computational cognitive neuroscience understanding brain mind based biologically computational models networks neuron units field number different including neuroscience computation cognitive psychology computational cognitive neuroscience difficult subject difficult well book intended support subject providing coherent principled introduction main ideas field course text researchers related areas want learn new field idea book factors positions asked cognitive neural network modeling unable find existing textbook see discussion factor computational algorithm called leabra combined coherent major established well neural network principles model cognitive phenomena leabra provided means computational implementation wide range cognitive phenomena enabling learn phenomena unified framework possible complicated computational mechanisms algorithm greater underlying biological properties neurons cortical networks algorithms enabling behavior brain link clearly understood factor completion computer simulation environment called pdp developed provides intuitive flexible models allows presentation easy simulation exercises book powerful serve researchers develop models own research text subject computational cognitive neuroscience potentially set write different component aspects computation cognition neuroscience existing focusing specific issues firing patterns individual neurons oriented mathematically computational properties networks abstract models cognitive phenomena field based large part wide scope issues involved biological computational properties cognitive function requires perspective hand hand greater details captured similar main existing text scope issues original field pdp parallel distributed processing later works present computational biological mechanisms cognitive perspective attempt modeling range cognitive phenomena read convey pdp present different algorithms ideas simulation exercises separate text play role understanding complex behavior models finally neuroscience short treatment models abstract biologically implausible mechanisms objective replicate scope original pdp integrated unified manner related biology cognition provided intuitive simulations step way necessary achieve scope focusing consistent set principles form computation neuroscience cognition leabra algorithm provided computational framework objective point write textbook based kind new computational algorithm standard field fact early similarly start answer critical question simple leabra principled widely algorithms field including error backpropagation hebbian competitive learning simple point neuron activation function text standard algorithms context principles shows leabra implements same principles biologically plausible way addition extra benefit principles mutually consistent mutually beneficial combining algorithm repeated isnt consistent set principles underlying models summary book attempt integrate range fields phenomena coherent relatively easily reader level result viewed integration existing knowledge found process putting ideas led emergent phenomenon whole greater sum parts come away sense computational cognitive neuroscience writing book feel brain think challenging science answer question new data constantly work researchers related fields brain record biological thought color images images results important techniques understanding neural cognition considerably different approaches cognitive neuroscience goal important cognitive neuroscience remain years come complex difficult understand sequences images brain thinking network regions interact complex ways changing patterns thought picture worth words language fails attempt capture computational models based biological properties brain provide important language understanding complexity example models capture flow information eyes letters words parts brain activated different word resulting integrated text understanding phenomenon models enable describe levels processing involved common set underlying mechanisms implemented computer tested ultimately understood book provides introduction sub known computational cognitive neuroscience simulating human cognition based biologically networks neuron units provide style textbook treatment central ideas field integrated computer simulations allow own explorations presented text chapter present overview basic history computational cognitive neuroscience followed overview subsequent chapters range basic neural computational mechanisms part range cognitive phenomena part including learning perception memory language level higher cognition whole idea cognitive neuroscience notion human thought explained same way science reducing complex phenomenon cognition simpler components underlying biological mechanisms brain continues standard method fields example matter reduced components helps explain properties different kinds matter ways interact similarly biological phenomena explained terms actions underlying natural think terms physical systems explaining cognition terms physical brain possible achieve form terms abstract components system argue forms explanation form explain previously thing terms familiar constructs understand definition word terms familiar words years explain human cognition different example cognition explained assuming based simple assuming works standard serial computer approaches idea look brain itself language principles upon explain human cognition seems likely brain ultimately responsible cognitive neuroscience components based physical human cognition brain physical importance physical science well captured following recall provide source science components theory real things example space clearly true physical actually measure components theories atoms etc interesting biological biology theories common components etc based theory based physical evidence structure possible develop theories biological function terms real underlying components etc measured similarly previous current theories human cognition based constructs attention working memory based analysis behaviors physical measured cognitive neuroscience forms cognitive explain cognitive phenomena terms underlying neurobiological components principle measured aspects science particularly study human cognition emphasis process reducing phenomena component pieces essential complementary process pieces larger phenomenon refer latter process simply say brain neurons explain billions neurons produce human cognition computational approach cognitive neuroscience critically important difficult verbal arguments human cognition complex phenomenon action large number components show implement behavior components computer test capable phenomena simulations crucial developing understanding neurons produce cognition true emergent phenomena emerge interactions obviously present behavior individual elements neurons whole greater sum parts importance areas science cognitive neuroscience recently relatively fast computers illustration importance say system composed components show components interact produce overall behaviors interact produce changes speed effects emerge interaction property component shows simple illustration importance understanding systems here sufficient say system composed components shown panel specify interact shown panel interaction important behavioral properties changes speed emerge example smaller drives larger achieves decrease speed increase same driving interact smaller produce opposite effect essentially means behavior emerge interaction clearly property individual similarly cognition emergent phenomenon interactions billions neurons physical computational cognitive neuroscience appear sound straightforward approach understanding human cognition extreme complexity lack knowledge brain cognition produces result researchers notion hierarchical levels analysis order deal complexity clearly levels underlying mechanism appropriate explaining human cognition example try explain human cognition directly terms atoms simple focus higher level mechanisms exactly level right level essentially level presented book represents best time approach towards thinking issue levels analysis suggested introduced notion computational levels analogy computer take example list numbers specify abstract terms computation performed numbers list next next etc abstract computational level analysis useful different exactly think executive summary level details actually occurs different adopt tradeoffs terms speed amount memory etc critically algorithm provides information actually implement specify details language variable etc details left level actually written particular computer particular language etc levels corresponding emphasis computational levels early artificial cognitive psychology cognitive science based idea ignore underlying biological mechanisms cognition focusing important computational cognitive level properties traditional approaches based assumption brain works standard computer computational levels important details underlying neurobiological implementation version emphasis level computational widely complexity biology psychology approach here assumes possible identify optimal computation function performed person context whatever brain somehow same optimal computation ignored example argued memory curves tuned expected frequency retrieval demands items stored memory view matter memory mechanisms work ultimately driven matching expected demands items turn assumed follow general sound case definition ends number assumptions including nature underlying implementation real independent basis short defined purely objective terms optimal situation depends detailed thing levels approaches appear suggest level largely irrelevant standard computers true effectively equivalent level issues affect computational levels analysis effect higher levels analysis already assumed general form proper credit shaping whole place parallel computers people beginning limitations computation algorithms assume standard serial computer based address memory entirely new algorithms ways thinking problems developed order take advantage parallel computation brain clearly parallel computer billions computing elements neurons simple ideas based standard computers end researchers emphasized level primary computational argued cognitive models detailed neurons resulting model contains important biological mechanisms approach complementary emphasize purely computational approach clear understanding biological properties important ends complicated models difficult understand provide insight critical properties cognition further models inevitably fail represent biological mechanisms possible detail sure important missing level adopt fully interactive approach emphasizes connections data relevant levels reasonable balance simplified model known biological mechanisms possible place bottom working biological facts cognition top working cognition biological facts approaches example useful take set facts neurons encode set equations computer see kinds behaviors result depend properties neurons useful think cognition particular case computational level principled basis implementation see well know brain well cognitive job kind neurobiological cognitive principled computational otherwise emphasized text basic levels analysis text intermediate level help order summarize approach avoid associations terminology adopt following hierarchy levels see essentially simple level physical hierarchy lower level consisting neurobiological mechanisms upper level consisting cognitive phenomena end able explain cognitive phenomena directly terms action underlying neurobiological mechanisms order help levels analysis intermediate level consisting principles presented text think brain cognition fully described principles play role shown side figure serve clear connection certain aspects biology certain aspects cognition understanding principles based level computational aspects cognition keeping discussion levels want avoid principles provide level description view computational level thinking data basic empirical levels cognition relevant principles shaped help good balance primary levels analysis essentially levels analysis position biological mechanisms operating level individual neurons explain relatively complex level high cognitive phenomena question basic neural mechanisms understanding undoubtedly product billions neurons include neurons simulations seen scaling issue way scaled model real brain important emphasize need scaling least partially issue limitations currently available computational resources possible put following arguments test future larger complex models constructed scaled models easier understand good place begin computational cognitive neuroscience approach scaling problem following ways target cognitive behavior expect obtain models similarly scaled compared actual human cognition show simulated neurons units model approximate behavior real neurons meaning build models multiple brain areas neurons areas simulated units argue brain quality same basic properties apply physical scales basic properties individual neurons show higher levels relevant understanding scale large behavior brain argument amounts idea neural network models performing essentially same type processing human particular task reduced problem detailed information content human equivalent course phenomena different scaled content dimension seems reasonable allow important properties relatively scale invariant example argue major area human cortex reduced small portion content actually pixel retina pixels important aspects essential computation information preserved reduced model reduced cortical areas connected imagine useful simplified model reasonably complex psychological phenomena argument individual effects neurons neurons likely roughly similar average effects neurons likely true assuming similar patterns weights areas neurons extent neurons communicating signals depend largely average firing matter debate current evidence rule possibility see details course encode information robust effects noise constant scaled model approximation original argument applies computational neural network algorithms including main text average firing based rate activation again next chapter finally reasons believe brain character likely least cortex effective properties range long connectivity similar local range short connectivity example short range long connectivity produce balance excitation inhibition virtue connecting excitatory inhibitory neurons model based properties range short connectivity cortical area describe scale larger model containing cortical areas simulated level reason basically same averaging neurons average behaves roughly same individual neuron levels description similar self means place short arguments provide basis models based neurobiological data provide useful accounts cognitive phenomena involve large widely distributed areas brain models described book issue remains open important question computational cognitive neuroscience following perspective provides overview important issues shaped field field computational cognitive neuroscience relatively easily large number related time research aspect cognition neuroscience computation potential important contribution field entire space book account relevant history field section intended provide brief overview particularly relevant context motivation approach specifically focus understanding networks simulated neurons lead interesting cognitive phenomena occurred initially again period present form main approach clear said neural network modeling approach provides crucial link networks neurons human cognition field cognitive psychology early following key perspective least associated new field emphasis internal mechanisms cognition particular explicit computational models simulating cognition computers problem solving mathematical dominant approach based computer metaphor human cognition processing standard serial computer time researchers modeling cognition networks simple neuron processing elements operating parallel computer metaphor approach book simple neuron models significant computational limitations unable learn solve large basic problems researchers studying network neural models critical field back real early psychological computational based activation dynamics networks backpropagation learning algorithm parallel distributed processing pdp established neural network models critically backpropagation algorithm limitations earlier models enabling essentially function learned neural network led new cognitive modeling goes name based backpropagation algorithm backpropagation represented step computationally step backwards biological perspective clear implemented biological mechanisms based backpropagation cognitive modeling clear biological basis causing researchers same kinds arguments computer metaphor approach computational level arguments discussed previously biological issues field essentially computational cognitive psychology based neural processing principles true computational cognitive neuroscience parallel influence neural network models understanding cognition rapid oriented biologically modeling identify categories type research divide biological models emphasize learning learning non types models include detailed models individual neurons information approaches processing neurons networks neurons original models hold considerable due underlying mathematical terms statistical research led important tends direct relevant cognitively issues network itself provides important principles see based biologically learning models focused learning early visual system emphasis hebbian learning importantly large body basic neuroscience research idea hebbian mechanisms operating neurons important cognitively areas brain hebbian learning computationally weak suffers limitations similar generation learning mechanisms widely cognitive modeling generally learn psychological tasks addition cognitive biological neural network research considerable work done computational end apparent mathematical basis neural networks common statistics computational connection further recently bayesian framework statistical applied develop new algorithms generally understand existing ones models developed point provide framework learning works reliably wide range cognitive tasks simultaneously reasonable biological mechanism principal researchers computational end field concerned theoretical statistical learning kinds issues cognitive biological ones short field state computational cognitive psychology primarily focused understanding human cognition close underlying biological focused information constructs computationally weak learning mechanisms close cognition learning focused computational level analysis involving statistical constructs close biology cognition think strong set cognitively relevant computational biological principles years time attempt integrate principles coherent overall framework brief overview provides useful describing basic characteristics approach taken book basically tried develop single coherent algorithm includes based backpropagation driven error learning hebbian learning principles central principles network interactive satisfaction constraint style processing underlying implementation direct ion channels behavior real neurons described incorporates number established well anatomical properties neocortex described detailed connections biology cognition algorithm way consistent well established computational principles recent shaped nature integrated algorithm development biologically plausible way implementing backpropagation learning algorithm resulting algorithm called generec consistent known biological mechanisms learning greater biological properties brain including interactivity allows realistic simulated neurons development mechanism neural competition powerful distributed representations combined interactivity learning way previously possible competition number important functional benefits required hebbian learning mechanisms approach based combination generec hebbian learning competitive interactive activation dynamics set properties leabra algorithm local driven error associative describing hebbian realistic biologically algorithm text leabra pronounced emphasizes balance different achieved algorithm consider leabra coherent framework computational cognitive neuroscience provides useful existing ideas help identify limitations problems need solved future following section discuss basic principles approach detail finally worth noting arguments proposed favor developing unified theories cognition apply notion developing unified neural network algorithm thought unified theory essential principles underlying cognition requires terms overall architecture theory cognition essential point here relatively easy relatively specialized theories specific phenomena taking range data constraints theory account data likely true model architecture clear process unified architecture common set ideas net case neural network models models generally constraints modeling process fact single set principles implemented leabra algorithm model wide range phenomena covered book measure discussion benefits unified model number general issues regarding benefits computational modeling cognitive neuroscience think benefits generally important potential problems associated well provide brief summary advantages problems here note points mutually fact tend load central issues specific derivative issues provides richer perspective advantages computational model forces explicit assumptions exactly relevant processes actually work example people tried explicit computational models object recognition didnt seem difficult problem story going implement model say didnt computational model deal complexity ways verbal arguments times interactions emergent phenomena model produce satisfying explanation otherwise hand arguments computational model control variables precisely real system replicate results precisely enables explore role different components ways otherwise impossible computational models enable data multiple levels analysis integrated related example computational models book show biological properties rise cognitive behaviors ways impossible simple verbal arguments theories cognition particular components theory end work theory example executive theory frontal pre cortex function executive explaining good brain areas explained well put box computational model provide insight works example providing explanation phenomenon phenomena complex interactions components obtain insight putting set principles assumptions implementing seeing happens parameters see kinds effects computational model sense real exists behaves novel running model novel context exercise verbal theory due lack specificity flexibility verbal constructs computational model forces consequences assumptions generally possible computational models general flexible account new data via acts verbal hand model modified account new data clear exactly changes easily evaluate resulting previous theory difficult people detect purely verbal theory hard time keeping computational model actually work computational model forces aspects problem otherwise ignored considered irrelevant ends aspects see problems useful exercise problems anyway problems properties relevant computational models require know relevant details end number arbitrary assumptions learn model partially correct assumptions end inevitably ends empirical question depends wrong assumptions influence results identify critical principles models behavior demonstrate relative assumptions work problem model viewed concrete principles end itself essential step principles clear taken theory easy computational model theory point fact model set data explanation data said model viewed concrete principles way model principles account data clear theory level principles interactions model approach complexity model advantages modeling place detailed behavior model terms principles nature models explain inevitably successful models true principles model properly understood took powerful model fit data tell principles model clear account data contribution now relationship principles data completely principle train neural network learn task data fact subjects learn task isnt saying apply model range data different tasks greater detail task detailed properties learning process principles work tested continue fit data constitute important advance end cortex understood relatively small number powerful principles seems rule possibility powerful model account lot data final chapter book revisit issues again benefit comes aspects human cognition particularly kinds neural mechanisms described text describe important aspects here order further connections cognition aspects cognition obvious average person nature own cognition tend emphasize aspects definition aware appear serial thought time focused subset things occurring brain fact undoubtedly standard serial computer model understanding human cognition point comparison discussion argue aspects human cognition cognition possible relatively understand cognition focusing difficult necessary keep place whole thing important notion space phenomena focused metaphor limits following light important ideas keep aspects cognition discuss gradedness interactivity competition learning cognitive neuroscience unable say useful experience phenomenon book note last chapter book specifically level higher cognition closely associated experience present set ideas models provide basic mechanisms principles developed rest book sequential discrete focused nature experience view properties due particular brain areas prefrontal cortex pfc hippocampus result emergent phenomena arise basic properties neural processing processing system chapter emphasizes continuum processing able same time simple case parallel processing thing same time essential understanding neural basis human cognition level analysis know neurons human brain contributes bit overall human cognition brain neurons important true figure based different brain removed tested simple tasks found didnt seem matter large amount solve simple tasks figure problem tasks simple solved number different mechanisms example lesion completely ability visual processing solve task later research showed part brain contributing own special way overall cognition particular cases contributions redundant good thing survival perspective function basic level lots examples typical experience parallel processing evident driving etc name effect speaking reading reading name last example particularly interesting processes reading speaking know lots examples parallel processing taking place cases simply aware multiple things going example look visual scene part brain processes visual information identify seeing part things people lesions brain areas things apparently way view world product bunch specialized brain areas operating fashion opened techniques brain obvious multiple brain areas inevitably activated cognitive tasks challenge parallel processing want understand cognition difficult figure processes sub eventually end sensible whole contrast cognition bunch discrete sequential steps task easier identify steps sequence body problem understanding interaction things simple number things operating same time mutually hard figure going virtue approach cognition presented book based start parallel processing providing powerful mathematical intuitive understanding interactions large number processing units neurons lead useful cognition example graded nature categorical representations middle item categories contrast discrete binary memory representations standard computers brain graded nature see next chapter neurons integrate information large number different input sources producing essentially continuous real valued number represents relative strength inputs compared inputs received neuron graded signal rate firing activation neurons function relative strength value result neurons good graded signals convey extent degree true example neuron convey object last likely research people categories shown tend things graded manner according close item example category graded activation values important representing continuous dimensions position force color coarse coding function basis representations shown here units shown graded activation signal roughly close point continuous dimension units preferred point defined point response gradedness critical kinds perceptual motor phenomena deal continuous underlying values position force color turns brain tends deal same way continuum different neurons represent different values continuum cases