocessofdeclineanddieout;

3）Deeplearningisabrand-newandrapidlyrisingfield.theGooglesearchindexofthreeconceptsince200408八月20231）Machinelearningj309八月2023Theconstructedmachinelearningsystembasedoncomputermainlycontainstwocoreparts：representationandgeneralization.Thefirststepfordatalearningistorepresentthedata,i.e.detectthepatternofdata.Establishageneralizedmodelofdataspaceaccordingtoagroupofknowndatatopredictthenewdata.Thecoretargetofmachinelearningistogeneralizefromknownexperience.Generalizationmeansapowerofwhichthemachinelearningsystemtobelearnedforknowndatathatcouldpredictthenewdata.08八月2023Theconstructedmach4SupervisedlearningInputdatahaslabels.Thecommonkindoflearningalgorithmisclassification.Themodelhasbeentrainedviathecorrespondencebetweenfeatureandlabelofinputdata.Therefore,whensomeunknowndatawhichhasfeaturesbutnolabelinput,wecanpredictthelabelofunknowndataaccordingtotheexistingmodel.09八月2023Supervisedlearning08八月20235UnsupervisedlearningInputdatahasnolabels.Itrelatestoanotherlearningalgorithm,i.e.clustering.Thebasicdefinitionisacoursethatdividethegatherofphysicalorabstractobjectintomultipleclasswhichconsistofsimilarobjects.09八月2023Unsupervisedlearning08八月2026Iftheoutputeigenvectormarkscomefromalimitedsetthatconsistofclassornamevariable,thenthekindofmachinelearningbelongstoclassificationproblem.

Ifoutputmarkisacontinuousvariable,thenthekindofmachinelearningbelongstoregressionproblem.09八月2023Iftheoutputeigenvectormark7ClassificationstepFeatureextractionFeatureselectionModeltrainingClassificationandpredictionRawdataNewdata09八月2023ClassificationstepFeatureext8Featureselection（featurereduction）CurseofDimensionality：Usuallyrefertotheproblemthatconcernedaboutcomputationofvector.Withtheincreaseofdimension,calculatedamountwilljumpexponentially.Corticalfeaturesofdifferentbrainregionsexhibitvarianteffectduringtheclassificationprocessandmayexistsomeredundantfeature.Inparticularafterthemultimodalfusion,theincreaseoffeaturedimensionwillcause“curseofDimensionality”.09八月2023Featureselection（featurered9PrincipalComponentAnalysis,PCAPCAisthemostcommonlineardimensionreductionmethod.Itstargetismappingthedataofhighdimensiontolow-dimensionspaceviacertainlinearprojection,andexpectthevarianceofdatathatprojectthecorrespondingdimensionismaximum.Itcanusefewerdatadimensionmeanwhileretainthemajorcharacteristicofrawdata.09八月2023PrincipalComponentAnalysis,10Lineardiscriminantanalysis,LDAThebasicideaofLDAisprojection,mappingtheNdimensiondatatolow-dimensionspaceandseparatethebetween-groupsassoonaspossible.i.e.theoptimalseparabilityinthespace.Thebenchmarkisthenewsubspacehasmaximumbetweenclassdistanceandminimalinter-objectdistance.09八月2023Lineardiscriminantanalysis,11Independentcomponentanalysis,ICAThebasicideaofICAistoextracttheindependencesignalfromagroupofmixedobservedsignaloruseindependencesignaltorepresentothersignal.09八月2023Independentcomponentanalysis12Recursivefeatureeliminationalgorithm,RFERFEisagreedyalgorithmthatwipeoffinsignificancefeaturestepbysteptoselectthefeature.Firstly,cyclicorderingthefeatureaccordingtotheweightofsub-featureinclassificationandremovethefeaturewhichrankatterminalonebyone.Then,accordingtothefinalfeatureorderinglist,selectdifferentdimensionofseveralfeaturesubsetfronttoback.Assesstheclassificationeffectofdifferentfeaturesubsetandthengettheoptimalfeaturesubset.

