机器学习与模式识别的关系英文版

Machine Learningand Pattern RecognitionAbstractRecently,machine learninghas developedrapidly ininformation field.Also jthas aclose relationshipwith pattern recognition.Machining learninghas been appliedto pattern recognition successfully.Therefore,the paperdescribesthe basiccharacteristics of machine learning and patternrecognition,which includesthe concepts,development,application andclassification.It alsoprovides anapplication perspectivefor understandingtheconcepts ofmachining andpattern recognition.Keywords:Machine LearningPatternRecognition

0.IntroductionMachine learningis oneof thecore problemsof artificialintelligenceresearch.Its applicationhasbeenthroughout allbranches of artificialintelligence,such asexpert systems,automated reasoningin the field ofnaturallanguage understanding,pattern recognition,computer vision,intelligent robotics.Just asits nameimplies,Machine learningis tolet thecomputerto learnsome wayto improve its performance.Pattern recognitioncan be seen as somethingwhich candivide differentobjects intodifferentcategories.Humans candeepen theirunderstanding ofthings throughcontinuouslearning,similarly thepattern recognitionsystem based onsimulating humanintelligence alsoneeds toimproveitsclassificationperformance throughmachine learningalgorithm improvements,so thecontactbetween machine learningandpattern recognitionis gettingcloserand closer.This articlewill explain the basicconcepts ofmachine learningandpatternrecognition,patternrecognitionanalysis inseveral machinelearningalgorithms.

1.Machine Learning

1.1The definition ofmachine learningCurrently,the accuratedefinitionofmachine learning:fbr certainassignmentT andperformance metricsP,if acomputer programto measuretheperformance ofP andalong withthe experienceof self-improvement on[then wecall thecomputer programis learningfromexperience E.

1.2The workingmechanism of the machine learning systemTheenvironment providescertain information to the learning partsofthe system,then thelearning partuses thisinformationtomodify itsknowledge base toenhance theperformance ofexecution part;The executiondoits workaccording theknowledgebase,also bringback theacquiredinformation tolearning part.The processcan beseenasa certainprocess thatthemachine learningsystem acquireknowledge automaticallywithinformation whichare providedby internaland externalenvironment.

1.3The design of themachine learningsystemThere aremainly twoparts thatneed betaken intoconsideration whendesigninga perfectmachine learningsystem:Model selection and design,Learning algorithmselectionanddesign.Different modelsdeterminedifferent objectivefunctions anddifferent learningmechanisms.Thecomplexity andcapacity ofalgorithm determinethe capacityandefficiency of thelearningsystem.Also thesize of training samplesand featureselectionproblem arethe keyfactors whichwill constrainmachine learningsystemperformance.

2.Machine learningalgorithm inpatternrecognitionPattern recognitionmeans thatwe shouldanalyze perceptionsignal.It is aprocess ofidentification and interpretation.Wedescribe thisprocess.can drawa picturetoI获取数据T预处理卜［特征生成卜特征选择T模式分类卜变巫机器学习The coreissue ofmachine learningis searchingproblems.As fbrdifferentapplication models,the researchershave designedsome differentsearchingalgorithms.Currently inthefield of patternrecognition,we oftenusegenetic algorithmsneural networks,support vectormachines,k-nearestneighbor methodand othermachine learningalgorithms.

2.1Genetic algorithmCharacteristicdimension is a majorproblem inmachinelearning,because thecharacteristics presentedfrom certainmodel havedifferentweights inreflecting thenature ofthings.But someshowed nosignificantcontribution tothe catagories,even redundant,so the feature selectionprocessis verycritical.Genetic algorithmcan solvethis problemto someextend asaoptimization algorithm.Genetic algorithmnot onlycan choosethefeaturethatnot onlyreflects theoriginal information,but alsohave asignificant impactonthe classificationresults.There arethree kindsof operationin GA.Selection-reproduction,crossover,as wellas mutation.We usuallydo asfollows:Choose Nchromosomes frompopulation SinN separatetimes.The probabilityof oneindividual beingchosen isPxi.Thecomputational formulaof Pxi:There isa chance that thechromosomes ofthe twoparents arecopiedunmodified asoffspring,or randomlyrecombined crossoverto formoffspring.Also thereisachancethata geneofachild ischanged randomly.Generally thechance ofmutation islow.GA havefour basicelements fromthe present:coding strategies;settinginitial population;designoffitness function;genetic operatorsdesign,chooseoperator,crossover operator,mutation operator,and thesehave beenaimportant pointsin improving.

2.2Artificial neural networksNeural networkisanew technologyinthefieldofmachinelearning.Many peoplehave heardoftheword,but fewpeople reallyunderstandwhat itis.The basicneural networkfunctions,including itsgeneralstructure,related terms,types andapplications.In patternrecognition applications,a classifierusing aneural networkisdesigned by a relativelysmall numberof neuronsconnected togetheraccordingto certainrules ofnetwork system,and eachneuron inthe networkhavethe samestructure.Neurons typically expressed asa multiple-input,single-output nonlinearelements,its structurecanbedesigned likethis:As alink learningalgorithm,neural networkfeatures are:parallelprocessing ofinformation,storage anddistribution ofstrong faulttolerance;self^leaming,self^organization andself^applicability.Through training,theneural networkcan automaticallyadjust itsnetwork configurationparametersto simulatethe nonlinearrelationship betweeninput andoutput,so whenwegive thenetwork someinputs,we canget theright classification.23Support vectormachinesThe sizeoftrainingsamples inmachinelearningsystem influencetheability ofgeneralization learningsystem.In machinelearning,supportvector machinesSVMs,also support vector networksare supervisedlearningmodels withassociated learningalgorithms thatanalyze dataandrecognize patterns,used forclassification andregression analysis.Given aset oftraining examples,each markedas belongingto oneof twocategories,an SVMtraining algorithmbuilds amodel thatassigns newexamplesinto onecategory orthe other,making ita non-probabilisticbinary linearclassifier.An SVMmodel isa representationof theexamplesas pointsin space,mapped sothat theexamples ofthe separatecategoriesare dividedbyaclear gapthat isas wideas possible.Newexamples arethen mappedinto thatsame spaceand predictedto belongtoa categorybasedonwhich sideofthegap theyfall on.In additiontoperforming linear classification,SVMs canefficiently performanon-linearclassificationusing whatis calledthe kerneltrick,implicitlymapping theirinputs intohigh-dimensional featurespaces.Supposing thatwe havea sampleset xi,yi,i=l,23・・・.n,ye{+1,—1},andinthe preconditiony[*X,.+5—1]20FTy叫/Satisfies thecondition andfind acategory surfacethat canminimize

2.

3.ConclusionMachine learningin away canbe understoodas anassuming spacewhichis definedby anymodels.Its coretechnology islearning howtouse thealgorithm tosearch inthe correspondinghypothesis space,inother words,search processis thelearning process.In patternrecognitionapplications.The classifierconstructed byneuralnetwork,supportvectormachine andk-nearest neighboris findinga categorysurfacein supposingspace thatcan makethe differenttypes ofsamplesin disjointregions aspossible aswe can.。