essentially placed points respond graded signals reflecting close current preferred value see type representation known coarse coding basis function representation actually precise indication particular location continuum weighted estimate based graded signal associated basis values important aspect gradedness fact neuron brain receives inputs neurons individual neuron critical functioning neurons contribute part graded overall signal reflects number neurons contributing well strength individual contributions fact rise phenomenon function increasing amounts damage neural explain saying neurons reduces strength signals eliminate performance entirely contrast standard computer tend fail obvious equally important aspect gradedness way processing happens brain familiar associated trying remember come mind immediately trying different things hit upon right thing psychologists tongue phenomenon tongue remember gradedness critical here allows brain bunch relatively weak ideas see ones stronger things ones weaker away similar bunch relatively weak factors add support idea single clear discrete reason computationally phenomena examples bootstrapping multiple constraint satisfaction bootstrapping ability system itself taking weak information eventually producing solid result multiple constraint satisfaction refers ability parallel graded systems find good solutions problems involve number constraints well discuss greater length basic idea factor constraint solution rough proportion graded strength importance resulting solution represents kind constraints roughly same direction number constraints remain sounds write equations works run simulations showing action way brain standard serial computer processing direction time lots things same time going backwards known interactivity processing think brain organized processing areas visual stimuli example processed simple level low way terms oriented lines present image subsequent stages features represented combinations lines parts objects objects etc least correct system interactivity amounts bottom top processing information flows simple complex complex simple combined gradedness interactivity leads satisfying solution number otherwise cognitive phenomena example well documented people faster accurate letters context words context random letters word effect finding unidirectional serial computer perspective letters identified words read context word help letter finding seems natural interactive processing perspective information higher word level come back affect processing lower letter level gradedness critical here allows weak estimates letter level activate word level comes back letter estimates overall representation word letters explanation word effect proposed interactivity important bootstrapping multiple constraint satisfaction processes described allows constraints levels processing bootstrap converge good overall solution ambiguous letters context words example interactivity level word processing level letter processing examples interactivity psychological involve stimuli ambiguous level context higher level processing example shown words constrain ambiguous stimulus look case saying competition good thing true brain evolution brain competition neurons leads selection certain representations strongly active context bootstrapping described analogy process survival idea important force shaping learning processing encourage neurons better adapted particular situations tasks etc argued kind competition provides sufficient basis learning brain find number important mechanisms biologically well known inhibitory interneurons provide mechanism competition areas brain central cognition cognitively competition evident phenomenon attention closely associated perceptual processing clearly evident aspects cognition phenomenon spatial attention demonstrated posner task good example here ones attention particular region visual space cue bar computer screen stimulus target presented cue opposite region space subject respond pressing key computer whenever detect onset target stimulus result target detected significantly faster location compared non happens faster move ones eyes kind internal attention result processing cue stimulus see results related ones accounted simple model competition neurons mediated inhibitory interneurons well nature versus debate development human inevitably terms genetic configuration brain results based experience learning important contributions fact advance understanding exactly genetic configuration learning interact produce human cognition understanding major goal computational cognitive neuroscience unique position able simulate kinds complex subtle exist certain properties brain learning process addition learning process learning occurs constantly cognition possible identify relatively simple learning mechanism appropriately initial architecture organize billions neurons human brain produce whole range cognitive functions exhibit obviously cognitive neuroscience reason text dominated properties learning mechanism biological cognitive environment operates results course focus importance genetic basis cognition feel context learning mechanism genetic parameters fully understood role itself shaping understood context emergent process takes learn important gradedness aspects biological mechanisms discussed problem learning considered problem change learn change way information processed system easier learn system responds changes graded manner way behaves allows system try new ideas ways processing things kind graded indication changes affect processing exploring lots changes system evaluate improve performance learning bootstrapping phenomenon described respect processing earlier depend number weak graded signals exploring possibly useful proceed further building look kind bootstrapping possible discrete system standard serial computer responds small changes way putting computer typically works right missing step typically provides indication well perform complete same thing true system relationships typically discrete systems typically provide effective learning present view learning dominated bootstrapping small changes idea think kinds learning discrete nature trial error kind learning familiar experience here discrete hypothesis behavior trial outcome error update hypothesis next time discrete find best implemented same kinds graded neural mechanisms kinds learning discrete kind learning associated particular discrete facts events appears brain specialized area particularly good kind learning called hippocampus properties learning discrete character discuss type learning further described book based relatively small coherent set principles introduced part text applied part range different cognitive phenomena principles implemented leabra algorithm exploration simulations integrated chapter form part text tried ensure lot book simulations emphasis book understanding set principles leads wide human cognitive phenomena provide detail aspects neuroscience computation cognition potentially related central present expect mathematical computational researchers cognitive psychologists find book detailed area provide coherent framework areas consistent well established facts domain provide useful means knowledge areas current debate upon choice presented relevant arguments data presented short existing relevant supporting arguments details necessarily presented idea book cases likely find relatively coherent clear picture rapidly maintain fit working memory reader follow rapid enable motion ideas book proceed rapidly cognition order overall picture emerge leaving reader facts tradeoffs cover large space existing neural network algorithms cover range computational algorithms provided interested reader further algorithms ideas covered here represent novel potentially provide important principles cognition approach introduce new principles existing ones found obviously case scope cognitive phenomena covered book said conclude principles anyway considered final inevitably rough domain level satisfaction present time overview approach sections clear leabra algorithm book incorporates important ideas shaped history neural network algorithm development book principles introduced simple clear manner possible explicit development ideas comes time implement explore ideas simulations leabra implementation consistency reader knowledge standard algorithms unified integrated perspective helps understand relationship learning work different algorithms favor integrated algorithm necessary fully understand practical level simple required understand algorithm wide interactive computer simulations relevant principles interact produce important features human cognition detailed step step exploring simulations provided set exercises evaluation purposes models wide range human cognitive phenomena presented perception memory language level higher processing controlled appropriate focus cognition consider perception form cognition emphasize processing takes place human neocortex typically referred simply cortex large neurons part brain plays important role cognition interesting property relatively area area same basic types neurons present same basic types connectivity patterns principally allows single type algorithm explain wide range cognitive phenomena terms detailed organization already mentioned main parts book part basic neural computational mechanisms part large scale brain area organization cognitive phenomena chapters part individual neurons networks neurons learning mechanisms chapters part perception attention learning memory language level higher cognition chapter large scale brain area functional organization chapter begins overview followed detailed table let reader know scope covered chapters appear end chapter read list reading provided key words defined extensively easy found index explorations evaluation answer key available required provide written answer good idea look consider answer important issues book original pdp distributed parallel processing considered field remain relevant collection important early neural networks basic cognitive neuroscience levels analysis suggest chapter chapter computational modeling artificial neural networks see neuron unit abstract computational models provides basic information processing mechanisms human cognition computationally neurons function dedicated specialized detectors smoke detector neuron integrates information different sources inputs single valued real number reflects well information matches neuron specialized detect sends output reflects results evaluation standard fire integrate model neural function output provides input neurons continuing information processing network interconnected neurons biologically neuron provides basic mechanisms necessary integration evaluation information receives results neurons self way processes information chapter provides overview level computational description neuron detector biological mechanisms underlie neural information processing focus neuron cortex consistent general focus book cortex sum biological mechanisms known activation function resulting output neuron called activation value actually bit known neural activation function leabra algorithm balance based biologically mechanisms hand keeping computational implementation relatively simple point neuron activation function biological details regarding basic dynamics information processing neuron spatial extent neuron single point simplifies computational implementation simulations chapter basic properties activation function arises underlying biological properties neurons show activation function understood terms mathematical analysis based bayesian hypothesis testing standard serial computer basic information processing simple memory manipulations retrieval brain based same kinds appear case order understand functions biological mechanisms neuron need come computational level description neuron purpose section standard computer point comparison standard computer memory processing separated distinct processing central processing unit information memory processed stored back memory contrast brain appears parallel distributed processing pdp processing occurs simultaneously parallel billions neurons distributed brain memory processing similarly distributed brain computational level description neural processing explain neuron provides memory processing functions distributed way producing useful neurons work central idea explain neuron detector simply put neuron detects existence set conditions responds signal extent conditions think smoke detector constantly air looking conditions indicate presence fire brain neurons early stages visual system constantly visual input looking conditions indicate presence simple visual features bars light position orientation visual scene higher visual system neurons detect specific sets objects emphasize useful view function neuron detector content exactly detecting well captured relatively things smoke detectors contrast detectors neuron considerably complex different inputs huge dynamic network neurons possible describe roughly neuron detecting need necessarily case see later neurons contribute overall computation number detecting different hard subsets combinations things neurons response context sensitive depends things otherwise example oriented bar detectors early visual system respond visual scenes bar light appropriately oriented detector result aspects scene context bar light further detector respond time scene viewed next result dynamic changes network determine ones focus attention way think context sensitivity neuron act dynamic detector plays multiple roles different situations model smoke detector analogy further detection obviously appropriate sensory processing describe processing motor output pathways brain purely internal processing neuron detect motor response output leads response abstract internal actions attentional system thinking appropriate word describe thought terms detecting appropriate conditions things virtue detector model easily complexity basic framework understood initially simple intuitive terms analogy simple smoke detectors detector model neuron emphasizes important properties emphasizes neurons dedicated specialized smoke detectors neuron memory cell computer hold arbitrary information neuron dedicated detecting specific set things matter difficult describe enables neuron simultaneously perform memory processing functions memory amounts conditions applies inputs order detect whatever detects processing way goes conditions communicating results neurons neurons traditional line specialize step process whereas computers build different parts themselves analogy helps explain useful emerge actions large number individual neurons captures achieved parallel processing compared serial processing specialized nature neural detectors important enabling refer representation neuron group neurons term history widely way roughly consistent idea neurons representation simply detects neuron detects oriented bar position said represent oriented bar position common refer pattern activity multiple neurons representation case refers things group detects term reflect properties neural response itself example distributed representation multiple neurons participate representing thing next chapter finally important note traditional view brain standard serial computer inconsistent biological facts rise detector model traditional production system model cognition potentially consistent least general here cognition simulated seen detectors ones fire activated detect appropriate configuration information active system result firing activation state appropriate way example production adding numbers detect presence digits active add fires digits sum state provides appropriate configuration activating production revisit issue see neurons implement production constraints dedicated specialized representations detector model function corresponding neural components detector model provides direct interpretation functions components neuron shown view neural functioning known fire integrate model simplest view neural functioning detector needs inputs provide information detection elaborated neuron receives inputs via synapses typically occur dendrites extend large cell body neuron human brain relatively neurons directly connected sensory inputs rest getting inputs earlier stages processing multiple levels detectors lead powerful detection work directly raw sensory inputs job difficult hard figure kinds input information neuron detection related sensory inputs regardless gets detector needs perform processing inputs relative contribution input value overall detection controlled weights implemented relative synapse inputs neuron aka synaptic synaptic strengths weights provide critical parameters neuron detects essentially neuron detect different patterns inputs input patterns best fit pattern weights producing detection response appear neural individual inputs treating part overall input pattern done combining integrating weighted inputs form measure degree input pattern fits expected happens properties dendrites resulting eventually membrane potential voltage cell body central part neuron reflects results combination detector needs evaluate extent combined input sufficient count detected worth communicating smoke want evaluation shouldnt gets want real fire amount setting threshold combined input responding threshold neuron mechanism cell body implementing threshold modified result neural activity threshold turned neuron saying finally detector needs communicate results processing form output neurons detected looking smoke output smoke neurons provide graded output signals reflect weak outputs low strong outputs high ones neural outputs communicated long process cell body called axon forms synapses neurons dendrites providing inputs described repeating chain neural processing relatively neurons actually produce physical output way forms true output entire network cognitive models details output simulated internal output representation captures necessary relevant output information order detector model neuron real consistent larger understanding networks detectors perform useful computations exhibit human cognition details larger picture subsequent chapters critical idea learning provide means getting bunch neural detectors useful learning neurons implemented weights synaptic provide main parameters neuron detects shaping weights learning shapes neurons detect turns powerful ways sure neuron learns detect end useful larger task performed entire network overall result network learning contains number detectors way produce proper outputs set inputs internal detectors related way humans way perform task learning shape networks neural detectors ways result good task performance understand individual detectors work typically understand direct mathematically way entire network behaves general level reason network neurons possible different ways detectors configured different combinations activity detectors understand consequences different detailed way put way neural networks complicated systems easily computational level researchers provide better ways learning processing networks inevitably requires significant limitations behavior network state potentially consider analogous situation standard serial computer say understood pieces basic solid put shape computer right thing kind greatly objective better computers number factors situation considerably better seem already mentioned important existence learning mechanisms derived sound mathematical basis provide network converge good solutions wide range tasks addition useful ways mathematically summarizing aspects networks behavior ways help understand behaves way principally energy functions covered next chapter finally large number principles explain important aspects behavior networks situations provide rich basis understanding principles explained book appropriate bottom line practice networks neurons successfully perform wide range cognitive tasks see subsequent chapters described function neuron detector now explore biological mechanisms enable neuron integrate combine information inputs mechanisms based movement atoms called ions neuron generating currents neuron electrical system understood basic principles electricity addition electrical need understand behavior concentrations areas relatively high low ions liquid process called diffusion basic ideas introduced applied towards understanding neuron works turns neuron different concentrations ions cell enables diffusion forces interesting basic result neuron integrates inputs controlled flow specific ions cell turn changes electrical potential voltage neuron drives neuron produce outputs function potential electricity behavior movement called charge basic matter purposes negative charge positive charge same magnitude opposite sign atoms equal numbers net charge ions atoms versa vice negative positive net charge typically difference indicated writing name ion corresponding number plus minus ions relevant neural functioning potassium calcium interesting thing charge rule negative charges close possible positive charges versa vice means larger concentration ions negative ions towards area happens electrical current simply movement charge place sign same charges ions area move away assuming positive negative ions ions liquid case neurons same type current caused positive ions leaving negative ions negative current opposite case positive ions negative charges leaving area positive current extent amount positive negative charge place called electrical potential reflects potential amount opposite charge area again negative charges rise negative potential positive charges positive potential note area starts opposite charges starts potential new opposite charges net charge potential area changes function current coming area play important role neuron behaves excited charges sketch ohms law action charges leads potential difference charge represented line drives current channel conductance ions move caused liquid small channels pass sense greater ions move greater amount potential required imagine top bunch higher potential ion relationship known ohms law current amount motion electrical potential possible same thing slightly convenient form terms called conductance conductance represents easily ions place labeled ohms law written terms forms basis equation describing neuron integrates information summarized brief neuron channels determine conductances type ion function input receives potential referred membrane potential updated computing current ohms law tell charges moving neuron membrane potential change applying ohms law again compute changes potential time model neuron computes turns due combined forces diffusion explained electrical potential ion respond membrane potential different way ion ends own unique way contributing overall current add currents overall current addition electrical potentials main factor causes ions move neuron force called diffusion recall electrical potentials caused concentrations positive negative ions location diffusion comes play concentrations put simply diffusion causes type distributed space time large concentration location diffusion acts spread concentration possible sounds simple underlying causes diffusion complicated results fact atoms liquid constantly moving results process tends average cause diffusion direct force electrical potential indirect effect well key thing diffusion type move gets large concentration type ion place equally large concentration ion same charge contrast electricity care different types ions positive charge same perfectly ions large concentration sketch diffusion action types ions move same direction independent due effects random motion fact direct force treat diffusion reliable effect electrical force convenient write similar force equations diffusion essentially same terminology electricity describe happens ions result concentration differences charge differences imagine box liquid separated closed large number particular type ion imagine blue concentration difference results diffusion potential cause ions move removed diffusion current generated ions move side case electricity reduces diffusion potential eventually well diffusion acts electrical conductance relationship analogous ohms law holds order compute current ion produce need way summarizing results diffusion forces ion accomplished special equilibrium point diffusion forces balance concentration ions exactly individual ions moving randomly point zero current respect type ion current function net motion ions course simple system electrical forces equilibrium point electrical potential actually zero different concentrations ions neuron resulting diffusion forces equilibrium point typically zero electrical potential absolute levels current involved generally relatively small assume relative concentrations ion cell remain relatively fixed time addition see neuron special mechanism maintaining relatively fixed set relative concentrations turns equilibrium point expressed amount electrical potential necessary effectively constant diffusion force potential called equilibrium potential reversal potential current changes sign side zero point driving potential flow ions drive membrane potential towards value equilibrium potential particularly convenient correction factor ohms law simply away actual potential resulting net potential call diffusion version ohms law applied ion ion basis current type ion see now put basic principles electricity diffusion work understanding neuron integrates information say bit neuron kind environment neuron cell cell body purposes think neuron liquid membrane generally prevents things leaving cell including ions interested membrane difference electrical charge neuron referred membrane potential practical purposes ions interested cross membrane specifically let special called channel imagine channels allow specific types ions cell opened closed mechanisms important later level conductance ion determined number open channels pass type ion neurons liquid environment brain called space similar interesting thought place certain amount results reasonable concentration ions reasons described ions lower concentration neuron space summary major activation ions channels now consider type ion turn assess electrical diffusion forces channels allow flow neuron summarized exists greater concentration neuron diffusion force neuron order neuron positive charge relative equilibrium potential positive typical value relatively low internal concentration produced active energy mechanism called potassium ions neuron smaller amount ions net effect produce negative resting potential potential holds inputs coming neuron positive ions cell net negative charge negative charge typically primary types channels pass important purposes excitatory synaptic input channel opened neurotransmitter glutamate released sending neurons activated dependent voltage channel opening dependent level membrane potential plays central role action potential described later channel implemented directly leabra algorithm summarize action potential mechanisms simpler implementation general plays central role excitation activation neuron ions exist greater concentrations neuron again diffusion force neuron negative charge diffusion force negative potential neuron equilibrium potential negative typical value right note same negative resting potential caused maintained low internal concentration additional mechanism necessary maintain concentration maintained negative resting potential itself main channel inhibitory synaptic input channel opened neurotransmitter released inhibitory interneurons activated note due equilibrium potential similar resting potential inhibition neurons effect current generated neuron starts excited membrane potential phenomenon described inhibition potassium exists greater concentrations cell previous ions diffusion force neuron positive charge neuron needs negative charge keep leaving equilibrium potential negative value typically tend maintain concentration due negative resting potential internal concentration bit potassium ions cell negative equilibrium potential concentration difference larger different types channels relevant purposes leak channel constantly open lets small amounts potassium turns channel lets small amounts equilibrium potential same ion same resting potential roughly dependent voltage channel effects excitation produced action potential larger amounts neuron excited again channel implemented directly algorithm deal details action potential generation type channel present function amount calcium ion present