论文.docx

论文.docx

《论文.docx》由会员分享，可在线阅读，更多相关《论文.docx（11页珍藏版）》请在冰豆网上搜索。

论文

1DefiningQuestions

Ascientificfieldisbestdefinedbythecentralquestionitstudies.ThefieldofMachineLearningseekstoanswerthequestion

“Howcanwebuildcomputersystemsthatautomaticallyimprovewithexperience,andwhatarethefundamentallawsthatgovernalllearningprocesses?

”

Thisquestioncoversabroadrangeoflearningtasks,suchashowtodesignautonomousmobilerobotsthatlearntonavigatefromtheirownexperience,howtodataminehistoricalmedicalrecordstolearnwhichfuturepatientswillrespondbesttowhichtreatments,andhowtobuildsearchenginesthatautomaticallycustomizetotheiruser’sinterests.Tobemoreprecise,wesaythatamachinelearnswithrespecttoaparticulartaskT,performancemetricP,andtypeofexperienceE,ifthesystemreliablyimprovesitsperformancePattaskT,followingexperienceE.

DependingonhowwespecifyT,P,andE,thelearningtaskmightalsobecalledbynamessuchasdatamining,autonomousdiscovery,databaseupdating,programmingbyexample,etc.

MachineLearningisanaturaloutgrowthoftheintersectionofComputerScienceandStatistics.WemightsaythedefiningquestionofComputerScienceis“Howcanwebuildmachinesthatsolveproblems,andwhichproblemsareinherentlytractable/intractable?

”ThequestionthatlargelydefinesStatisticsis“Whatcanbeinferredfromdataplusasetofmodelingassumptions,withwhatreliability?

”ThedefiningquestionforMachineLearningbuildsonboth,butitisadistinctquestion.WhereasComputerSciencehasfocusedprimarilyonhowtomanuallyprogramcomputers,MachineLearningfocusesonthequestionof

howtogetcomputerstoprogramthemselves（fromexperienceplussomeinitialstructure）.WhereasStatisticshasfocusedprimarilyonwhatconclusionscanbeinferredfromdata,MachineLearningincorporatesadditionalquestionsaboutwhatcomputationalarchitecturesandalgorithmscanbeusedtomosteffectivelycapture,store,index,retrieveandmergethesedata,howmultiplelearningsubtaskscanbeorchestratedinalargersystem,andquestionsofcomputationaltractability.

AthirdfieldwhosedefiningquestioniscloselyrelatedtoMachineLearningisthestudyofhumanandanimallearninginPsychology,Neuroscience,andrelatedfields.Thequestionsofhowcomputerscanlearnandhowanimalslearnmostprobablyhavehighlyintertwinedanswers.Todate,however,theinsightsMachineLearninghasgainedfromstudiesofHumanLearningaremuchweakerthanthoseithasgainedfromStatisticsandComputerScience,dueprimarilytotheweakstateofourunderstandingofHumanLearning.Nevertheless,thesynergybetweenstudiesofmachineandhumanlearningisgrowing,withmachinelearningalgorithmssuchastemporaldifferencelearningnowbeingsuggestedasexplanationsforneuralsignalsobservedinlearninganimals.OverthecomingyearsitisreasonabletoexpectthesynergybetweenstudiesofHumanLearningandMachineLearningtogrowsubstantially,astheyarecloseneighborsinthelandscapeofcorescientificquestions.

Otherfields,frombiologytoecomonicstocontroltheoryalsohaveacoreinterestinthequestionofhowsystemscanautomaticallyadaptoroptimizetotheirenvironment,andmachinelearningwilllikelyhaveanincreasingexchangeofideaswiththesefieldsoverthecomingyears.Forexample,economicsisinterestedinquestionssuchashowdistributedcollectionsofself-interestedindividualsmayformasystem（market）thatlearnspricesleadingtopareto-optimalallocationsforthegreatestcommongood.Andcontroltheory,especiallyadaptivecontroltheory,isinterestedinquestionssuchashowaservo-controlsystemcanimproveitscontrolstrategythroughexperience.Interestingly,themathematicalmodelsforadaptationintheseotherfieldsaresomewhatdifferentfromthosecommonlyusedinmachinelearning,suggestingsignificantpotentialforcross-fertilizationofmodelsandtheories.

ThefollowingsectionsdiscussthestateoftheartofMachineLearning,asampleofsuccessfulapplica-tions,andasampleofopenresearchquestions.

2StateofMachineLearning

Herewedescribesomeoftheprogressinmachinelearning,aswellasopenresearchquestions.

2.1ApplicationSuccesses

OnemeasureofprogressinMachineLearningisitssignificantreal-worldapplications,suchasthoselistedbelow.Althoughwenowtakemanyoftheseapplicationsforgranted,itisworthnotingthataslateas1985therewerealmostnocommercialapplicationsofmachinelearning.

*Speechrecognition.Currentlyavailablecommercialsystemsforspeechrecognitionallusemachinelearninginonefashionoranothertotrainthesystemtorecognizespeech.Thereasonissimple:

thespeechrecognitionaccuracyisgreaterifonetrainsthesystem,thanifoneattemptstoprogramitbyhand.Infact,manycommercialspee