sensory input pattern storage感觉输入模式存储

AMODELOFHUMANMEMORYJuly2008Memorypatternsareprobablynotstoredinthebrain’sneurons.Why?•Theytaketoolongtotrain.•New“on-the-fly”trainingdisturbsoldknowledge.•Oldknowledgemustberetrainedinordertobemaintained.Wherewouldtheoldknowledgebestored?Catch22.•Onceneuralnetworksaretrained,thetrainingpatternsarediscarded.Humanmemorydoesnotdiscardtrainingpatterns.•Humanmemoryrecordsinputpatternsingreatdetail.Cite:BélaJulesz’sworkwithrandomdotstereogramsandeideticsubject.SENSORYINPUTPATTERNSTORAGE•Sensoryinputpatternscomefromeyes,ears,tactile,olfactory,vestibularandothersensors.•Incomingpatternsarestoredinemptyfolders,wherevertheyarelocated.•Sequencesofpatterns,likevideosarestoredinthesamefolder.•Visual,auditory,andothersensorypatternsthatwerereceivedatthesametimearestoredinthesamefolder.•Only“interesting”inputpatternsarestoredinthememory.Theyremainfortherestofone’slife.Aneideticstoreseverything,interestingornot.•Sensoryinputpatternsgotoshorttermmemoryand,ifinteresting,aretransferredtomainmemoryandrecordedforlife.•Shorttermmemory(oftheorderofafewseconds)isusedinthedeterminationofwhatisinteresting.Theproblemsolvercanalsodecidewhatisinteresting.•Autoassociativeneuralnetworksareusedinthepatternretrievalprocess.•Patternretrievaloccursinresponsetopromptpatterns.•Promptpatternsmaycomefromsensoryinputs.•Thememoryisorganizedintheformofindependentsegmentstomakepossibleaverylargestoragecapacity.SEGMENTNNNSENSINGSTRONGTRAININGMUXSEGMENTN+1PROBLEMSOLVERHIT?HITLINESENSORYINPUTNNSENSINGSTRONGTRAININGMUXHIT?VCACSCSHORTTERMMEMORY“INTERESTINGINPUT”SWITCHSENSORYINPUTLINEMEMORYINPUTLINEVC=VISUALCORTEXAC=AUDITORYCORTEXSC=SENSORYCORTEXSENSORYINPUTPATTERNSTORAGEINBORNPATTERNSTORAGE•Inbornknowledgeintheformofpatternsispre-loadedinthedevelopingbrain’smemoryandremainsintactthroughoutone’slifetime.•Examplesofinbornknowledge:(a)Abirdbuildinganestinvolvescomplexconstructionin“safe”placessuchasrooftops,treetops,telephonepoles,etc.(b)Babyhorsewalkingandfindinglunchwithinhalfhourofbirth.(c)Humanbabysucking,crying,peeingandpooping.•Itisconjecturedthatthememorystoragemeansforinbornknowledgeisthesameasforsensoryknowledgegainedduringalifetime.•Itisconjecturedthatthememoryretrievalmeansforinborndataisthesameasforsensoryinputdata.•Inbornpatternsarestoredinfoldersin“memorysegment0”.MEMORYINPUTLINESEGMENT0NNSENSINGMUXMEMORYINPUTLINEINBORNPATTERNSSENSORYINPUTLINEHITLINESHORTTERMMEMORY“INTERESTINGINPUT”SWITCHINBORNPATTERNSTORAGEVCACHIT?SCVC=VISUALCORTEXAC=AUDITORYCORTEXSC=SENSORYCORTEXTHOUGHTPATTERNSTORAGE•Thoughtpatternsarealsostoredinmemory.Storagemeansandretrievalmeansforthoughtpatternsarethesameasforsensoryinputpatterns.•Thoughtpatternscomefromthe“problemsolver”.•Thedesignoftheproblemsolverisnotyetpartofthisstudy,butcouldbethoughtofasamechanismbasedonArthurSamuel’scheckerplayer.•Thoughtpatternsarealwaysinterestingandstoredinemptymemorylocations.•Storageofthoughtpatternstakesprecedenceoverstorageofsensoryinputpatterns.SEGMENTNNNSTRONGTRAININGMUXSEGMENTN+1STRONGTRAININGMUXPROBLEMSOLVERMEMORYINPUTLINETHOUGHTPATTERNSTHOUGHTPATTERNSTORAGENNPATTERNRETRIEVALSYSTEM•Patternsstoredinmemorycanberetrievedwithoutknowledgeoftheirstoragelocation.•Autoassociativeneuralnetworksarepartoftheretrievalmechanism.•Autoassociativeneuralnetworksaretrainedbyusingtheirinputpatternsasbothinputanddesiredresponsepatterns.Theyaretrainedtoproduceoutputsthatarereproductionsoftheirinputs.•Oncetrained,autoassociativenetworksproducesmallinput/outputdifferenceswhenpresentedwithpatternsthatweretrainedin,butlargedifferenceswhenpresentedwithpatternsthatwerenottrainedin.DéjàVu?Hitornohit?•Autoassociativenetworksaretrainedwithallthepatternsstoredintheconnectedmemoryfolders.•Theautoassociativenetworksarepromptedwithsensoryinputpatternsorthoughtpatterns.Visualinputpatternsforexamplearerotated,translated,scaled,brightened,contrasted,etc.bythe“visualcortex”VCwhileattemptingtomakeahit.Ifthereisahit,thehitpatternissavedandcomparedwiththecontentsofalltheconnectedmemoryfolders.Thepatternsofthefoldercontainingthehitpatternareretrievedandsenttotheproblemsolverwhichisthememoryoutput“customer”.•Thesepatternsinturnmaybeusedaspromptstoretrieveotherfolders.Thistypeoffeedbackcouldcausea“chainreaction”resultingintheretrievalofmanyinterrelatedfolders.(Ihavebeenspeakingwithsomeonefortenminutesbutwhatishisname?Oh,nowIremember.ItsJonathanJones.)SEGMENTNNNSENSINGSTRONGTRAININGPATTERNRETRIEVALSYSTEMMUXPROBLEMSOLVERHIT?SENSORYINPUTPATTERNSPROMPTLINEVCACSCTHOUGHTPATTERNSOUTPUTBUFFERHIT?SEGMENTN+1NNSENSINGSTRONGTRAININGMUXHIT?BUFFERHIT?RETRIEVEDPATTERNSLINERETRIEVEDPATTERNSPROMPTSIGNALPROMPTPATTERNSSHORTTERMMEMORYVCACSCPROMPTLINEVC=VISUALCORTEXAC=AUDITORYCORTEXSC=SENSORYCORTEXTRAININGDURINGNON-REMSLEEP•Itisspeculatedthattheautoassociativeneuralnetworksaretrainedduringnon-REMsleep.•Multiplexerssensethememoryfolders,sequentiallyfeedingthepatterncontentstotheautoassociativenetworksfortraining.•Thetrainingprocessinitiatesautomaticallyoncethebrainisin“sleepmode”.Thiscontinuesthroughoutthenightduringperiodsofnon-REMsleep.SEGMENTNNNN