neuron extended periods activity channel produces accommodation effect active neurons discussed further last section chapter general plays largely role neuron calcium present concentrations neuron larger huge concentrations neuron diffusion force positive internal potential back potential order due relatively large concentration differences involved note extra positive charges electrical potential acts strongly ion ions net charge difference leabra algorithm explicitly simulate important channels tend influence activation cell membrane potential cause things happen example nmda channel glutamate released excitatory neurons critical learning mechanisms described subsequent chapter accommodation effect opposite sensitization effect depend presence ions neuron measure neural activity ions enter neuron gated voltage channels presence indicates recent neural activity exist small amounts neuron concentration provides reasonable indication average level neural activity recent time period useful things happen neuron covered major ions channels neural processing now position put equation reflects neural integration information recall result equation updating membrane potential variable voltage membrane ohms law compute current ion channel add currents type ion channel now need know things equilibrium potential fraction total number channels ion open present time maximum conductance result channels open ions channels let current channel diffusion ohms law described total conductance fraction open times maximum conductance times net potential difference membrane potential present time equilibrium potential basic channels activation work neuron excitatory synaptic input channel inhibitory synaptic input channel leak channel total net current channels said current affects membrane potential movement charges decreases net charge difference causes potential place following equation updates membrane potential model based previous membrane potential net current time constant simulator typical value potential change capturing corresponding slowing change neuron primarily result cell membrane details particularly relevant here fact slow changes understanding behavior neurons useful think increasing membrane potential resulting positive current excitation shows according electricity increasing membrane potential results negative current match relationship potential current simply change sign current model add previous membrane potential course mathematically equivalent captures intuitive relationship potential current now refer order notation finally mathematical terms say current temporal rate change variable derivative membrane potential computed net current membrane potential updated excitatory inputs time conductance conductance equation provides means integrating inputs neuron show here different values conductances different ion channels well see moment actually compute conductances function inputs neuron equation point simulate response neuron fixed input providing values conductances shows graph net current membrane potential starting current rest potential responding different inputs come form value starting time step value constant inhibitory conductance run example important thing note figure membrane potential response excitatory input level dependent strength conductance excitation compared leak conductance conductances present membrane potential provides basis neurons subsequent output described tell stronger inputs put neuron threshold responding weaker clearly threshold sub interesting note fact value clearly corresponds change derivative back towards settles equilibrium value equilibrium point perfect balance forces system remains equilibrium membrane potential important phenomenon determining long present patterns network analysis presented showing relationship biology hypothesis testing basically reasonably short period time excitatory inhibitory channels opened effects dependent voltage channels subsequently activated membrane potential settle new stable value reflects new balance forces neuron new equilibrium membrane potential net current return zero individual currents particular channels zero non values add zero net current present changes taking place membrane potential expected current mathematical derivative membrane potential useful able compute value equilibrium membrane potential configuration conductances clearly equation computing value relevant equilibrium membrane potential equation equation computing membrane potential membrane potential function depends itself provides updating membrane potential tell directly value membrane potential settle provided constant input neuron easily solve equilibrium membrane potential equation noting equal zero value change obvious setting equation equal zero solving value appears places equation equilibrium value membrane potential solve value longer function time state now equation directly solve equilibrium membrane potential fixed set inputs see form equation understood terms hypothesis testing analysis detector benefit analysis equation shows membrane potential basically weighted average inputs conductances obvious follows membrane potential towards driving aka reversal equilibrium potential channel direct proportion fraction current channel total current example lets examine simple case excitation drives neuron towards membrane potential arbitrary units leak inhibition drive towards lets assume channels same conductance level total current excitation total neuron move way towards excitation reversal potentials normalized values maximum conductance values channels simulated leabra based biologically constants including resting potential firing threshold accommodation hysteresis currents discussed greater detail note value range based finally remaining issue regarding units values computing simulations typically normalized values range biological values weve here normalized values easier common axis meaningful related easily probability values see shows table basic parameters simulations biological normalized values performed minimum dividing range excitatory inhibitory channels open close function synaptic input coming neuron leak channels open turns way computing inhibitory input neuron described next chapter simplifies algorithm capturing main functional contribution inhibition need consider compute excitatory synaptic input point typical cortical neuron excitatory synaptic inputs come synaptic channels located dendritic tree large structure dendrites primary input region neuron synaptic inputs onto single neuron synaptic input individual channels typically neuron receives inputs number different brain areas different groups inputs called projections case inputs different projections different parts dendritic tree way compute excitatory input sensitive level projection structure allowing different projections different levels overall impact neuron allowing differences expected activity level different projections case models automatically important component excitatory input model comes bias input likely neurons individual differences leak current levels differences biology rise differences biases overall level differences computationally important allow neurons effectively different different sense threshold discussed next section activation strong inputs units active closely matching weight pattern require weak inputs units active closely matching weights types useful necessary solving particular tasks important include differences model adapt learning order problem hand see further discussion order keep implementation simple actually know biological mechanism responsible implement bias input way artificial neural network models additional bias input term input equation specifically introduce bias weight determines amount bias input modified learning weights network leabra algorithm levels involved computing overall excitatory conductance discussed critical understand details computation come back necessary general points understood point excitation single input product sending activation times weight value inputs combined averaging similar standard way computing input unit neural network simple sum individual inputs divide total number inputs obtain average sum result number values represent lot excitation values represent excitation rest details ways different projections combined levels computing excitatory synaptic input individual input synapse collection inputs same area dendritic tree averaged arbitrary scaling parameters bias input shown property active conductance equal bias weight scaled total number inputs equivalent input value sum scaled inputs including bias taking place shows levels computing excitatory inputs individual synaptic channel associated input collection inputs same projection bias input entire set inputs entire tree described basic computation involves determining fraction inputs active time projection level structure affects relative scaling set inputs compared determining fraction step compute fraction excitatory synaptic input channels open input detailed structure synapse computation biological level summarize number channels opened sending unit fires function weight overall synaptic efficacy strength connection receiver assume maximum possible level efficacy connection represent weight proportion maximum number case model individual spikes simply equal weight value whenever sending unit fires otherwise rate code activations see activation sending unit represents firing frequency weight value expected fraction channels open time single input next individual synaptic conductance come same input projection normalized following way overall fraction open channels projection computed simply dividing sum individual input conductance total number inputs required convenient introduce scaling factors affect value due activity constraints imposed inhibitory conductances brain area represented layer simulation models characteristic expected activity level expected proportion neurons active time level fraction total number neurons expressed vary significantly layer layer useful inputs factor differences due expected activity levels otherwise layer lower average activity contribute lower fraction active channels influence neuron cases significantly simplifies simulations automatically automatically differences computing input conductance fraction simply divide projections now apply mechanism determining relative strength different projections done relative scaling parameter simulator associated projection represented scaling parameters normalized sum projections neuron ensures excitatory conductance same range relative contributions changing function scaling net conductance projection index goes input projections neuron implementing projections relative scaling form scaling useful form absolute scaling different inputs well example see network performs absence particular input projection changing scaling inputs needs absolute scaling parameter affects input projection question accomplished absolute scaling parameter simulator directly projection input value typically set necessary scale input bias weight otherwise large impact relative synaptic inputs dividing total number input connections seems work well bias weight roughly same impact normal synaptic input impact bias weight modified absolute scaling parameter relative scaling parameter effect finally level projection conductances simply added bias input overall excitatory conductance neuron total excitatory input commonly called net input neuron computational models net input includes inhibitory inputs included here referred text additional aspect way net inputs computed reflects general feature dendritic processing averaging time net input term rapid cause network otherwise propagate information reliable fashion time averaging reflects way dendrites respond inputs due part fact relatively large membrane surface acts here averaging time implemented time constant simulator typically modified version input net equation follows aspects structure dendritic tree level projection organization time averaging point neuron model dendritic effectively single point considerable simplification effects detailed structure actual neuron currents generated dendrites cell body known including dendrites share basic properties way properties dendrites basic effects properties potentials generated synaptic currents experience time cell body further turns dendrites number active gated voltage channels input signals way eliminate included simple model active channels communicate output spikes produced cell body see next section back dendrites useful learning based overall activity postsynaptic neuron researchers emphasized complex interactions occur inputs dendritic excitatory inputs occur aspects dendrites effects well demonstrated actual neurons inconsistent keep simple model see discussion simple model described leabra algorithm allows scaling averaging time otherwise assumes synaptic inputs combine linear fashion addition phenomenon temporal mentioned particularly relevant understanding model basic idea here effective drive neuron inputs come roughly same time result relatively large proportion open excitatory channels large excitatory input contrast same number excitatory inputs spread longer time window excitatory input away leak current add produce large current typically code rate leabra essentially detailed timing issues associated neural spiking mechanism temporal present model time window temporal net input time averaging parameter prior section explained neuron integrates information different sources reducing single value membrane potential section neuron membrane potential results neurons evaluation process actually simple neuron applies threshold sending signal neurons membrane potential threshold remaining otherwise implemented explicitly algorithm describe biological mechanisms rise threshold describe nature signal form spikes produced biological neurons followed widely approximation terms code rate individual spikes represents expected rate spiking leabra model implement spikes code rate latter commonly important property output signal version underlying membrane potential emphasizes differences right threshold collapsing differences relatively strong potentials finally discuss details synapses provide sending receiving neuron good computational practical sense neuron apply threshold membrane potential communicating neurons same reason sense run people see fire time smoke detector sense neuron communicate fact detected conditions looking threshold ensures significant events communicated generally relatively rare consistent principle sparse activations discussed next chapter biological level takes resources neuron communicate neurons sense important events takes computational resources simulated neurons communicate threshold faster simulations threshold arises biologically action voltage gated channels mentioned previously gated voltage channels open membrane potential exceeds threshold value open allow ions neuron resulting further excitation gated voltage channels open complementary gated voltage channels open act neuron result spike activity action potential membrane potential rapidly goes comes back membrane potential back channels tends resting potential slightly causes period following spike unable fire spike membrane potential back threshold level again period means effectively fixed maximum rate neuron fire spikes illustration thresholded activations membrane potential exceeds threshold case activation occurs spiking activation function spike membrane potential reset code rate version zero coded rate activation value results according equation detailed mathematical equations derived describe way gated voltage channels open close function membrane potential provide detail need model modeling spikes simple threshold mechanism results positive activation value membrane potential exceeds threshold zero otherwise see activation value depend spiking code rate mechanism described illustration principal aspects output system including axon action potential spike separated spike combination active properties spike start axon place called axon large concentration relevant gated voltage channels value membrane potential point threshold applied determining spiking neuron spike communicated length axon combination different mechanisms explicitly simulated model describe active mechanism amounts reaction chain involving same kinds gated voltage channels distributed length axon spike start axon increase membrane potential bit further axon resulting same spiking process taking place think effect active mechanism relatively speaking relatively slow requires opening channels neurons sections axon propagate spike mechanism due properties similar present dendrites propagation faster purely electrical process suffers neuron way active spiking mechanism signal called order axon covered called typically concerned level timing issues level typically ever takes spike axon relevant biological details perspective model important result ability rapidly information large number neurons implementation spiking process leabra simple spike whenever membrane potential exceeds threshold membrane potential reset resting sub level subsequent time step parameter spike processed receiving neurons next time step order simulate extended temporally effects single spike postsynaptic neuron extended multiple cycles comparison code rate version described simulator simple averaged time version firing rate call code rate equivalent activation called simulator computed period updates cycles follows number spikes time period total number cycles gain factor value better range rate code activations result ensure range modeling individual spikes typically convenient model rate spikes particular level excitation consistent simple view detector neuron data fire integrate model neural processing see discussion order need function takes membrane potential present time expected firing rate associated membrane potential assuming remain constant spiking membrane potential reset reflects balance inputs neuron computed membrane potential update equation described previous section simple version spike mechanism main factor determines spiking rate time takes membrane potential return threshold level reset previous spike see unable write form closed expression time function non membrane potential simulations summarized reasonably accurately function plus form threshold value arbitrary gain parameter expression means positive component zero negative interestingly function same form compute itself similar bayesian interpretation terms comparing thresholded value constant null hypothesis represented number see written simply clear relationship function standard sigmoidal shaped logistic function typically abstract neural network models difference presence logistic sigmoidal logistic function applied directly net input membrane potential simplified model neuron purposes analysis case extensively studied important mathematical properties comparing function actual spiking need take account presence noise spiking model note simulated spiking neuron fire spikes completely regular constant input see inconsistent fact detailed spiking neurons random difference explained part timing spikes coming neuron inputs discrete randomly constant expect system containing spiking units exhibit appropriate take additional step ensure code rate function properly accounts presence noise adding noise membrane potential directly adding noise processing processing results need averaging large obtain reliable effects produce modified code rate function directly incorporates average effect noise result units reflect expected average effects noise illustration point new function line produced adding values normalized gaussian centered point times original function solid line noisy activation function threshold written showing effects gaussian noise case noise gain standard membrane potentials range averaging noise activation function done distributed gaussian noise function activation function illustrated simply amounts shaped gaussian noise function times neighborhood points surrounding including point activation function adding new value point new values reflect probabilities neighboring points point noise added result operation shown call new function plus noisy noisy function important effects noise shape activation function curve threshold gradually curves zero starting threshold point original function important neurons graded overall activation function advantages gradedness discussed note means activity associated threshold sub membrane potentials due noise sending threshold effect noise reduces gain activation function average spiking rate function equilibrium membrane potential threshold threshold written constant excitatory inhibitory conductances compared noisy function same conditions note due effects resulting time parameters standard values shown previous figure essential form function obviously well captured noisy function now position compare noisy function simulation actual spiking rate results seen shows good overall fit noisy activation function simulate average expected effects spiking neuron averaging relationship reasonable assume coded rate unit represents effects small spiking neurons context scaling issues discussed coded rate units models summarizing actual neurons spiking neurons detailed high resolution simulations illustrates important characteristics output function neuron hold rate code discrete spiking versions function roughly linear small potentials threshold potentials higher maximum value large membrane potential due spiking case increasingly effects period potential threshold reset following spike property sigmoidal shaped functions logistic similar kinds suggested neural spiking mechanism synaptic effects important consequences sigmoidal function emphasizes differences variable neuron threshold differences well threshold further gain parameter sensitive region threshold thresholded membrane potential spiking function appears relatively large gain factor unit highly sensitive small values threshold relatively differences larger values effects important put neurons networks diagram synapse action potential causes button causing bind presynaptic membrane release neurotransmitter postsynaptic receptors allow ions nmda cause postsynaptic chemical processes take place produced terminal via complete biological picture neuron describe main properties synapses here directly simulated model underlie important features synapse sending neurons axon receiving neurons see diagram excitatory cortical neurons synapse axon buttons sending side dendritic receiving side inhibitory interneurons typically own dendrites reason called neurons synapses directly onto dendrites excitatory neurons onto types neurons terminal buttons projections interestingly information happens neurons via synapse chemical process involving release neurotransmitter whereas essentially electrical action potential coming sending neuron terminal button opening gated voltage channels bring terminal possibly internal binding neurotransmitter membrane terminal process fully understood detailed level upon binding membrane release synaptic excitatory neurons release glutamate inhibitory neurons release released postsynaptic receptors receiving neuron causes open allow ions enter described results chemical processes postsynaptic neuron receptor provides primary excitatory input via ions nmda receptor important learning allowing ions enter trigger chemical processes lead learning glutamate receptor important learning activating chemical processes synapses inhibitory neurons receptors main provide main inhibitory input allowing ions enter excitatory channels opened resulting change postsynaptic membrane potential called excitatory postsynaptic potential similarly individual inhibitory inputs called order maintain release gets bound membrane later new new important via released taken back axon terminal allowed mechanisms keep opening activating receptors cause problems affect stages process including receptor activation postsynaptic chemical processes activated receptors important studying components complex biological system synapse number properties affect way behaves function prior activity commonly noted effect spikes coming reasonably rapid stronger presumably due presence membrane result prior release likely extended high firing synaptic resources resulting increased numbers release fails released spike contribute neural output function couple important features biology synapse emphasized number ways different components synapse affect overall efficacy strength information receiver net effect summarized weight neurons modification results learning main presynaptic components weight number released action potential amount assumed vary efficacy mechanism main postsynaptic factors include total number channels neurotransmitter presynaptic release efficacy individual channels researchers argued shape postsynaptic important impact conductance electrical signals synapse whole exactly factors modified learning matter considerable debate appears multiple pre postsynaptic side things important consequence synaptic biology type cortical neuron typically type neurotransmitter turn activates particular types postsynaptic receptors means note true cortical neurons types specialized subcortical neurons excitatory inhibitory same time neurons typically provide excitatory inputs inhibitory inputs neurons artificial neural network models separation excitation inhibition individual units communicate positive negative signals biologically implausible least models cortex leabra model explicitly incorporates division basic properties specialized inhibitory mechanism provides inhibition simulating way inhibitory interneurons work standard units communicate excitation summary point neuron activation function showing activation flows sending units weights resulting excitatory net input inputs including bias weight excitatory input combined inhibition leak compute membrane potential activation thresholded sigmoidal function membrane potential now covered major components computation level individual neuron including computation excitatory inputs weighted function sending unit activity integration excitatory inhibitory leak forces conductances thresholded activation output refer collection equations point neuron activation function provide fairly accurate actual neuron effectively single point space factors detailed neural major steps function summarized note order emphasize essential aspects function details scaling different input projections represented computation excitatory net input gain parameter noise shown activation function simulator variable actually represents now proceed explore properties function computer simulations now simulator explore properties individual neurons simulations directory see read introduction pdp simulation environment start following chapter leabra bring windows large windows graphlog displays shows simulation configured consists single input unit single receiving unit seeing here single input turned again response receiving unit look units response occurs convenient graphlog displays information time allows multiple variables viewed same time allows multiple runs different parameters compared point ahead iconify window addition graphlog windows see windows control panel contains parameters actions exploration take moment parameters click parameter view brief description control panel buttons bottom addition purpose special buttons vary simulation simulation sure read functions buttons remember buttons control panel menu project window select plots produced simulation shown excitatory leak currents operating simple neuron conductances reversal potentials currents shown control panel select activation function parameter keep default time press button control panel plot current parameters see lines plotted time steps cycles axis note standard normalized parameters default switch biological values later lets focus line red shows net input unit starts rapidly remaining time steps goes back again recall net input name total excitatory input neuron net simulation sending unit sends value later manipulate value control panel control magnitude net input default value timing input controlled parameters control panel total number cycles controlled excitatory input form line orange shows net current unit expected shows excitatory current input comes inhibitory input goes line plotted own special axis going shown orange lines share axis input net red note axis line coded color right variable buttons line yellow shows membrane potential