09八月2023Recursivefeatureelimination13Classificationalgorithm

DecisiontreeDecisiontreeisatreestructure.Eachnonleafnodeexpressesthetestofafeaturepropertyandeachbranchexpressestheoutputoffeaturepropertyincertainrangeandeachleafnodestoresaclass.Thedecision-makingcourseofdecisiontreeisstartingfromrootnode,testingthecorrespondingfeaturepropertyofwaitingobjects,selectingtheoutputbranchaccordingtotheirvalues,untilreachingtheleafnodeandtaketheclassthatleafnodestoreasthedecisionresult.09八月2023ClassificationalgorithmDecis14NaiveBayes,NBNBclassificationalgorithmisaclassificationmethodinstatistics.Ituseprobabilitystatisticsknowledgeforclassification.Thisalgorithmcouldapplytolargedatabaseandithashighclassificationaccuracyandhighspeed.09八月2023NaiveBayes,NB08八月202315Artificialneuralnetwork,ANNANNisamathematicalmodelthatapplyakindofstructurewhichsimilarwithsynapseconnectionforinformationprocessing.Inthismodel,amassofnodeformanetwork,i.e.neuralnetwork,toreachthegoalofinformationprocessing.Neuralnetworkusuallyneedtotrain.Thecourseoftrainingisnetworklearning.Thetrainingchangethelinkweightofnetworknodeandmakeitpossessthefunctionofclassification.Thenetworkaftertrainingapplytorecognizeobject.09八月2023Artificialneuralnetwork,ANN16k-NearestNeighbors,kNNkNNalgorithmisakindofclassificationmethodbaseonlivingexample.Thismethodistofindthenearestktrainingsampleswithunknownsamplexandexaminethemostofksamplesbelongtowhichclass,thenxbelongstothatclass.kNNisalazylearningmethod.Itstoressamplesbutproceedclassificationuntilneedtoclassify.Ifsamplesetarerelativelycomplex,itmaybeleadtolargecomputationoverhead.Soitcannotapplytostronglyreal-timeoccasion.09八月2023k-NearestNeighbors,kNN08八月17supportvectormachine,SVMMappingthelinearlyinseparabledatainlow-dimensionspacetohigh-dimensionspaceandmakeitlinearlyseparable09八月2023supportvectormachine,SVM0818Crossvalidation,CVThebasicideaofCVisgroupingtherawdatainasense.Onepartistakenastrainset,theotherpartistakenasvalidationset.Primarily,theclassifieristrainedwithtrainset,andthenusevalidationsettotestthereceivedmodelbytraining.09八月2023Crossvalidation,CVThebasic19K-foldcross-validationIn

k-foldcross-validation,theoriginalsampleisrandomlypartitionedinto

k

equalsizedsubsamples.Ofthe

k

subsamples,asinglesubsampleisretainedasthevalidationdatafortestingthemodel,andtheremaining

k

?

1subsamplesareusedastrainingdata.Thecross-validationprocessisthenrepeated

k

times(the

folds),witheachofthe

k

subsamplesusedexactlyonceasthevalidationdata.The

k

resultsfromthefoldscanthenbeaveragedtoproduceasingleestimation.Theadvantageofthismethodoverrepeatedrandomsub-samplingisthatallobservationsareusedforbothtrainingandvalidation,andeachobservationisusedforvalidationexactlyonce.10-foldcross-validationiscommonlyused.09八月2023K-foldcross-validation08八月220Leave-one-outcross-validation,LOOCVWhen

k

=

n

(thenumberofobservations),the

k-foldcross-validationisexactlytheleave-one-outcross-validation.09八月2023Leave-one-outcross-validation21confusionmatrixTP——goldstandardandtestaffirmsufferfromcertainillness;TN——goldstandardandtestaffirmnotsufferfromcertainillness;FP——go