资源描述：
《+ nonlinear dynamic recurrent associative memory for learning bipolar and nonbipolar correlated patterns》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、IEEETRANSACTIONSONNEURALNETWORKS,VOL.16,NO.6,NOVEMBER20051393NDRAM:NonlinearDynamicRecurrentAssociativeMemoryforLearningBipolarandNonbipolarCorrelatedPatternsSylvainChartier,Member,IEEEandRobertProulx,SeniorMember,IEEEAbstract—Thispaperpresentsanewunsupervisedattractora
2、tscalingparameter[10],[11](see[12]foracomparativeanal-neuralnetwork,which,contrarytooptimallinearassociativeysisofanoptimalprojectionmodelinaHopfieldtypenetwork).memorymodels,isabletodevelopnonbipolarattractorsaswellAlthoughthosevariousmodelshaveabetterperformanceasbipol
3、arattractors.Moreover,themodelisabletodeveloplessthanthesimpleHebbianalgorithm,theyneverthelesshavethespuriousattractorsandhasabetterrecallperformanceunderrandomnoisethananyotherHopfieldtypeneuralnetwork.Thoseproblemofspuriousattractorsandlackthecapacitytodevelopperforma
4、ncesareobtainedbyasimpleHebbian/anti-Hebbiannonbipolarattractors.Inallpreviousmodels,theoutputisboundonlinelearningrulethatdirectlyincorporatesfeedbackfromainahypercubethatlimitstheunit’svaluesto1or1.More-specificnonlineartransmissionrule.Severalcomputersimulationsover,V
5、idyasagar[13]demonstratesthatHopfieldtypenetworksshowthemodel’sdistinguishingproperties.usingastepfunctioncanonlydevelopstableattractorsathyper-IndexTerms—Associativememory,dynamicmodel,neuralnet-cubecorners.Consequently,thosemodelstypicallydevelopex-work,unsupervisedlea
6、rning.tremebehaviorthatrestrictsitscognitiveexplanation.AmorepowerfulmodelwouldbeabletodevelopattractorsanywhereI.INTRODUCTIONwithinahypercubequadrantinsteadofonlyatitsextremities.Toaccomplishthis,researchershaveusedadifferenttypeofTTRACTORneuralnetworks(e.g.,[1]and[2])
7、defineatransmissionruleusingmultiplelimitoutputfunction[14],[15].Aclassofformalmodelsthatareusuallyusedasautoasso-Althoughthissolutionyieldsgoodresults,itdoessowithanin-ciativememory.Thekeymechanismcommontoallattractorcreaseinthelearningrulecomplexity.Moreover,thepropose
8、dneuralnetworksisthepresenceofafeedbackloop.Feedbackmodelismoresensitivetonoisethanitsbinarycounterpart.enable