starts resting potential increases excitation decreases back rest goes line green shows activation value coded rate activation function results membrane potential goes roughly button line graph precise value point back again note activation rise net input cycles due time takes membrane potential reach threshold verify occur right threshold value line blue turned default simply value code rate activation function come spiking units now parameters control panel explore properties point neuron activation function focus controls amount excitatory conductance simulation proportion excitatory input channels open input turned determines overall net input value general interested seeing unit membrane potential reflects balance different inputs coming here excitation leak output activation responds resulting membrane potential happens increase press buttons see effects decrease difference neural response changes magnitude away initial value important aspect point neuron activation function different runs top happen naturally want clear log press clear button clear graphlog reason graph goes blank somehow button things systematically parameter range point membrane potential threshold recall normalized units noisy activation function soft threshold switch setting control panel hard threshold exploring parameter places value unit threshold think better way finding value hint remember equation equilibrium membrane potential particular set inputs compute exact value excitatory input required reach threshold showing note leak channels open input inhibition present here ignored determined empirically value hint play value leak conductance button control default parameters see happens increase decrease leak conductance response unit change change question compute exact amount leak current necessary put membrane potential exactly threshold explain results terms relationship leak excitation equilibrium membrane potential equation now sense unit responds different currents computes resulting membrane potential reflects balance currents explore role reversal potentials happens change leak reversal potential sure pay attention aspects results time input active conclude relationship resting potential leak reversal potential happens change excitatory reversal potential changing value places value essentially same activation value default parameters same approach solve exact value default parameters show point good idea conductance reversal potential parameters influence resulting membrane potential demonstrate real difference behavior unit switch normalized reversal potential values based biologically ones click button switch biological values bring new graphlog display click now membrane potential plotted scale larger scale scaling gain activation function identical look simulation see exactly switch normalized biological parameters way unit behaves simulator controlled called unit particular part overall project via menu project window select appear select side side going parameters top bottom easily see different values sense differences click parameter explanation parameter values control panel previous exercises small subset full see locate section explore way unit computes activation output main objective understand relationship spiking code rate activation functions same project previous section press begin previous section know changing level excitatory input affect membrane potential resulting activation value lets explore relationship spiking activation function set press presentation period now see caused spiking mechanism membrane potential threshold activation spikes membrane potential reset resting sub potential reflecting spiking mechanism potential back process itself spikes firing rate hard graph useful click graph line plotted blue shows value code rate equivalent rate spike value function spike train see try changing effect spiking rate compared default closely resulting values end spike train compare activations computed function switch values tell good approximation rate spikes produced actual spiking function details spiking coded rate activations explained differences parameters shapes functions particular spiking function dependent setting membrane potential update rate parameter determines fast membrane potential reach threshold whereas noisy function care parameter membrane potential equilibrium value default parameters simulation set slow membrane potential updating activation dynamics easier see time results step activation function spiking model partially adding noise important aspect spiking real neurons timing spikes random overall rate firing remains obviously evident single constant input far results regular firing introduce noise adding small randomly generated numbers membrane potential see kind effect multiple spiking inputs coming cell now note additional noise plays similar role noise function noisy function case noisy function incorporates averaged effects noise here actually adding random numbers themselves behavior set variance noise control panel otherwise default parameters set tell spike timing random looking run happens multiple runs top look overlapping runs determine spike timing random regular hint try lower values noise compare exhibits greater variability measured number different possible multiple runs stimulus train detailed spike timing explain result interesting understand relatively small amount variability introduced membrane potential noise variance produce relatively amounts spike timing way see noise membrane potential membrane potential plots last spike noise membrane potential point constant value rest noise slightly small causes large effects random spike timing threshold string small membrane potential delay spiking string threshold effectively small differences large distinction small values membrane potential now lets return explore properties noisy code rate activation function compared possible functions well comparing noisy non version noisy linear function difference membrane potential output change excitatory input curves functions different noisy compared happens increase note activations fit range see curve stops gradually relationship noisy vary function level excitation activation obvious observed indication natural property functions present linear function functions approximate linear function lower levels excitation done simulation completely simulator window selecting project locate project window select explored basic equations point neuron activation function now explore basic function neuron detecting input patterns see particular pattern weights simulated neuron respond input patterns level neuron neuron respond pattern best fits weights graded manner patterns close weight pattern provides insight point neuron activation function works way previous simulation start following chapter leabra previous project open project window project find main control panel press button window comes enviroview window shows different events environment presented unit order measure detection responses patterns presented units later networks units contained environment environment real world provides external inputs individual pattern contained event represents distinct possible state environment see environment case digits represented simple grid pixels grid picture elements pixel event digit drive corresponding input unit provides synaptic input simulated neuron now press control panel bring window showing network containing input event activations unit receives inputs sure operation view pattern weights synaptic strengths receiving unit input idea unit detect press button lower hand left side window scroll bar find bottom list click receiving unit single unit input grid now see input grid pattern weight pattern receiving unit connections input units weight value displayed corresponding sending input unit weight pattern determine extent input patterns activate receiving unit see action present different input patterns determine unit responds weights locate control panel window upper right screen epoch process presentation events environment network sure operation process control panel press button process start event press button process event environment digit presents input pattern onto input layer network setting units directly corresponding values event pattern updates activations network equilibrium activations effectively reasonable reached units see process known settling viewing final results settling network window viewing weights network window see resulting activation values noted weights remain constant events need view activations network window select button lower left hand part window scroll back act top see pattern digit input layer clamped pattern event representing digit note receiving unit shows activity value getting lets proceed remaining digits observe responds inputs continue press digit presented seen receiving unit activated digit presented activation zero digits expected receiving unit acts detector pattern responses reflected shows activation unit function event digit number axis now lets try understand exactly unit responds key understand relationship pattern weights input pattern display weights current input pattern clicking button already selected sure select button hold key left button select receiving unit already selected case click click now see unit display divided left activation right weight value note activations provided environment maximum tell difference weights activations now epoch process followed digit report number input units weight input unit active easily display find variability numbers digits activation value receiving unit reflect variability better variable examine order view underlying variability now click variable graphlog window general relationship plot variable numbers computed previous question equations explain exactly net input computed results values plotted graph digit remember click line graph obtain exact values order work verify couple digits recall default case need know input layer projection set ignore scaling parameters set projection anyway bias weights ignored ignore averaging time looking values settling result working now detailed understanding net excitatory input neuron reflects degree match input pattern weights observed activation value ignore graded information present input signal now explore change information activation signal locate value leak current set sufficient strength excitatory inputs strongest best input patterns happens pattern receiving unit activity reduce value note hit button control panel apply new value run epoch process digits happens values explain effect changing terms point neuron activation function consequences different response patterns units output receiving unit try possible advantages higher lower values clearly important difference neuron inputs tradeoffs associated different levels brain kind problem neurons code input neurons high threshold low threshold types providing corresponding advantages specificity response bias weights important parameter determining behavior see next chapter value leak current partially inhibitory input plays important role providing dynamically level inhibition excitatory net input ensures neurons generally right range useful information neurons dependent neurons important consequences imagine explorations note section contains abstract mathematical ideas essential understanding subsequent depth perspective primary ways expressing computational level description via mathematical manipulations turns mathematical language probability statistics particularly appropriate describing behavior individual neurons detectors relevant parts language introduced here see provide interesting explanation basic form point neuron activation function described particular statistics relevant here hypothesis testing general idea hypothesis relevant data evidence want determine well hypothesis supported data provides language same basic operation detector performs data input processing performed detector hypothesis thing things detector detects likely present data identify important hypotheses detector hypothesis detected thing null hypothesis thing labeled want compare relative probabilities hypotheses true produce output reflects extent detection hypothesis null hypothesis current input result come detailed follows probability current input data written simple function functions relationship hypotheses data written here resulting probability function strong support detection hypothesis null hypothesis function familiar psychologists choice mathematical psychology models number years well explore functional form onto equilibrium membrane potential equation already see same weighted average quality way putting objective analysis important issues want evaluate extent believe hypotheses important possible actually measure objective probabilities hypothesis true particular set data typically impossible number reasons settle subjective definition probability refers objective fact case probabilities simply numbers means happens true means happens true intermediate values mean meaning happens time average getting intermediate probabilities correspond intermediate values value means true distinction fully time important ultimately concerned here valued real averaged time numbers consistent simple vertical line detector detects presence vertical line amounts inputs active input assumed inputs driven way visual signals vertical line visual inputs inputs tend light system noisy inputs active inactive possible states world vertical line detector frequency number times state occurs world hypothesis line exists world null hypothesis exist numbers inputs data frequencies show states inputs active likely hypothesis true versa vice bottom line contains total number states computing probabilities purposes hypothesis testing framework simple detector example shown detector receives inputs sources assume driven world vertical line present detectors likely activated hypothesis represented detector vertical line actually present world represent hypothesis variable hypothesis true null hypothesis vertical line present world represented states order compute objective based frequency probabilities opposed subjective probabilities example need table states world frequencies occurring shows table states define world purposes example state consists values variables world hypotheses data inputs frequency associated state indicates times state actually occurs world note states frequency meaning occur world simple able include possible states here note hypotheses mutually true same time essentially same table shown compute probability occurrence interested simply number times occurs table dividing total number states table compute probabilities needed hypothesis testing clear objective basis relevant equations inputs present table large due huge number different unique combinations input states example inputs binary actually true neurons worse table requires inputs hypotheses roughly inputs result typical inputs cortical neuron main reason need develop subjective ways computing probability terms applied understanding realistic neuron relevant probabilities computed table data basic probabilities interested computed directly world state table overall probability hypothesis true written short found finding states adding corresponding frequencies dividing result total frequency count states computation illustrated result next probability current input data receiving inputs present time compute need pick particular data state lets choose data case inputs active shows probability data state short occurs time hypothesis times true world finally need know times hypothesis true data present obviously relevant indicates predictive data hypothesis true called joint probability hypothesis data written data shows detector primarily interested predictive data hypothesis true gets inputs know clearly interested joint probability hypothesis data right information input data tend think hypothesis likely true data problem properly space computing probabilities joint probability tells occur compared possible states want know hypothesis true receive particular input data conditional probability hypothesis data written defined follows example data want know matching tells inputs active indicates chance level hypothesis vertical line present true basic information well correlated input data hypothesis comes joint probability critical restricting space events table considered computing probability subset entire table dividing now total subset table information appropriate context cases particular input data actually occurred basic equation want detector solve turns way subjective probabilities require computing kind conditional probability called likelihood likelihood opposite conditional probability conditional probability data hypothesis bit think computing probability data inputs experiment based hypothesis thing sure perfect sense think likely data based assumptions hypothesis words likelihood simply computes well data fits hypothesis comes same joint probability hypothesis data different way time scope cases hypothesis true determine fraction total particular input data state expect receive data time hypothesis true tells likely predict getting data hypothesis true straightforward world state table neurons explained find way computed way key step likelihood terms computed directly input data itself reference objective probabilities table effect idea easily measure hypothesis implemented neuron simply way inputs integrated weighted definition hypothesis place probabilities subjective frequencies events world mathematical manipulations probabilities consistent self ultimately computing likelihood integrating weighted input values inputs connected detector weight integrated resulting likelihood value shows likelihood value computed function integration different input values simple example input equal weights important model real neurons relationship objective based frequency probabilities complicated effectively ignored here setting total likelihood ends appropriately normalized sum total input weights activities inputs equation came place constructed equation produce same likelihood values compute world state table now good time verify equation produce same likelihood values cases computed world state table course likelihood equation consistent actual based frequency probabilities impossible didnt world state table exactly situation now considering kind simple assumption relationship input data weights resulting likelihood value typically assumption operation likelihood term particularly kind direct computation inputs relatively easy define likely particular input generated neuron configured particular set weights assuming likelihood function now need figure way write terms likelihood functions following steps take note definition likelihood new way expressing joint probability term appears back last equation known formula provides starting point whole field known bayesian statistics allows write called posterior bayesian terminology terms likelihood times prior called prior basically indicates likely hypothesis true seen data hypotheses plausible true reflected term favor simpler hypotheses likely necessary application here prior terms end constants actually measured least underlying biology terms normalized probability data turns turn term involving likelihood terms null hypothesis again want likelihood terms relatively simple equations hypothesis null hypothesis mutually hypotheses considering write probability data terms part hypothesis plus part null hypothesis amounts computing top bottom separately adding results overall result formula joint probabilities turned conditional probabilities simple conditional probability definition following formula resulting now expression easily terms hypotheses equation showed simple form reflects likelihood favor hypothesis form biological properties neuron implement general null likelihood serves strength likelihood term reasonable overall probability term world state table simple example null likelihood computed likelihood computation turns neural equivalent null likelihood constant try simple example function activity neurons communicated via inhibitory inputs computations frequency table nonetheless reasonable serve important computational roles described subsequent chapters now position compare equation equilibrium membrane potential hypothesis testing function developed reference equilibrium membrane potential equation here general idea excitatory input plays role likelihood support hypothesis inhibitory input leak current play role support null hypotheses considered null hypothesis analysis easy extend ignore leak current time inhibitory input play role null hypothesis now order compare biological equation hypothesis testing equation need appropriate values reversal potentials resulting membrane potential same range probabilities assume excitatory input drives potential towards inhibitory leak currents drive potential towards sense considering complete support excitation hypothesis result probability complete absence support excitation leak inhibition result probability values biological equation following relationship particular equations identical following excitation hypothesis neuron detecting inhibition null hypothesis assumed fraction channels open likelihood value excitation inhibition essentially assumed already computing likelihood function sending activations times weights baseline conductance levels represent prior probability values respectively provides satisfying level computational interpretation biological mechanism neuron integrating information way good statistical sense true actual values relevant biological parameters scale apparent relationship probabilities important thing form equation balance excitatory detection hypothesis inhibitory null hypothesis forces due form applied linear scaling values anyway finally full equation leak current reflecting case different independent null hypotheses represented inhibition leak see detail inhibition dynamic null hypothesis changes function activation units network leak constant null hypothesis sets basic minimum standard detection hypothesis compared far simple view neuron detector shown consistent biological properties mathematical description neural function based hypothesis testing detector model importance graded processing learning provide basis thinking neuron performs relatively simple task idea huge amount biological complexity present neuron amount information processing power neuron potentially exhibit example dendrites neuron potentially perform complex processing neural inputs including further sequence output spikes neuron potentially convey huge amount information varying timing spikes systematic ways individual neuron complex relatively simple detector number fundamental problems idea complex processing neural level learning requires graded response order bootstrap changes described introduction later neuron viewed performing lots discrete communicating via precise spike timing changes learning likely robust powerful learning algorithm neurons learning mechanism difficult organize networks neurons perform effectively brain robust noise damage example constant brain results significant movement neural undoubtedly kinds noise processing known sources noise due biological mechanisms neurotransmitter release addition substantial effects detailed firing properties individual neurons effect cognition graded catastrophic further known high levels damage neurons sustained effects cognition individual neuron contributing relatively simple way cognition neuron receives inputs neurons sends output signal computing detailed inputs attention detailed spike timing individual inputs complexity difficulty organizing large number further complexity detail somehow reduced single output signal end provide fraction total input neurons clear point complex processing end large number biological properties neurons consistent fire integrate model neuron detector hypothesis idea detailed spike timing important further neurons tend respond graded way noisy versions stimuli consistent detailed tend neural response sensitive specific appears finally bottom line able model wide range cognitive phenomena simple style detector neurons additional complexity appear necessary point note section mechanisms described applicable limited range phenomena active default simulations reader come back later find need mechanisms addition integration evaluation functions basic point neuron activation function neurons complex activation dynamics enable neuron way responds subsequent inputs function prior activation history thought form regulation self neurons response simulations explore later complex dynamics enter picture operate longer time typically typically interested initial activation state produced response input pattern complex longer term aspects response simulations ignore additional dynamics assume additional dynamics present simulations run settling long observe effects forms regulation self known accommodation hysteresis accommodation causes neuron active active same amount excitatory input hysteresis causes neuron active remain active period time excitatory input forces otherwise resolved hysteresis operate time period neurons active term short tendency remain active hysteresis active longer accommodation term longer accommodation results tendency network switch different interpretation input pattern see subsequent explorations addition hysteresis accommodation dynamics term longer threshold adaptation mechanism ensures neuron constantly active active neuron constantly active value activation threshold increased likely able remain active conversely threshold goes neuron active time threshold adaptation active neuron essentially accommodation adaptation active neuron hysteresis form sensitization active neuron gets increasingly inputs likely active details processes following sections potential biological accommodation hysteresis accommodation arise effects potassium channels activated response membrane potentials voltage gated channels increased concentrations calcium ions neuron result sustained activation gated channels increased current effectively larger leak current membrane potential strongly back towards rest accommodation arise lasting long effects certain types inhibitory channels hysteresis arise gated voltage channels open neural membrane potential sustained level recall results excitation neuron attempt include detailed biological mechanisms model adopt simple understand easy implementation dynamics same basic equations different parameters accommodation hysteresis delayed effects accommodation hysteresis accomplished basis variable represents activation pressure relevant channels basis variables updated function activation value neuron according following function basis value time average activation state time constant difference accommodation hysteresis time constant faster hysteresis typically accommodation typically actual activation channel function basis variable basis variable gets activation threshold value conductance channel begins increase different time constant basis lower threshold value conductance decreases again finally conductances accommodation hysteresis added conductances point neuron equations contribute overall current experienced neuron see accommodation conductance computed similarly hysteresis conductance computed same equation different parameters default values reversal potentials channels shown open project directory looks pretty simulation explored earlier plots variables basis conductance values accommodation hysteresis contains parameters control self channels control panel locate overall control panel note bottom parameters stimuli press button observe activation result accommodation hysteresis turned expect cycles lasting note particular apparent effect prior activation response latter now lets turn accommodation click button control panel graph log window click variable accommodation net conductance accommodation displayed graph press increases unit active reach activation threshold set see field control panel unit needs remain active bit longer order basis variable point activation way achieve set time stimulus later set press again now observe accommodation conductance starts increase stimulus goes unit inactive takes basis variable decrease threshold field takes long time next input stimulus comes cycles strong accommodation current unit activated input cycles basis variable finally goes threshold starts decrease unit active unit network units received input stimulus activated stimulus units active immediately