《modelling cell assemblies》由会员分享，可在线阅读，更多相关《modelling cell assemblies（20页珍藏版）》请在金锄头文库上搜索。
1、ModellingCellAssemblies ChristianR?Huyck MiddlesexUniversity c?huyck?mdx?ac?uk Abstract NeuralNetworksareverypopularcomputationalmodelsthatare generallysaidtobeinspiredbyhumanneuralfunctioning?However? neuronsinmostneuralnetsfunctionquitedi?erentlyfromnatural systems? Thispaperdescribesanovelarchite
2、cture?theCANT?Connections? AssociationsandNetworkTechnology?modelwhichisdesignedto functionlikenaturalneuralsystems?It?rstelucidatestheimportant aspectsofthemodelandhowtheyrelatetonaturalsystems?Thebasic ideaisderivedfromD?O?Hebb?sideaoftheCellAssemblywhichis theneuralequivalentofaconcept? Thepaperg
3、oesontodescribethreeinstancesofthemodel?In eachcase?stimulusispresentedtothenetworkbydirectlyactivating neuronsinthesystem?Inanygivenrunofthenetwork?thisshould leadtoactivationoftheappropriateCA?Whenlearningisinvolved? earlyrunsonthesystemwillnotleadtoactivationofaCA?Instead? theearlyactivationwilll
4、eadtothesystemlearningtheCA? CAsareinitiallydeterminedbyappropriatelevelsofactivationin thenetwork?Latertheyaredeterminedbystatisticalanalysisofthe activationpatterns?The?rstexperimentshowsthattheCANTmodel generatesCAsgivenawidevarietyofparametersettings?Learningis essentialinneuralmodels?theseconde
5、xperimentshowshowlearning candevelopaCAwheremereparametersettingcannot?Innatural neuralsystems?multipleCAsmustexistasnaturalsystemsmusthave multipleconcepts?Thethirdexperimentshowshowasimplenetcan containtwoCAs? Thepaperconcludeswithadiscussionoffuturedevelopmentsof themodel?Theseincludeduplicatingp
6、sychologicalandneurobiological data?ThiswillrequirefurtheranalysisofCAs?learningalgorithmsand developmentofhigherorderformalisms? ? ?IntroductionandBackground NeuralNetworksareverypopularcomputationalmodelsthataregenerally saidtobeinspiredbyhumanneuralfunctioning?However?neuronsorunits inmostneuraln
7、etsfunctionquitedi?erentlyfromhumanneurons? Anaveragehumanneuron?resforabrieftime?andthenfatigues?Ifsuch aneuronweretorepresentaconcept?wecouldonlythinkofthatconcept forabrieftime?Sincewecanthinkofagivenconceptforatleastseveral seconds?asingleneuroncannotrepresentanentireconcept?Aconcept mustberepre
8、sentedbyacollectionofneurons?Howmightsuchaconcept berepresented? D?O?Hebbproposedasolutiontothisproblem?areverberating circuitofcells?HebbcalledthisreverberatingcircuitaCellAssembly?CA? Somephysiologicalevidenceexists?forinstance?Abeles?hasexperimented withelectricalprobesinthebrainsofmonkeys?Therea
9、re?probesina small?cm?areaofthebrain?Theelectrodesshowrepeatingpatterns ofactivationwhenthemonkeyissensingorperformingaspeci?cevent? Whiletherehasbeenafairamountofworkinpsychology?andneuro? biologytoshowtheexistenceofCAs?therehasbeenlittlecomputational modellingofCAs?Thecomputationalmodelinghaseithe
10、rbeenfrom amoreabstractlevelthanneurons?orhasbeenofaverylimitednature ? Thispaperdescribesseveralinstancesanovelarchitecture?theCANT ?Connections?AssociationsandNetworkTechnology?model?Thenexttwo sectionsdescribethegeneralCANTmodel?Onesectionisonthebasic neuronandoneisonthenetandCAs? Theneuronistheb
11、asisofthemodel?Theneuronhasaxons?connec? tions?acurrentactivation?anactivationthresholditmustreachbefore ?ring?itisconnected?viaaxons?tootherneuronsinadistance?biasedfash? ion?theactivationofaneurondecays?and?ringneuronsfatigue? ThesectionontheCANTnetandthepropertiesofitsCAsisnext? CAsshouldfollowas
12、tandardcourseofactivation?Theyshouldemerge fromaninitialnetviatrainingandalocalizedlearningrule?Thissection alsodescribesthestructureofthenetusedintheexperimentsdescribedin thispaper? Thefourthsectiondescribesthreeexperiments?The?rstexperiment showstherobustnessofthemodel?Thesecondintegrateslearning
13、?The thirdshowstwoCAsinasinglenet? The?fthsectionfurtherdiscussestheresultsoftheexperiments?It elaboratestheproblemsencounteredduringtheexperimentsandhowthey relatetogeneralproblemswithCAs? Theconcludingsectiondiscussesthelong?termandshort?termplansfor ? themodel?Itconcludeswiththeoverallgoalsofthem
14、odel ?BasicNeuralModel Thebasisofthemodelistheneuron?Verycomplexmodelsofneuronsexist ?eg?However?thesemodelstendtotakequitesometimetorun andsincemanyneuronsareneeded?asimplemodelhasbeenemployed? CANTattemptstomakegoodtradeo?sbetweencomputationale?ciency andneurologicalvalidity?capturingtheimportantc
15、omputationalaspectsof theneurone?ciently? TheCANTmodelhassixneuralproperties? ?ConnectionStrength ?Activation ?ActivationThreshold ?VariableConnectivity ?Decay ?Fatigue The?rstthreearequitecommoninNeuralNetworkmodelsandthelast threearelesscommon? ?ConnectionStrength ACANTneuronhasconnectionstoothern
16、eurons?whicharesimilarto connectionsinsidebiologicalneuralsystems?Connectionsareunidirectional? Likemostneuralnetsimulations?theconnectionstrengthmayvarybased onalearningrule?Theconnectionmayhavepositiveornegativestrength? Continuousactivationissimulatedbytimesteps? Theaveragebiologicalneuronisactivatedbyabout?otherneurons? andinturnactivatesabout?otherneurons?Eachbiologicalneuron hasseveralaxons?whichhavefeettosendactivationtootherneurons?The CA
