資源描述：
《a learning algorithm for boltzmann machines 1985 3542外語英文電子書》由會(huì)員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在教育資源-天天文庫。
1、COGNITIVESCIENCE9,147-169(1985)ALearningAlgorithmforBoltzmannMachines*DAVIDH.ACKLEYGEOFFREYE.HINTONComputerScienceDepartmentCarnegie-MellonUniversityTERRENCEJ.SEJNOWSKIBiophysicsDepartmentTheJohnsHopkinsUniversityThecomputotionolpowerofmassivelyparallelnetworksofsi
2、mpleprocessingelementsresidesinthecommunicationbandwidthprovidedbythehardwareconnectionsbetweenelements.Theseconnectionsconallowasignificantfractionoftheknowledgeofthesystemtobeappliedtoaninstanceofaprob-leminoveryshorttime.Onekindofcomputationforwhichmassivelyporo
3、llelnetworksappeartobewellsuitedislargeconstraintsatisfactionsearches,buttousetheconnectionsefficientlytwoconditionsmustbemet:First,asearchtechniquethatissuitableforparallelnetworksmustbefound.Second,theremustbesomewayofchoosinginternalrepresentationswhichallowthep
4、reexistinghardwareconnectionstobeusedefficientlyforencodingthecon-straintsinthedomainbeingsearched.Wedescribeagenerolparallelsearchmethod,basedonstatisticalmechanics,andweshowhowitleadstoagen-erallearningruleformodifyingtheconnectionstrengthssoastoincorporateknowle
5、dgeoboutotaskdomaininonefficientway.Wedescribesomesimpleexamplesinwhichthelearningalgorithmcreatesinternalrepresentationsthotoredemonstroblythemostefficientwayofusingthepreexistingconnec-tivitystructure.1.INTRODUCTIONEvidenceaboutthearchitectureofthebrainandthepote
6、ntialofthenewVLSItechnologyhaveledtoaresurgenceofinterestin“connectionist”sys-lTheresearchreportedherewassupportedbygrantsfromtheSystemDevelopmentFoundation.WethankPeterBrown,FrancisCrick,MarkDerthick,ScottFahlman,JerryFeldman,StuartGeman,GailGong,JohnHopfield,JayM
7、cClelland,BarakPearlmutter,HarryPrintz,DaveRumelhart,TimShallice,PaulSmolensky,RickSzeliski,andVenkatara-manVenkatasubramanianforhelpfuldiscussions.ReprintrequestsshouldbeaddressedtoDavidAckley,ComputerScienceDepartment,Carnegie-MellonUniversity,Pittsburgh,PA15213.
8、147148ACKLEY.HINTON.ANDSEJNOWSKIterns(Feldman&Ballard,1982;Hinton&Anderson,1981)thatstoretheirlong-termknowledgeasthestrengthsoftheconnectionsbet