produce different representation input accommodation provides means network respond subsequent inputs based prior activity point explore parameters play determines strong overall accommodation current try see weak completely unit done exploring press button return default parameters now lets add hysteresis click buttons graph log click button field see unit remains active time input stimulus goes back cycles additional period activation causes accommodation current activated eventually turns unit result accommodation saw unit activated immediately input now play parameters see units response properties summary see considerable potential complex dynamics emerge interactions different channels fact actual neurons channels complex dynamics suggests basic point neuron model lot simpler real thing evolution time dynamics considerable typically ignore complexity difference aspects behavior modeling simulations later text away changing balance excitatory inhibitory currents coming neuron accommodation hysteresis change threshold activation threshold adaptation useful units participate overall representational scheme network network algorithms known well model depend large degree unit active roughly same percentage time leabra model brain depend strongly idea basic idea force units represent different aspects input environment force active roughly same number different input patterns unit tend focus different subset patterns whole space covered shown powerful computational idea important limitations assumes inputs different relevant types distributed time likely true real world basically type model predict went novel environment ability represent back units active represent present current environment order amount activity view threshold adaptation upper lower activation frequency precise target value addition themselves wide strength typically weak implementation threshold adaptation depends average running activation value time constant computing average typically small average takes account activity long time period equation applied end settling applicable activations different input patterns average activity value unit exceeds wide lower upper units activation threshold appears constant modified maximum threshold value typically unit active level typically minimum threshold value typically unit active lower typically average activation threshold back standard value typically rate threshold typically actual simulations specify single parameter simulator standard threshold open project threshold adaptation process control panel press start due excitatory input activation average orange line increases value point threshold yellow line increases units activity point average threshold again slightly allowing unit active again repeating cycle continue result units average activity substantially now play parameters see kind effects threshold adaptation process biological functional properties neuron consistent detector constantly information available looking conditions match specialized detect view neural function different standard serial computer serves basis comparison whereas standard computers relatively general purpose computational neuron relatively specialized dedicated detecting particular set things refer things neuron detects representation emphasize individual neurons representations typically difficult describe simple verbal terms smoke smoke detector neurons exist huge numbers operate parallel whereas standard computer operates serial performing operation time good reasons think neurons perform relatively simple computation arguments mathematical description neuron detector framework bayesian statistical hypothesis testing produces same form mathematical activation function actually neurons learning shapes neurons detect different things plays critical role sure detectors working parallel accomplish sensible particular task solve sum detector model neuron provides good intuitive model function help sense underlying biological properties detectors neurons need receive combine information number different input sources relevant biological consists channels synapses contain particular types channels allow atoms ions neuron different ions different concentrations neuron leads generation electrical current channels open allow ions flow concentration cell neurons excited ions enter cell synaptic channels receiving areas called dendrites synaptic channels opened neurotransmitter known glutamate released sending presynaptic neuron different inputs provide different amounts activation depending neurotransmitter released channels postsynaptic receiving neuron open result different synaptic synaptic strengths different inputs refer weights critical determining neuron detects contrast excitation neurons ions enter neuron channels opened neurotransmitter released inhibitory interneurons basic negative current caused positive ions potassium leaving neuron via leak channels open simple equation way overall electrical voltage cell known membrane potential integrates currents valued real number equation same form derived principles based computational level detector model neuron point neuron activation function leabra algorithm detector integrated information evaluate evidence strong conclude detected communicate care membrane potential results integration turn determines neuron produce action potential spike causes neurotransmitter released ends neurons sending projection axon form synapses onto neurons dendrites see action potential thresholded meaning occurs membrane potential start axon called axon gets certain critical value called threshold means neurons communicate detected level biologically computationally leabra simulate individual action potentials typically rate code valued real number represents frequency rate cell produce spikes based membrane potential spiking rate units activation value called activity output unit useful small simulations spike timing causes problems present larger networks averaging number units reduce impact noise rate code intended represent scaling assumption individual units model represent number roughly similar neurons average impact spiking neurons rate code leabra spike rate function initially roughly linear threshold approaches maximum firing rate biologically determined number factors including period spike essentially impossible fire rate synapses release neurotransmitter activation function important computational power neural networks producing stable activation states interactive networks neurons number self channels affect way neuron responds based prior history activity different mechanisms included leabra essential aspects algorithm simulations accommodation causes neuron active active same amount excitatory input hysteresis causes neuron active remain active period time excitatory input obviously accommodation typically operates time period finally term longer threshold adaptation mechanism implemented way ensure neuron constantly active active way active neurons similar simple accommodation mechanism capacity neuron sensitive sensitization active recently neuron neuron provides basic unit processing network neurons required accomplish simple tasks able describe essential computation performed neuron terms detector model simple computational metaphor applies computation performed entire network adopt approach understanding networks work chapter identify explore important principles general behavior networks next chapter show learning responsible setting detailed weight values specify unit detects shape behavior networks according principles build upon developed chapter begin summary general structure patterns connectivity cortex neocortex biological basis general types networks cortical areas terms neuron types general patterns connectivity rise generic cortical network structure modeling kinds different psychological phenomena explained previous chapter excitation inhibition separated cortex implemented different types neurons different patterns connectivity separation useful understanding basic principles network function unidirectional feedforward processing information via excitatory interactions performs information processing transformations essential cognition bidirectional connectivity advantages unidirectional connectivity cortex requires inhibition control positive excitatory feedback effects cortical inhibition summarized inhibitory function explore functional case excitation context inhibition end clear forms interaction separable overall network behavior possible summarize overall effects interactions terms constraint satisfaction networks achieve state activation simultaneously satisfaction external constraints environment internal constraints patterns weights connecting neurons cortex neocortex forms part brain humans relative data idea cognition takes place important remember cortex depends critically subcortical brain areas proper functioning cortex divided number different cortical areas specialized different kinds processing areas critical recognizing objects appear process spatial information areas perform language processing higher level planning etc able single algorithm writing book large part due fact cortex fairly consistent general structure applies different cortical areas specialized processing takes place similar general network structure simulated common computational framework properties structure section detailed description biological properties cortex entire details considerably here general neurons identified cortex excitatory neurons release excitatory neurotransmitter glutamate inhibitory neurons release inhibitory neurotransmitter see details primary excitatory neurons neurons larger number different inhibitory neurons excitatory neurons constitute roughly total number neurons cortex apparently responsible information flow form networks range longer projections different areas cortex subcortical areas following discussion connectivity focused excitatory neurons contrast inhibitory neurons project small areas cortex consistent role providing local inhibitory feedback mechanism see cortical neurons organized distinct layers cortex neuron types found layers layer place neurons found neurons typically found layers cortex identified anatomical important understanding detailed biology cortex purposes picture considering functional layers input hidden output layers information typically order see illustration term layer refer functional layers term cortical layer based biologically layers input layer corresponds cortical layer receives sensory input way subcortical brain area called thalamus information retina sense output layer corresponds cortical layers sends motor outputs wide range subcortical areas including basal ganglia simple layer interpretation cortical structure consistent general connectivity patterns provides useful starting point modeling direct excitatory connectivity shown open connections inhibitory interneurons indicated connections operate cortical layer receive same types excitatory connections excitatory neurons lines indicate connections exist consistent flow information input hidden output limited data difficult determine important connections transformation input output mediated hidden layer directly visible connected cortical non input output areas functional layer corresponds upper cortical layers cortical layer largely hidden layer thought internal model environment provides useful typically elaborated processed basis driving outputs network function inputs example hidden layer internal model represents according categories red orange yellow green etc continuous representation provided visual sensory inputs categories relevant networks output color words provide useful basis outputs raw input itself goes cognition thought terms elaborate internal models next chapter find models developed learning useful essentially same functional cortical layers suggested long supported different types data anatomical connectivity different areas suggests information coming input layer next primarily hidden layer output layer see addition firing properties neurons different cortical layers shows hidden output layer neurons complex responses input layers view cortex presented considerably simplified order focus distinctive structure reality cortical areas process sensory input same ones produce motor output large number areas direct sensory input direct motor output presents accurate picture structure main different kinds cortical areas correspond functional layer types input hidden output different types areas emphasizes corresponding functional layers input areas well developed cortical layer including different layers sub layer primary visual input area output areas well developed output layers hidden areas reduced input output layers explore ideas function later chapters larger scale version cortical showing different types cortical areas input area developed well layer receiving sensory input thalamus producing motor output directly hidden area called level higher association area receives input input areas hidden areas sends outputs output areas hidden areas reduced input output layers primarily via layer connectivity output area motor control area subcortical brain areas drive motor system real layer larger layer lines indicate layers connections reduced importance area lines again represent connections exist consistent output hidden input model information flow picture view cortical structure sensible information comes network specialized areas layers processed potentially long sequence internal processing areas hidden layers results output drive motor system add important complexity otherwise simple picture excitatory connections bidirectional cortex cortical layer number inhibitory neurons function discussed importance features clear finally possible role connectivity thalamus subcortical structures hidden areas simple view discussed later chapters previous section showed cortex excitatory interactions clearly primary importance section explore basic types computations networks neurons excitatory interactions perform start simpler case unidirectional feedforward connections rare cortex basic computations generalize bidirectionally connected case certain brain areas play important role relevant hippocampus see transformations possible type connectivity basic cognition digit network input layer digit images hidden layer units represent different digits already covered computation performed unidirectional information forward direction feedforward network discussed role neuron detector explored simulations basically feedforward connectivity allows sequences detectors process input signal detecting different patterns function weight values connections case explored input signal consisting images different digits grid unit detects extent input digit looks imagine multiple units layer hidden layer containing detectors different digits image digit presented unit represents digit activated digit network discussion explorations order understand nature transformation performed hidden units digit network layers detector units generally need develop terminology describing differences input representations hidden representations input layer activated pattern looks natural assume input layer represents digit sense representation specific digit same input layer different activity pattern represent digit image entirely hidden unit set weights specifically configured detect pattern seems different type representation provided input difference specificity critical understanding benefits detectors hidden layer specifically say unidirectional flow information digit image input layer hidden layer results transformation image specific representation hidden layer emphasizes distinctions images different digits distinct categories same time collapsing differences category treating images same digit same imagine whole subset input patterns correspond digit digits hidden unit representing digit groups images same category images important try say way transformations emphasize distinctions distinctions relevant items represented distinctions digits case digit network emphasized distinctions relevant distinctions different instances same digit pretty easy effects transformation performed single unit plot response unit input patterns exercises considerably difficult entire hidden layer units useful cluster plot similarities patterns activity hidden layer cluster plot groups patterns groups patterns similar similarity typically measured distance recall computed distance points measured dimensions easily generalized number dimensions units present hidden layer adding squared difference terms dimension cluster plot based distance matrix shows patterns visual form cluster plot similarity information easy see example clusters plot similarity information captured underlying representations problem cluster plot similarity relationships different patterns indirect measure transformations performed individual units necessary reduce high hidden unit representation somehow meaning information bit time learning certain features cluster plot arise underlying properties individual unit transformations cluster plots feedforward transformation digit images digit categories shows cluster plot input digit images shows cluster plot hidden layer digit detectors shown images complex patterns digit similarity separated digit category labels specifically represent digit note items cluster example cluster plot simple case digit network hidden unit exactly representing digit shown compared cluster plot input images themselves shown note axis plots index different patterns axis shows distance distance items cluster length horizontal line coming common vertical line axis note axis scaled sure look actual values assume different plots same length particular plot hidden units distance away squared sum distance put single cluster contrast complex pattern similarities input layer reflects amounts overlap shared pixels images sketch summarizing results previous figure showing transformation overlapping patterns digit representations input layer specific categorical digit representations hidden layer circle input represents collection pixels digit image overlap indicating similarity produced shared pixels hidden layer representations overlap digits clearly specifically represented order interpret plots terms transformation input patterns need clear idea transformation accomplish case lets say want transformation digit completely separate distinct put way want hidden unit specifically represent single digit digits similarity relationships digits want essentially similarity structure digits representation equally distinct exactly see course actually kind transformation want least fairly transformation input patterns take issue types transformations generally better following section sketch transformation shown intended roughly capture similarity structure shown cluster plots function overlap representing individual digit images accurately impossible dimensions cluster plots digits different images digit shows cluster plot noisy digit images shows cluster plot hidden layer digit detectors shown images note length zero lines digit clusters sub here indicates exactly same pattern zero distance input represents image distinct pattern digit hidden layer differences digit category distinctions categories transformations defined emphasize collapse important benefit type representation shown previous able tell different images digit somehow equivalent beneficial digit looking purposes mathematical details color size particular digit example shown different versions digit image original noisy versions input pattern image same digit different digit different hidden layer versions same digit length zero horizontal lines labels right vertical cluster line cluster plot indicate cluster identical identical distinct versions different digits now simple way different versions digit cluster seems pretty transformation clusters simple example kinds transformations actually useful adding bit complexity steps transformations result powerful processing cluster plots letters digit network shows cluster plot input letter images shows cluster plot hidden layer digit detectors shown images hidden layer units activated letter inputs representations overlap meaning hidden layer distinction patterns good example specificity transformation tendency collapse irrelevant distinctions seen presenting images entirely different things digit network example presented images letters network expect sensible hidden layer representation contrast based image input representation simple grid perfectly capable representing letters clear categorical distinctions illustrates phenomenon showing digit hidden layer letters case collapsing happens hidden units specifically configured detect digits simply respond letters exception hidden unit digit responding sufficiently similar letter resulting single cluster hidden unit activity things done problem units respond easily distinctions remain due fact weights specifically tuned letter patterns distinguish process patterns activation central cognition view cortical networks sensory information ways produce specific representations relevant distinctions necessary survival collapsing irrelevant ones aspect unidirectional connectivity bidirectional case see later now explore ideas presented start project chapter leabra see network digits network window called display activity states network time panel visible addition standard pdp project windows see information lets examine network looks layer digit images layer hidden units representing digit select button lower left window need scroll click different hidden units see weights exactly match images digits units represent units referred matching units respond proportion match input specific weight weights developed learning mechanisms described typically specific single patterns resulting distributed specific hidden unit representations individual units participate representation multiple different input patterns important issue next section pretty obvious unit respond digit input patterns lets explore nonetheless lets view unit activities network selecting button next locate process control panel overall control panel hit button followed button presents input pattern network updates activations network equilibrium activations effectively reached units see proceed activities net inputs hidden units digits now display press button control panel select run digits essentially same pressing process control panel allows select input patterns present observed unit activated matching digit presented digits presented continue important understand role bias weights simulation digit images input patterns different numbers active units observed detector exercise previous chapter press control panel select different net input activation levels corresponding hidden units activations shown appear roughly similar here bias weights differences overall activity level coming inputs see select window see pattern different values see effect press hit affect hidden unit activities now activations consistent number active units input patterns turn bias weights back window isnt already selected explain bias weights contribute producing originally observed hidden unit activities run network biases again now cluster plots shown hit button window containing cluster plot similarity relationships digit images shown compare amount overlap activated pixels digit images click control panel select cluster plot results iconify events window done window iconify button window now again selecting time cluster plot shown note values greater hidden unit activation values purposes small differences activation values units otherwise reflected cluster plot present purposes interested binary patterns activation different units detailed differences activation units putting cluster plots context underlying unit activities discussion previous section concrete reading point useful next step try running case multiple instances digit call case run selecting press see appropriate hidden unit active version digits small levels activity units observed cases compare different digit images same digit noisy version digit additional units active now explore effects selectivity units behavior control panel shows leak conductance hidden units set value case detector example changing parameter affect specificity units responses happens generally hidden activations reduce value inputs affect cluster plot hidden unit activities goal network same hidden representation version same digit different representations different digits changing specificity units responses affect network set leak conductance back now see network responds letter inputs digits press button control panel select layer network controls top scroll display back start letter top display single button good choice fine visible response came hidden unit letter input similar press end forward fast button grid log continue scroll next press pressing inputs now see digit units respond letter stimuli based experiences previous question expect happen cluster plot hidden responses letter inputs leak current value right say hidden representation good letter identity information find setting letter information clear digits example explored highly specific matching hidden representations provided useful simplicity particularly realistic powerful type representation general term representations unit represents input pattern local localist referred cell representations brain neuron uniquely represents ones detector model neuron associated type representation clear detector function apply complex difficult describe local non representations alternative localist representations distributed representations individual units participate representation multiple input patterns addition multiple units typically active input pattern saw good examples types representations leak current explorations previous section useful think distributed units representing features input patterns whole input pattern composed number features particular feature pattern encoded pattern active units known based feature representation course notion detectors case units need clearly defined easily described representations distributed representation effective case coarse coded representations continuous dimensions see features arbitrary underlying dimension unit graded response distributed representations general tend specific localist representations units participate representing multiple items features represented units serve enhance distinctions collapse localist case difficult see distributed case main advantages distributed representations localist ones follows total units required represent number input patterns representation shared units otherwise unit pattern required distributed representations provide natural means encoding similarity relationships different patterns function number units common pattern overlap network distributed representations respond appropriately novel input patterns appropriate novel combinations hidden units impossible localist networks require entirely new unit multiple units representation robust damage representing continuous dimensions distributed coarse coded representations accurate equivalent number localist representations lot information contained relative activities set units whereas localist units represent different values units distributed representations allow bootstrapping small changes critical learning see clear distributed representations critical key properties neural networks described chapter researchers maintain localist representations preferred part simplicity useful previous section remaining models book distributed representations discuss important difference sparse distributed representations generic distributed representations finally evidence cortex distributed representations researchers presented visual stimuli systematically number different dimensions size shape color etc activities neurons areas cortex process visual inputs cases aware shown cortical neurons exhibit relatively broad tuning curves means respond stimuli range different parameter values consistent coarse coding stimulus dimensions kind abstract hard based feature representation level higher based object representations neurons early visual system forms evidence suggest encode oriented bars light see again consistent based feature representation now explore difference localist distributed representations start project chapter leabra open project window select project looks similar previous containing network grid log press same localist network now pick see new network appear same place distributed network contains hidden units lets explore network examining weights hidden units select notice units configured detect parts features digit images entire digits case localist network imagine units active whenever features present input test idea now verify firing patterns hidden units sense features present different digits case hidden unit firing digit fires left right match weight pattern actually important encoded hidden unit middle horizontal line detector actually serve multiple roles simple case kind complexity attempt describe content detected neurons imagine weights complicated pattern values start feel complicated neurons responses cluster plot distributed networks hidden unit representations compare cluster plot input patterns cluster plot localist hidden units terms visual similarity digits representations capture hint remember clusters plot similarity information captured underlying representations explain consistent general level specificity distributed representations relative localist ones account differences distributed hidden units compared input patterns select test distributed network letter inputs cluster plot resulting hidden units compare cluster plot letter inputs distributed network provide good representation letters think case relate answer specificity distributed representations previous question distributed network achieves useful representation digits number hidden units localist network number hidden units number input units greatly input representation explain achieved binary representation units active pattern number units required represent digits ways bidirectional excitatory connectivity behaves same way unidirectional feedforward connectivity explored transformation goes ways addition able produce digit category image digit feedforward networks explored bidirectionally connected network produce image digit digit category produce image based digit involves processing going top bottom explored lateral connectivity connections units same layer part pattern activate parts process goes name pattern completion pattern full processing part closely related phenomenon mutual support lateral connectivity activations interconnected units strongly activated phenomena described general term dynamics networks activations appear particular final configuration range initial activity patterns range initial lead same final pattern called explore cases proceed amplification properties bidirectional connectivity lets begin exploring bidirectional case same kind transformations unidirectional open project previous digit networks same windows etc main control panel called examine network note connecting layers going coming back indicates bidirectional connectivity view connectivity selecting button network window click hidden units see familiar digit image now click individual units input image see receive hidden units well note different input units receive different numbers sending units reflects different types pattern overlap digit images easier way see happen activated hidden unit view sending weights selecting sending weights look receiving ones appropriate digit image verify fact weights click notice difference display remember weights important need determine level symmetry future interesting thing weights see unit activates same things activate kind consistency things activated bidirectional networks important see later now run network see weights produce replicate previous feedforward results press select present images input layer digits notice network displays activations settling process changing equilibrium come later note grid log shows input hidden activation states now results run exactly same feedforward network weights additional effect here input units clamped pattern event actually computing activations otherwise attention weights pressing run network digit category units hidden layer digit images input resulting input patterns driven top weights units saw dynamics activation updating network window settling pattern otherwise difficult tell difference runs produce basically same patterns activity layers tell hidden units clamped same activation value whereas value input images presented images slightly lower activity value driven digit category units simple exercise demonstrates bidirectional connectivity enables information flow transformations computed bottom top number important issues surrounding phenomenon different digit images correspond same categorical hidden unit exploring bit happens multiple combinations hidden units active press enviroview event pattern displayed pattern contains input presented hidden layer units moment lets run describe happens input layer activations digit categories activated sure note subtle differences activation account result change value enhance differences activation present explain helps kind enhancement differences generally useful cognition now click left button digit categories input way select combinations digit categories activate sure press changes take effect trying different combinations now explore pattern completion network bidirectional connections single layer units top bottom processing network exhibits lateral processing difference same underlying processing mechanisms units involved somehow whereas top bottom hierarchical relationship open project addition usual windows see network window single network environment enviroview window event click left button single event display pattern window clicking pattern change network complete usual begin examining weights network select click different units see units image digit interconnected weights value units weights appear presenting part image result activation remaining parts test press button main control panel presents input units determined event pattern shown view window activated notice viewing activations updated settling tell ones clamped environment result fundamental difference pattern completion phenomenon types excitatory processing examined far simply result units detecting pattern activity amongst activated pattern sufficiently matches pattern encoded weights pattern completion particularly useful thinking recall memory cue distinctive image etc triggers ones memory results recall related event discussed further note previous simulations called soft inputs event pattern presented additional excitatory input neurons directly setting activations corresponding values latter form previously called hard results faster processing soft soft necessary case units layer need update activation values function weights order produce pattern completion pattern weights think number units need clamped order produce pattern completion full now test answer units event pattern network longer produces complete pattern press view clicking units necessary completion parameter lower number value parameter allows completion input active happens activate inputs part pattern weights layer configured support representation pattern think difference similar new pattern pattern phenomenon closely related pattern completion mutual support happens activity set units excitatory produces extra activation units amount extra activation provides useful indication strongly interconnected units enables better interference noise inhibition see effect simulation clicking entire pattern input network unit part set leak current hit note units part interconnected pattern experience mutual support able overcome relatively strong level leak current unit weights suffers significantly finally note preceding simulations highly simplified basic phenomena underlying mechanisms clear see start learning algorithms networks complex deal large number patterns encoded same units resulting behavior powerful difficult understand detail same basic principles work examples involving bidirectional connectivity controlled avoid problems bidirectional connectivity positive feedback positive feedback rise useful phenomenon amplification excitatory signals neurons result activation strengths amplification critical explaining aspects cognition including word effect discussed introduction later positive feedback problem results spread excitation entire network producing effectively unit activated inhibitory mechanisms described next section necessary take full advantage bidirectional amplification consequences section explore couple simple amplification action see lead ability bootstrap weak initial activation fully active pattern see cause spread activation network begin exploring simple case top amplification weak bottom input via bidirectional excitatory connections open project see network layers input bidirectionally connected hidden layers unit layer graph right plot activations layer hidden units time network settles response input bootstrapping phenomenon activation hidden unit strong start activating hidden comes back hidden producing strong activation press control panel see graph window here activation coming top unit relatively weak initial activation unit resulting strong activation units example bootstrapping unit activate unit place receive additional top excitation increase strength leak current observe resulting activation now activate bootstrapping amplification occurs decrease leak current bottom pattern relatively strong anyway bootstrapping amplification simple case bootstrapping amplification provides insight level word information presumably higher cortical processing stream level letter information come back bootstrap activation corresponding level letter stimuli revisit example detail note example organized according bottom top processing same principles apply lateral connectivity simply move unit same layer parts interconnected pattern activated previous example illustrated benefits bidirectional excitatory amplification bootstrapping occur distributed representations amplification lead activation units particular overlapping connections required implement distributed representations allow excitation spread see spread inhibition resort increasing leak current prevent activation spread problem here saw example leak current impossible bootstrap representations activity benefits bidirectional excitatory amplification available example provides strong motivation next section inhibitory interactions open project network here previous example now multiple units layer lets examine connectivity button click units notice receive corresponding input input units called connectivity notice left right units receive uniquely left right units now click left right units observe connectivity symmetric left unit receives left center right units connectivity pattern representing separable features input objects consisting features labeled simulation think object features feature overlap objects shared feature causing locate control panel press unique object see network settle looking grid log right showing activations comes back activate feature good example pattern completion phenomenon top activation lateral activation fails presented non feature active default leak current level try setting leak current see center hidden feature active describe happened explain happened terms connectivity network try parameter space increments see find level strong activation activate ambiguous center input feature pressing increments strong activation common feature weak activation features explain results finally set leak current press button try notice full pattern presented network activates level object units active find value reasonable solution observed explorations bidirectional excitatory connectivity interesting amplification pattern completion processing easily away further see networks tend strongly respect small parameter changes activated area important property networks later chapters network behavior supported present point neuron activation function see particular property sigmoidal plus function provides necessary upper positive feedback gain parameter changes threshold value contributes character units simulation far leak current played central role determining networks behavior leak current far excitatory input coming neurons consistent role described seen likelihood null hypothesis bayesian hypothesis testing framework problem situation leak current constant easily respond dynamic changes activation network manipulate role dynamic excitatory input played inhibitory inputs neuron produced inhibitory interneurons described neurons sample general level activation network provide dynamically amount inhibition based activation level think general role inhibitory neurons controlled air prevents network getting active sample air inhibitory neurons sample activity network inhibitory neurons detect network getting active produce inhibition increased activity turn gets conversely detect activity provide inhibition principle necessary inhibition same kind set point behavior roughly same activity level see provides convenient reasonable approximation effects inhibition explorations see effects different parameters understood analogy basic types inhibitory connectivity excitation shown open connections inhibition ones shows feedforward inhibition driven input layer activity shows feedback inhibition driven same layer note inhibitory interneurons typically themselves well forms connectivity involving inhibitory neurons cortex rise feedforward feedback inhibition see example feedforward inhibition inhibitory interneurons driven directly input layer inhibition hidden layer neurons hidden layer neurons receive amount inhibition function level activity input layer projects excitatory connections hidden layer feedback inhibition occurs same layer inhibitory neurons producing negative feedback note inhibitory neurons themselves providing negative feedback control own activity levels turn important related ways forms inhibitory connectivity act control excitation feedforward inhibition excitation coming layer layers take account provide result feedforward inhibition hidden layer excitatory neuron receive roughly amounts excitation inhibition think prevent neurons ever getting active place acts kind neurons particularly strong excitatory weights current input pattern able overcome feedforward inhibition comparison feedback inhibition level excitation layer itself provides negative feedback prevents excitation units observed previous section standard same air see following explorations types inhibition necessary order speed simulations summarize effects inhibitory interneurons computing inhibition function directly function amount excitation layer need explicitly simulate inhibitory interneurons themselves simplest effective inhibition functions forms take kwta function described see following explorations combined effects feedforward feedback inhibition rough set point behavior overall activity level layer levels activation set point back set point value characteristic kwta functions implements directly setting target set point units total active case function prevent network getting active result function allows network active units active isnt excitation coming network discuss particularly useful functional properties inhibition important general functional consequences inhibition inhibition leads form competition neurons case feedforward inhibition strongly activated neurons able overcome inhibition case feedback inhibition neurons active better able inhibitory feedback activity contributes inhibition neurons competition thing network provides mechanism selection finding appropriate representations current input pattern selection process natural selection based competition natural resources results evolution itself selection process network occurs moment moment line basis longer time periods interaction learning mechanisms described next chapter learning context competition produces evolution representations value competition long recognized artificial neural network models showed feedforward inhibition result form feedforward pattern completion return finally tried strong mapping selection takes place neural network rely basic competition learning mechanisms understanding process way viewing effects inhibition terms sparse distributed representations produced appropriate levels inhibition particularly kind inhibition produced kwta inhibition function representations distributed level inhibition multiple units active time sparse inhibition strong prevent relatively small percentage units active according parameter kwta sparseness understood terms specificity underlying representations explorations difficult unit active specific unit fires typically means fires specific obviously extreme case sparse representation localist representation sparse distributed representations intermediate fully distributed localist representations represent advantages types representations emphasized sparse distributed representations particularly useful representing things world simple form argument goes follows things world generally share number underlying features things world discussed sense represent things terms distributed representations composed underlying features large space possible features assuming features relatively specific relatively relevant thing situation units representing features active network built produce sparse distributed representations better suited representing things world particularly important context learning discussed next chapter finally way viewing inhibition sparse distributed representations terms balance competition needs take place distributed representation multiple units represent thing complete competition localist complete fully distributed generally good balance begin exploration open project usual windows including overall control panel network contains input layer projects hidden layer excitatory units layer inhibitory neurons inhibitory neurons activation level hidden layer units thought inhibitory units hidden layer own layer purposes simulation inhibitory units total hidden units found cortex commonly roughly inhibitory neurons excitatory neurons outputs contribute inhibitory conductance neuron excitatory conductance set activation parameters different inhibitory neurons discussed lets begin usual viewing weights network select units weights random inhibitory units fixed constant value notice hidden layer excitatory units receive input inhibitory units inhibitory units receive feedforward connections input layer feedback connections excitatory hidden units well inhibitory connections themselves control panel parameters determine relative contribution feedforward feedback inhibitory pathways applies feedforward weights input inhibitory units inhibitory units controls scaling inhibitory connections back onto inhibitory neurons themselves see important parameters arbitrary relative scaling parameters described important importance properties different types inhibition now lets select view activations network window press control panel see input units activated random activity pattern cycles activation updating hidden inhibitory units active activation appears controlled inhibition excitation input layer note level leak current small excitation performed inhibition leak current case previous simulations average activity hidden inhibitory layer units plotted graph window overall average activity hidden units graph window clearly shows inhibitory units activated advance hidden units important simulation incorporates difference excitatory inhibitory neurons rate updating controlled parameters see excitatory neurons updated inhibitory faster faster updating inhibitory neurons allows rapidly adapt changes overall excitation level important function feedforward inhibition important practical point update rate constants important advantage simplified inhibitory function described next section rate constants set relatively slow order prevent behavior see set largely time scale excitatory neurons update activity smaller steps inhibitory neurons better able set parameters back default press button now manipulate parameters control panel determine roles producing observed activations lets start conductance inhibitory current excitatory units scales level inhibition coming excitatory neurons clearly predict plays important role change effect average level excitation hidden units inhibitory units increase happens now explain pattern results set back now lets see happens manipulate analogous parameter inhibition coming inhibitory neurons expect results similar obtained inhibition upon inhibitory neurons interesting consequences try good idea run comparison see excitatory activation inhibitory level roughly same try value now excitatory activation level increases inhibition again remains same difficult phenomenon understand think here ways understanding going seems straightforward reducing amount inhibition inhibitory neurons result activation inhibitory neurons look activity inhibitory neurons true increasing inhibition results lower activation feedback inhibition starts hidden units active inhibitory activity same level runs sense greater activation inhibitory units case hidden units causing lower activation result activation inhibitory units coming feedback hidden units imagine reduced activation inhibitory neurons increased activation lower activation level hidden units remain lower activation levels inhibition goes back original activation level way explain noting dynamic system balance excitation inhibition imagine time excitatory hidden units start bit active turn activate inhibitory units easily themselves turn provides extra inhibition advance hidden units effectively played level changes activations absolute levels explain see evidence looking absolute levels intuitive details way understanding effect inhibition inhibitory neurons terms location relative output place close constant distance away far away output strongly driven output analogous parameter larger values result higher levels activation greater hidden layer smaller values set back important point take away explorations number different ways changing parameters achieve roughly same resulting level hidden unit activation means actual biological system difficult reason result greater activation levels backwards underlying parameter change rise course tell different parameters typically involve level low biological inhibitory inhibitory connectivity strength empirically individual neurons certain parameter changes evidence know now lets influence feedforward versus feedback inhibition overall effects inhibition weve observed far set effectively feedforward excitatory inputs inhibitory neurons input layer affect behavior excitatory inhibitory average activity levels explain result hint think effects feedforward inhibition next set feedback inhibition now happens try finding value activity level initial default system differ initial system explain pattern results kinds inhibition useful important things inhibition changes weight values learning typically units learn levels excitatory input develop greater variance input patterns patterns providing strong excitation producing natural result specialization units representing things test current inhibitory mechanism changes simulating effects learning units excitatory weight values higher level variance press return default parameters case networks weights produced generating random numbers mean variance mean case baseline comparison now click control panel select weights same mean variance gaussian distributed values produces higher variance excitatory net inputs units hidden layer increase total overall weight strength increase variance larger weights mean press see difference overall excitatory level observed greater level excitation weights compared weights verify system change increasing things simpler far exploring relatively easy case inhibition network bidirectional excitatory connectivity clearly requires inhibitory feedback saw explorations now lets try running network bidirectionally connected hidden layers select back default parameters comparison network bidirectional excitatory connectivity examining weights usual note layer inhibitory neurons receive excitatory projections back layer enabling feedforward type impact activity hidden layer now new network graph log shows average activity hidden inhibitory layers note initial part point hidden layer begins active same layer activates back layer inhibitory neurons active excitatory neurons overall activity level remains control substantially different earlier simulations see inhibition important bidirectionally connected networks set parameter reduce amount inhibition excitatory neurons note relatively small impact initial feedforward portion activity curve hidden layer active network activated fit set parameter back final exploration point provides motivation summary inhibition function presented next section here want explore happens activity levels different overall levels excitatory input presented network press button enter value change input pattern units active default now activity level substantially different previous case difference observe increases activity level system appears relatively robust changes overall input excitation show effect demonstration comes relatively small differences initial activity level hidden units compared subsequent level input hidden layer approximate set point behavior system tends produce relatively fixed level activity regardless magnitude excitatory input inhibition function described next section explain general terms system exhibits set point behavior change activation function units spiking activation function saw appropriate combination feedforward feedback inhibition rise controlled activation excitatory neurons bidirectionally connected possible summarize effects inhibitory neurons setting inhibitory current excitatory neurons directly inhibitory function based level excitation layer avoid need explicitly simulate inhibitory neurons connectivity significantly amount computation required simulation addition able way avoid need slow time constants updating units processing resources simplest understand inhibitory functions known take kwta function idea here set point property inhibitory system set wide layer level inhibition number units achieve threshold equilibrium membrane potentials level inhibition rest remain threshold order implement function need able compute amount inhibitory current put unit threshold present level excitatory input interestingly networks learn better computation performed excitatory input minus contribution bias weights sense biologically bias input visible inhibitory interneurons excitatory input presumably functional level allows bias weights affect overall levels activation inhibition computation trying produce fixed point set activation level necessary equation familiar exercises equilibrium membrane potential equation compute threshold level leak current same thing here inhibitory conductance threshold written threshold membrane potential value represents excitatory input minus contribution bias weight value compute wide layer inhibitory conductance setting values units layer positions list units level excitatory conductance words wide layer inhibitory threshold placed active units layer ensures unit remains threshold unit expressed formula constant determines exactly place inhibition units value typically enables unit reasonably far inhibitory threshold depending terms distributed possible level excitation units layer plotted related excitatory net input axis order number units layer axis kwta function places wide layer inhibitory current value active units shown lines shape distribution significantly affects extent highly activated units rise threshold figure note simple relationship inhibition excitation assumed purposes figure shows normal distribution active units reasonably inhibition strongly activated units threshold resulting small inhibitory excitatory activated units strongly active units resulting large activated units shows distribution excitatory activation net input units layer large impact well active units able overcome inhibition computed kwta function strongest activity produced clear separation active units active functional property inhibition function activation reflects active units excited property weakly activated units tend activated position threshold reasonable leak current prevent units active parameter represents upper level activity lower levels activity occur response weaker excitation accurate name function wta version kwta function provides greater flexibility regarding precise activity level produced providing relatively upper tradeoff lack exact activity level worth advantages network able bit flexibility representations version called based average kwta inhibitory function simulator inhibition computed function top units remaining units function units specifically inhibition computed function average threshold conductance inhibitory terms top units written average threshold conductance inhibitory remaining units same formula putting inhibitory conductance values time value typically simply tends work better simulations illustration based average kwta inhibitory function computed wide layer inhibitory current value placed average values top units average remaining values simpler kwta function entire distribution average terms whereas kwta values units case inhibition similar place simple kwta function results lower level inhibition simple kwta units active results higher level inhibition simple kwta units active depending underlying excitation involved units active inhibitory function see learning algorithm shaping representations network bit appropriate overall levels activation different cases addition flexibility advantage scheme resulting global inhibitory conductance based units layer stable robust activation updating reasons version kwta function generally preferred simpler layer activation sparse unit active time simpler kwta function works better generally case explorations now lets explore kwta inhibition functions compare behavior previous networks already open open already open reset parameters default values button press choose select bidirectionally connected network standard activation graph case actual inhibitory neurons now select choose press notice roughly same level activity results inhibitory activity zero units function activity function kwta function effectively perfect job appropriate level inhibition required further activation hidden layer starts earlier faster parameter now select based average kwta function again order test set point behavior kwta functions run network levels addition standard types kwta function notice functions exhibit stronger set point behavior inhibitory unit based inhibition based average kwta showing slightly variability overall activity level kwta functions explicitly set point whereas inhibitory units roughly produce point set behavior remember kwta functions approximation effects inhibitory neurons identical fashion order see main advantages kwta functions select try find update parameter increments maximum result significant behavior value found compare value parameter based unit inhibition advantages faster updating think kwta fast update rate based unit inhibition order property simple kwta function apply set leak current value prevents weak excitation activating units allows strong excitation produce activation select value units allows excitation activate layers explorations inhibition return digits example revisit issues originally explored benefits inhibition better sense inhibition specifically simple kwta function performs case representations random activation patterns open project essentially identical project explored simple kwta inhibition effect controlled parameter maximum number hidden units strongly active bias weights turned default localist network reasons clear default parameter expected localist result single hidden unit strongly active input note leak current relatively weak contributing selection active unit setting running now increase parameter strongly activated units identical excitatory net input values meaning inhibition placed right placed right units threshold results weak activation shown due effects noise noisy activation function described situation shown active unit well ones way network produce active units continue increase parameter observe provides increasingly distributed patterns hidden layer total number active units due effects discussed advantage controlling number active units kwta inhibition function error trial manipulation parameter precise direct control outcome parameter set apply regardless changes network parameters affect parameter now lets explore kwta inhibition distributed based feature network choose force distributed network single hidden unit active difficult achieve setting units similar levels excitation unit activities well feature representing similarity structure digits value parameter produces cluster distinctions hidden patterns actually units active mechanism pattern found happens activity levels reduce leak current speaking kwta inhibition function single takes wta function simple case competitive learning algorithm excited units activity set rest zero version idea developed same kinds bayesian discussed case units activated extent likelihood generating input pattern larger units activity unit took form likelihood measure unit comes conditional probabilities hypothesis associated different units mutually hypothesis null hypothesis mutual assumption significant limitation type model saw distributed representations obtain considerable power enabling things represented multiple units assumption inconsistent distributed representations single wta models localist representations units graded soft activation values simpler models important mathematical understanding level entire network generally possible kwta function related form activation function simple wta function network builds ideas proposed here single chosen now neighborhood units active activation function distance learning networks exhibit interesting properties built tendency treat neighboring items similar ways revisit ideas model early visual processing learning important limitation network representation full power distributed representations match different units represent different combinations features see later build neighborhood bias lateral connectivity units effectively same thing network flexible kwta framework described approach similar explicit lateral connectivity finally number models constructed units communicate excitation inhibition directly units model goes name interactive activation competition bring number principles discussed chapter cognitive phenomena word effect described previously interactive activation same thing bidirectional excitatory connectivity provided top bottom processing model development kinds bidirectionally connected networks properties discussed models important limitations fact special inhibitory neurons clearly separation excitation inhibition found brain kind direct inhibition units good distributed representations multiple units active competing case unit active stable balance excitation inhibition required keep units active same time units getting active separate inhibitory neurons inhibition units layer point point excitatory neurons inhibition results consistent activation dynamics distributed representations saw earlier explorations explored distinct effects excitation inhibition now position think global level analysis bidirectional excitation inhibition seen part larger computational goal overall perspective called constraint satisfaction network seen simultaneously trying number different constraints imposed via external inputs environment weights activation states network itself mathematically shown connected bidirectional networks sigmoidal activation functions extent constraints original demonstration point due applied ideas towards understanding network behavior crucial energy function physical system energy function associated provides global measure energy system depends strength interactions connections atoms system system higher moving faster higher energy interactions constraints system part energy system function strong constraints extent system system constraints higher energy takes energy constraints think constraints takes energy system higher constraints greater extent higher energy system constraints lower state energy turns nature settle lower energy state satisfying constraints simple example energy function sum distance function objects separated energy function system dimensions objects closer distance value obviously gets smaller meaning constraint objects closer gets according energy function see energy functions typically squared form distance function network neurons state constraints thought lower energy apply energy functions networks find simple act updating activations units network results same kind settling lower energy state satisfying constraints standard form network energy function follows represent sending receiving unit activations respectively weight connecting important weights connecting units symmetric influences unit unit enables network find single consistent state units weights symmetric unit same state obviously lead consistent global state constraints represented function extent activations consistent weights large weight units units strongly active contribute smaller energy value minus sign magnitude sign energy term negative value energy function called harmony here units said contribute greater harmony aka lower energy strongly active connected large weight terminology network settling acts increase overall harmony activation states see occurs lets take simplest case linear unit computes activation follows take derivative harmony equation respect units activation value note term appears sum term harmony equation term same meaning updating activations network units same thing harmony network analysis network sigmoidal units similar point neuron activation function see work mathematically energy harmony equation needs entropy term reflects extent activation states units network middle range values resulting overall equation called based harmony free energy based results adding additional term typically affect relationship different states words harmony fairly redundant measures restrict focus simpler harmony term point neuron activation function leabra result increased activity units connected strong weights simpler cases function tend increase overall harmony network understand general processing networks performing constraint satisfaction working overall harmony exercises explore idea empirically local minimum energy function noise system represented global minimum least better local minimum noise membrane potential activation values play important role constraint satisfaction basically noise helps keep things getting think form noise system getting optimal sub state similarly noise added activations units prevent network getting optimal sub state fails constraints optimal states see optimal sub states called local local depending harmony energy respectively compared global optimal possible states maximum harmony value possible simplest networks typically settle local noise likely find better local actual neurons naturally occurring noise precise timing output spikes code rate functions discrete spikes source noise noisy function builds noise shape function see relatively rare cases network local problems add additional noise back activation function switch spikes ambiguous stimuli truly equal possible interpretations see cube example noise necessary break constraint satisfaction problems need resort special technique called simulated slow quality high neural network simulated gradually reducing level noise network settles idea early processing want explore wide range different activation states search better relatively high levels noise good state want reduce noise level way maximum simulator controls noise level function number processing cycles performed settling illustration dimensions units kwta inhibition restrict search space smaller subset possible patterns represents network kwta inhibition explore possible combinations activation states shows effect kwta inhibition restricting activation states explored faster effective constraint satisfaction point set kwta inhibitory function important impact constraint satisfaction performance network basically constraint satisfaction form parallel search network number different possible states finding constraints fairly unit updating parallel search parallel sequentially huge number distinct states sequence viewed way role kwta inhibition restrict search space possibility going states wide range different overall activity levels network kwta inhibition produce states relatively range overall activity levels see illustration advantages restricting search space network settle faster reliably good states assuming good states allowed kwta algorithm idea sparse distributed representations described learning means network important possible representations range activities allowed kwta function see sparse distributed representations useful cognitive tasks studied later chapters network kwta inhibition settle faster reliably good constraint satisfaction solutions begin exploration simple semantic network intended represent small set relationships different features represent set world case representing features color size toy network contains information number individual able information general common unique tell consistency feature harmony function useful total constraint satisfaction level network particular configuration feature inputs network perform pattern completion way information particular individual individuals simple network covered chapter feature values individual exploration knowledge embedded network summarized knowledge encoded simply setting weight representing individual corresponding feature value individual groups features values column table represented distinct layers own layer inhibition addition identity units name units own separate layers well inhibitory function here important network flexibility actual number active units layer parameter set layers see start project weights network verify weights implement knowledge shown table now locate enviroview right click button see network displayed window inputs soft clamped corresponding network units button lets verify present individuals name input recall information individual form pattern completion single unique input cue see layer units enviroview already default locate overall control panel press see network activates appropriate features ahead try name activations sure press button enviroview changes sure click previous clicking again activity input now lets see network general information versus level information set individuals select network activates features typical explain reason different levels activation different features activated useful information now lets constraint satisfaction ideas notice graphlog network window bring run network again input selected graph shows value harmony cycles network settling notice expected value appears increase settling indicating network increasingly satisfying constraints activations updated now lets specific network activating color input addition see initial harmony value slightly larger reflecting greater excitation present input final harmony value significantly lower likely network able well put way easy constraint resulting harmony large plus applies things harmony lower number different ways network ahead present different input patterns see kinds responses network reasonable response set constraints provided input pattern interesting try figure activation spread network settles network enviroview useful purpose showing state network default cycles updating cube seen looking looking now lets explore constraint satisfaction processing ambiguous stimuli example cube shown viewed cube orientations people tend back forth viewing way versus rare view same time words tend form consistent overall interpretation ambiguous stimulus consistency reflects action constraint satisfaction system favors interpretations constraints imposed possible interpretations open project network window see units cube representing possible interpretations left cube corresponding right units layer simple kwta inhibition operating parameter active time constraints usual lets examine weights notice unit connected local neighborhood gets active tend activate consistent interpretation entire cube active same time interpretation cube activating via inhibition competing active return viewing network press view competition process action running interpretations receive equal weak amounts excitatory input see network settles point units active strength back forth cube eventually fully active remains inactive try times note cube random eventually observe case part cube activated entire cube active happens note plot harmony value graph log substantially correspond consistent solution cube inconsistent partial satisfaction weight constraints lower harmony full satisfaction constraints cube noise added membrane potential playing important role simulation break cube interpretations see lets manipulate level noise try following values report differences observed settling behavior network different values tell noise process try playing different activation functions setting parameter finally important psychological aspects cube stimulus people tend possible interpretations occurs neurons activated interpretation eventually allowing competing units active process neurons getting called accommodation well established property neurons covered button turn property network runs cycles observe least cube next neurons cortex typically described different layers neurons primary groups functional layers see input cortical layer hidden cortical layers output layers cortical layers input layer receives information senses via thalamus output neurons motor control outputs wide range subcortical areas hidden layer serves output network response input signal means internal model provides useful typically elaborated processed basis driving outputs network large amount interconnectivity layers forms data support idea information flows primarily input hidden output via excitatory neurons further excitatory neurons bidirectionally connected information flow backwards pathways inhibitory neurons exist cortical layers receive same types excitatory inputs excitatory neurons layers outputs far typically number excitatory neurons relatively close themselves inhibition appears provide kind local feedback mechanism excitatory neurons excitatory neurons distinct ways unidirectional feedforward connectivity pattern set neurons set versa vice typically found cortex properties feedforward processing generalize common bidirectional case unidirectional excitatory connectivity input activity patterns way emphasizes certain distinctions specifically representing certain patterns inevitably collapsing possible distinctions cognition understood terms process developing specific representations emphasize relevant distinctions collapse irrelevant ones extreme case specific representations localist representations specific distributed representations number properties models apparently cortex bidirectional aka recurrent interactive connectivity common cortex important functional properties found simple unidirectional connectivity neurons affect themselves via connections emphasize symmetric case same weight value relatively simple understand compared case capable performing unidirectional transformations enables top processing similar propagate information units layer leads pattern completion partial input pattern presented network excitatory connections activate missing pieces pattern bidirectional activation propagation typically leads amplification activity patterns time due mutual excitation neurons important effects due bidirectional excitatory connections including mutual support top support biasing bootstrapping phenomena described general term dynamics network appears particular activation state bidirectional excitatory connectivity allows useful dynamics processing neurons activations neurons communicate enabling dynamically influence type processing performed see next chapter bidirectional connectivity important communicating error signals drive learning bidirectional excitatory connectivity next driving own monitor output feedback basically positive feedback causes system signals maximum maximum firing rate neuron problem cortex well known study order control positive excitation cortex inhibitory neurons activate inhibitory synaptic channels see excitatory neurons order balance excitatory inputs forms inhibition present cortex feedforward driven level excitation coming layer feedback driven level excitation layer itself combination forms inhibition results set point behavior occurs point excitation leads inhibition resulting excitation excitation leads inhibition resulting excitation system preferred level excitation explicitly simulating inhibitory neurons summary function kind set point inhibition directly resulting greater simplicity simulations kind inhibition called take kwta set point parameter neurons total neurons active time addition controlling excitation inhibition results form competition neurons competition produces selection pressure activation dynamics learning results evolution survival adaptation representations finally produces sparse distributed representations sense terms aspects general structure world combined effect excitatory inhibitory interactions understood terms constraint satisfaction activity patterns network settle time way satisfaction constraints internal external network process understood mathematically energy function overall level satisfaction network show net effect activation propagation increase overall measure finally explore role inhibition consistency settling time associated large constraint satisfaction problems cognitive neuroscience provide information cognitive functions different cortical areas good choice provides information anatomical properties cortex detector model neural function learning provides primary mechanism setting parameters determine neuron detects parameters principally weights synaptic connections neurons including bias weights learning depends individual level neuron mechanisms specify parameters level network principles developed previous chapter produce overall network behaves appropriately environment importance mathematically treatment learning algorithms driven critical mathematical played key role learning algorithms developed levels analysis developing ideas learning occur human cortex begin known biological mechanisms underlie learning mechanisms known terms long term ltp long term ltd refer weights transient non long term manner form ltp found cortex associative hebbian form depends pre postsynaptic neural activity hebbian learning viewed performing model learning objective develop good internal model important statistical structure environment type learning called organizing self explicit feedback environment hebbian learning perform well task learning objective produce specific output patterns particular input patterns contrast driven error learning algorithm delta rule task learning direct target output actual output patterns target information labeled learning see valid sources require constant presence generalization delta rule called backpropagation allows errors occurring layer backwards earlier layers enabling development useful intermediate representations overall task easier solve original mathematically direct backpropagation algorithm biologically implausible bidirectional activation propagation communicate error signals called generec consistent known properties ltp overall biology cortex further allows error signals occurring affect learning combination based hebbian model learning based generec task learning typically producing better results finally learning mechanisms necessary address important difficult forms task learning known sequence temporally delayed learning study biology learning level long term ltp term describe findings order contrast forms inevitably transient knowledge refers increase measured excitation controlled stimulus onto receiving neuron frequency high typically cause long lasting effect increases synaptic efficacy now know considerable amount biological mechanisms underlying ltp related phenomenon ltd long lasting synaptic efficacy common form ltp cortex known mediated nmda ltp nice connection way nmda receptor works functional characteristics form ltp functional characteristics generally summarized term associative means activity states pre postsynaptic neurons important enabling occur association neurons ltp evidence suggests pre postsynaptic neurons relatively strongly activated occur consistent idea active representations strongly called hebbian learning covered detail next section conditions ltd occurs clear appears somehow reduced zero non activity pre postsynaptic neurons lead ltd focusing ltp case moment observed explained fact nmda receptor open postsynaptic membrane potential sufficiently excited cause ions move opening nmda receptor channel otherwise excitatory neurotransmitter glutamate bind receptor opening channel ions pass presynaptic activation glutamate release postsynaptic activation present order nmda channels open nmda channels open allow calcium ions enter postsynaptic neuron low base concentration neuron new calcium able trigger complex chemical processes ultimately results modification synaptic efficacy weight primary excitatory input receptors receptors discussed number factors pre postsynaptic lead modification overall synaptic efficacy debate factors important far resolved appear pre postsynaptic factors involved ltp learning rules depend nature events trigger synaptic modification mechanisms actually implement sketch biological mechanisms lead ltp nmda channels open postsynaptic membrane potential sufficiently glutamate released presynaptic neuron allows calcium ions postsynaptic neuron triggers complex chemical processes ultimately result modification synaptic efficacy ltd well understood presented data ltd occurs synapse active lower level required trigger ltp explanation finding effective opening nmda channels channels opened time results lower concentration calcium ions triggers different chemical processes ultimately end reducing synaptic efficacy shows illustration relationship ltp ltd threshold ltd higher ltp threshold relationship consistent hypothesis nature complex chemical further form ltd consistent ltd necessary model learning necessary task learning explained relationship ltp ltd amount increased calcium leads ltd larger amount leads ltp biological picture complicated number reasons receptors channels play role ltp example evidence glutamate receptors number ways calcium enter postsynaptic neuron non nmda dependent voltage calcium channels ltp number dopamine etc ways understood appropriately experiments evidence standard occurs natural activation patterns cortex generally ltp sensitive particular combinations activation signal properties timing frequency etc ways explored empirically important biologically plausible driven error learning impact nmda receptor show preserved measured learning nmda generally difficult interpret nmda significant effects behavior addition preserved learning observed accounted simple nmda story role nmda non factors gated voltage calcium channels receptors story least level biological mechanisms end involved mediated nmda described number biological data suggests associative learning occurring cortex generally state data serves need computational models explore kind synaptic modification rules lead effective overall learning details appear difficult extract biology sketch objective model learning produce internal model represents important features world complicated true state world hidden see bunch projections world state somehow produce corresponding state model section develop level computational motivation general goal learning see goal accomplished biological mechanisms discussed previous section call goal model learning order emphasize idea learning towards developing internal models world shows illustration internal model captures important features world difficult thing basic idea kind underlying structure regularities natural constants general characteristics world somehow represented order function properly example highly reasonably accurate representation effects seems fairly learn things least ones apparent took understand fundamental role hidden think model learning fairly easy automatic case least cognitive trying human brain fundamental problems model learning nature sensory underlying structure world amount information senses seem senses large low quality information highly processed order produce apparently world experience see quality problem introducing appropriate built start biases organize information information problem addressed biasing learning favor simpler parsimonious models end kinds information favor representing relevant information form techniques recognized way science itself works explicit extension model learning process same problems problem world via senses receive series relatively limited dimensional sound etc thought mathematical term dimensional higher matrix onto lower dimensional think real world projections dimensional high world onto dimensional lower sensory process world states back world job model learning difficult information projection onto senses characterized situation saying problem input data sufficiently constrain interpretation difficult large number possible internal models fit sensory input data equally well put way difficult know real underlying causes noise important way model learning problem easier integrating individual experiences single ambiguous noisy time science known believe phenomena reliably demonstrated different individual experiments different etc law large numbers noise averaged away integrating statistics large sample see integration process critical successful model learning naturally performed slowly adding small weight changes resulting weights represent statistics large sample experiences network ends representing stable patterns emerge wide range experiences world experiences enable development good internal model example averaged pixel images experienced retina big needs prior biases kinds patterns particularly informative organize structure representations way sense general structure world biases provide reasonable fit properties actual world model learning task easier reliable individuals example lets imagine task learning control complex new explicit help available playing know advance kinds actions generally sets likely control actions words set appropriate biases structure internal model details experiments playing time contrast complete resort randomly systematically depending trying corresponding responses slowly actions button task difficult space possible contingencies greater appropriate biases model learning easier essential keep mind biases wrong completely learning take longer biases trying subset important ones tried example biases came prior experience same kinds learning processes type biasing happen model learning networks interested kinds biases present start learning assume evolution built years good set biases human brain facilitate ability learn world essential role genetic learning emphasized neural network learning algorithms characterized completely blank learning systems reflect dynamic genetic biases based experience learning difficulty genetic biases typically obvious easily useful contrast kinds genetic important biasing neural network learning typically discussed psychologists emphasize genetic contributions emphasized think terms people specific knowledge representations knowledge solid things generally neural network terms require building detailed pattern weights relatively implausible information contain difficult expressed biologically development contrast biases areas generally connected areas biases fast area learns compared inhibition presumably relatively easily encoded expressed actions factors development subtle differences biases lead important differences learning easy exact nature biological biases shape learning see general aspects biology networks specifically role inhibition biology learning specifically associative hebbian character serve important biases model learning role biases learning long statistics goes name variance bias rely biases strongly learn actual experiences end getting wrong model think experienced important new favor familiar ones hand rely experiences strongly assuming experiences end different models reflect lot model variance assuming real underlying state world means lot wrong models proper biases experience true generally optimal solutions history research contains examples biases critical sense phenomena otherwise best known examples bias science negative ones ability understand world truly science provides example nature biases biases science parsimonious bias favor simplest explanation phenomenon primary practical advantages developing parsimonious models world results greater generalization application models novel situations think way bunch specific facts world trying extract simpler essential regularity underlying facts novel situation relevant example encode situation away part part see model learning favor developing relatively simple general models world ways related role associative learning inhibition positive correlations exist elements feature line model learning represent correlations main idea model learning based learning correlations environment understand useful pay attention correlations consider individual pixels picture elements image line see pixels active line present input producing positive correlation activities correlation reliable present different input images extent reliable world tends produce lines edges objects model learning pay attention correlations general seems world things relatively stable features tree individuals face eyes nose features reliable correlations input further model strongest reliable features components correlation matrix well see next section hebbian learning cause units represent strongest correlations environment detailed analysis hebbian learning well explore simple case shown simulation exploration see single unit hebbian learning rule explained detail learns represent correlations present pixels line begin open project input layer single receiving hidden unit graph log showing weight values addition usual windows present single right diagonal line see effect hebbian learning hidden units weights lets look initial weights hidden unit selecting variable view network window clicking hidden unit see random looking pattern weight values click back control panel see activation right diagonal line click back see units weights learned represent line environment click again see entire random weights line representation looking weights simple point exploration hebbian learning tend cause units represent stable things environment particularly simple case single thing environment serves point well explore interesting cases lets try understand mathematical basis hebbian learning general mathematical framework understand hebbian learning causes units represent correlations environment called principal components analysis pca name suggests pca representing major principal structural elements components correlational structure environment arbitrary pca refers principal components correlation sequential order strongest last term refer strongest components focusing principal components correlational structure framework holds developing reasonably parsimonious model provides useful mathematical level overall analysis understand effects learning further see pca implemented simple associative hebbian learning rule mediated nmda synaptic modification described follows develop particularly useful form hebbian learning performs version pca rely combination top mathematical bottom weights adapted find nice levels analysis order fundamental computations clear start simplest form hebbian learning problems algorithm actually simulations important keep mind different versions developed common extracting representing principal components correlational structure environment hebbian learning lets focus simple case single linear receiving unit gets input number input units imagine environment produces patterns input units certain correlations input units lets consider simple case environment line shown repeatedly presented set input units linear receiving units activation function weighted sum inputs see diagram usual input units reasons clear subsequent equations chapter variables implicitly function current time step input pattern writing variable things read finally lets assume weights unit learn time step input pattern according simple hebbian learning rule weight change pattern depends activities pre postsynaptic units follows learning rate parameter index particular input unit expression weight change enables update weights follows want know going happen weight value result learning input patterns easily expressed sum time again time different input patterns now lets learning rate set value total number patterns input arbitrary constant anyway turn sum average notation indicates average expected value variable patterns equation formula index sum inputs linear function activities input units find bit weight changes function correlations input units new variable correlation matrix input units correlation defined here expected value average product activity values time familiar standard correlation measure away mean values variables taking product result ignore additional time assuming simplicity need assumptions later activation variables zero mean unit variance correlations computed via simple hebbian learning algorithm main result changes weight input unit weighted average different input units correlation input units particular input unit see diagram computation strong correlations exist input units weights units increase average correlation value relatively large interestingly run learning rule long weights dominated strongest set correlations present input strongest set next strongest increasingly large simple hebbian rule learns strongest principal component input data demonstration simple hebbian algorithm units perfectly correlated completely input pattern computed change weights added next time step computed simple concrete demonstration learning rule shown input units single linear output unit different input patterns units perfectly correlated perfectly units zero mean assumed notice correlated units determining sign magnitude hidden unit activation ensures weights keep increasing ones weights run patterns notice weights correlated ones increase rapidly remains small due mathematically say simple hebbian learning rule weights towards strongest correlation matrix seen notation simplification weight itself expected value assuming changing relatively slowly matrix serves update function simple linear system state variables represented weight well known results state variables dominated strongest component update matrix problem simple hebbian learning rule weights large learning continues obviously good thing relatively simple weight updating weights remain end exactly algorithm similar explain influential version hebbian learning achieves weight algorithm proposed developed following modified hebbian learning rule away portion weight value order keep large learning rule complex see considering simple case input pattern avoid need average multiple patterns look stable value weight learning long time pattern order simply need set equation equal zero tell equilibrium weight values reached note same find equilibrium membrane potential weight input unit end representing proportion inputs activation relative total weighted activation inputs keep weights bound finally primarily based same correlation terms previous simple hebbian learning rule rule computes principal component input data involved problem rule fact simple hebbian pca well linear activations work properly point neuron activation function thresholded subject effect large activations substantial problem pca extension case multiple hidden units consider happen added multiple hidden units same activation function learning rule unit analyzed end learning exact same pattern weights strongest principal component input correlation matrix algorithms find general ways problem involve introducing kind interaction hidden units order different things problem here form weight update rule overall activation dynamics hidden units important relationship activation dynamics learning level perfect sense unit behaves activation dynamics going affect learns see repeatedly approach explored solving problem redundant hidden units introduce specialized lateral connectivity units configured ensure subsequent units end representing sequentially weaker components input correlation matrix explicit imposed hidden units unit ends representing principal component strongest correlations next unit gets next strongest etc call sequential principal components analysis spca solution reasons level computational principles available data neurons encode information sequential pca spca performed small images natural scenes principal component blob upper left subsequent components following right square large grid shows grid receiving weights hidden units common layer input units figure important level computational issue spca assumes input patterns share common set correlations represented principal component individual patterns sequentially distinctions represented subsequent components amounts assumption hierarchical structure central tendency shared environment individuals special cases overall principal component contrast seems world lots separate categories things exist roughly same level seen shows spca algorithm applied small images natural scenes scenes contain lots line different orientations positions sizes largely images world see principal component big blob average lines individual image pixel different lines correlations present line away completely thing left general correlation close pixels general close pixels tend similar values blob subsequent components essentially divide blob shape sub representing residual average correlations exist away big blob conditional pca cpca performed small images natural scenes results properties simple cells early visual cortex respond bars light specific orientations way visual system seems represent images images line different orientations sizes etc neuron responding small coherent category line properties neuron fire strongly short lines degree activity graded fashion lines differ preferred line examples weight patterns shown see description network produced weight patterns problem spca computes correlations entire space input patterns meaningful correlations exist particular subsets input patterns example somehow restrict application pca hebbian learning rule images lines roughly particular orientation size length etc present end units encode information essentially same way brain units represent correlations pixels particular subset lines way expressing idea units represent conditional principal components pca computation subset input cases consistent ideas regarding importance relatively sparse representation unit active subset input cases see related ideas issue precisely specify conditions unit perform pca complex avoid moment simply assuming activity receiving units determined external source turns units input contains things representing turns see later inhibitory competition play important role conditionalizing process task based learning further weights learned pca procedure itself conjunction inhibitory competition result organization self representations selective weights cause units active features represent present input section develop version hebbian learning specifically purpose performing conditional pca cpca type learning see resulting form learning rule updating weights similar normalized pca learning rule presented emphasized critical difference learning rule itself activation dynamics determine individual units participate learning different aspects environment developing cpca learning rule develop slightly different way objective hebbian learning better suited cpca idea problematic assumption linear activation function previous versions pca cpca rule consistent notion individual units hypothesis detectors activation states understood reflecting underlying probability existing environment cpca learning rule starts taking conditionalizing idea assumes want weights input unit represent conditional probability input unit active receiving unit active write form simplified notation continue call learning rule achieves weight values cpca algorithm learning objective analyzed slightly different context competitive learning algorithm important characteristic cpca represented weights reflect extent input unit active subset input patterns represented receiving unit typical characteristic aspect inputs weights large typical small think weights reflecting kind correlation input unit receiving unit conditional probabilities correlation terms conditional probability means zero correlation equally likely receiving unit active values larger indicate positive correlation likely receiving unit values indicate negative correlation likely receiving unit note least important difference conditional probabilities depend direction compute general whereas correlations come same way regardless way compute fact cpca computes correlation input receiving unit interesting pca based computing correlation input receiving unit pca receiving units linear activation directly reflects input activations output input correlation ends reflecting correlations different input units see cpca sensitive input correlations capable reflecting important conditionalizing factors determine receiving unit active competition amongst hidden units following analysis show following weight update rule achieves cpca conditional probability objective represented again learning rate parameter equivalent forms equation shown order emphasize similarity learning rule normalized pca learning rule shown showing simpler form main difference latter square activation times weight activation times weight expect produce roughly similar weight changes difference way works note activations cpca positive probability values difference activation affect sign weight change form emphasizes following interpretation form learning weights match value sending unit activation close possible difference weighted proportion activation receiving unit receiving unit active weight occur effectively receiving unit care happens input unit itself active receiving unit active lot input units activation set weight match individual weight changes slow learning rate averaged weight come approximate expected value sending unit receiver active words next section shows weight update rule shown implement conditional probability objective mathematically analysis significantly ability understand follows analysis based same technique understand normalized pca learning rule work backwards weight update equation equation setting zero solving equilibrium weight value show weights converge case conditional probability need relevant variables terms activations sending receiving units assumed represent probabilities corresponding units active consistent analysis point neuron function terms bayesian hypothesis testing presented expression represent probability receiving unit active particular input pattern presented represents corresponding thing sending unit total weight update computed possible patterns probability pattern occurs set zero order find equilibrium weight value solve resulting equation value results following now interesting thing note here actually definition joint probability sending receiving units active patterns similarly probability receiving unit active patterns equation point clear fraction joint probability probability receiver definition conditional probability receiver right end noted cpca hebbian learning rule essentially same learning rule competitive learning algorithm found pca soft competitive learning style networks see information algorithms commonly form hebbian learning simulation models addition pca idea number different ways effects learning rule competitive learning algorithm causes weight move towards input data clusters showed informative addition presenting conditional probability analysis learning rule roughly argued useful weights towards input data clusters shown idea clusters data important represented units network easy see strongly correlated input patterns tend form clusters sense way looking pca idea incorporates additional assumption multiple separate clusters different hidden units specialize representing different clusters conditionalizing idea unit learns patterns somehow relevant cluster competitive learning conditionalizing occurs via simple take competition described unit allowed active input pattern units weights come represent cluster likely cluster automatically resulting specialization units subsets input according clusters process organizing self learning mentioned previously problem simple competitive learning network single active hidden unit provides localist representational basis powerful distributed representations require feature cluster present input time clearly limits algorithm kwta activation function see allows multiple units representing input pattern unit representing features present simultaneously input same time sufficient competition units appropriately responses meaningful subset input patterns indicated inhibitory competition plays important role conditionalizing units learning important analysis hebbian learning related pca idea idea information idea here model learning develop representations amount information input patterns turns principal component correlation matrix information possible single unit causes units output amount variance input patterns variance information idea information placed context certain constraints taken extreme result development representations capture information present input result relatively parsimonious representations better consider role hebbian learning context tradeoff information complexity representations tradeoff represented framework known minimum description length clear inhibitory competition results specialization helps produce parsimonious models overall information capacity hidden layer works balance information objective emphasized form hebbian learning cpca extracting principal component subset input patterns receiving unit active learning rule itself addition inhibitory competition pressure develop relatively parsimonious model input patterns principal component relatively informative far capable representing information present input patterns different ways hebbian learning studied context cpca learning rules viewed weights effectively divided factor amounts probability receiving unit active case cpca due conditional probability form possible weight order control found necessary getting model hebbian learning early visual cortex produce receptive fields individual neurons specialize representing inputs initially receiving inputs see adding contrast enhancement factors cpca learning rule provide benefits increasing selectivity representations main limitation weights end bound whatever placed turns useful practice graded intermediate weight values necessary cases simple cases studied earlier version leabra algorithm combination turns combined soft weight bounding necessary keep weights graded learning rule implements cpca algorithm described due effects soft weight bounding algorithm perform well described beginning chapter likely biological mechanisms underlying weight changes cortex showed generally support hebbian associative type learning cpca learning rule slightly complex simple product sending receiving unit activations requires further explanation see account general characteristics weight changes cpca learning rule same basic mediated nmda ltp mechanisms described previously easier reference cpca equation pass seeing biology implement equation lets assume weight value well consider effects different weight values moment assumption general categories weight changes produced cpca sending receiving units strongly active weight increase ltp easily account associative nature mediated nmda ltp receiving unit active sending unit ltd occur explained nmda channels open function postsynaptic activity small amount presynaptic activity causing small zero level calcium lead ltd possible postsynaptic activity activate gated voltage calcium channels provide weak concentrations calcium necessary ltd presynaptic activity receiving unit active likelihood magnitude weight change goes zero explained nmda channels lack activation gated voltage calcium channels lead postsynaptic calcium weight changes finally effect cpca learning different values weights summarized follows weight large further increases happen likely larger weight smaller magnitude decreases show opposite pattern conversely weight small increases likely larger magnitude opposite holds decreases general pattern exactly observed empirically amounts ltp ltd upper lower respectively thought form soft weight bounding upper lower case soft manner slowing weight changes return issue context task learning later section exploration revisit simulation ran beginning chapter see single unit learns response different patterns correlation activity set input patterns conditionalizing activity receiving unit shape resulting weights emphasize feature present subset input patterns find need introduce additional factors learning rule order emphasis effective factors important organizing self case explored subsequent section begin open project want weights hidden unit learns select variable view network window click hidden unit now select control panel choose environment window events right diagonal line left sets correlations exist simple environment manipulate percentage time receiving unit active conjunction events order conditional probabilities drive learning cpca learning algorithm simple case explored earlier right line presented setting probability left lines probability view probabilities frequencies associated event locate parameter view window upper left hand side select thing displayed event display right side window see event means receiving unit active time conjunction right diagonal line time left note completely irrelevant times receiving unit inactive conjunction patterns weight change case frequencies conditional probabilities associated event events arbitrary probability appearing environment parameter control panel determines frequencies events environment event set lets set value ahead iconify environment window continuing important keep mind exercises single receiving unit multiple receiving units looking same input patterns want unit specialize representing correlated features environment lines case manipulate specialization conditional probabilities weighted towards event now press button control panel run network sets epochs randomly event event event value cpca hebbian learning rule applied event presentation weights updated see display weights network window updated epochs graph log display value right weights red left weights orange notice learning weights units active central unit present events weight weights event expected cpca learning rule causes weights reflect conditional probability input unit active receiver active experiment different values verify holds different probabilities parameter control panel corresponds cpca learning rule determines rapidly weights updated event change affect general character weight updates displayed network window explain happens explain importance integrating multiple experiences events learning set parameter back explored different values effectively selective receiving unit type event taking advantage conditional aspect cpca hebbian learning effectively conditionalizing representation input environment frequency events occurred environment think frequency receiving unit active events now want compare conditionalizing aspect cpca pca algorithm lets assume event equal probability appearing environment setting simulate effects standard form pca receiving unit effectively entire environment cpca receiving unit active lines present environment result lead weights weight pattern suggest existence separate diagonal line features existing environment compare blob solution natural scene images discussed shown set simulate hidden unit controlled way come right diagonal line input left result lead weights explain result informative case explored previous question extend architecture training network represent environment fully way explain answer simple environment weve far realistic assumes mapping input patterns categories features typically want represent lets switch environment pressing environment notice now different versions left right diagonal lines upper lower addition original center lines environment spread types right lines general category right lines left lines set see right lines weights left lines weights result illustrates couple problems cpca learning algorithm units represent categories features single instances weights end due fact receiving unit active different input patterns conditional probabilities individual pattern relatively small happens receiving unit perfect selectivity category features right left case crucial early phases learning reasons clear later differences weight magnitude input units selected category right categories left small sense cpca algorithm actual conditional probabilities emphasize selectivity receiving unit small overall weight values reduce dynamic range weights end inconsistent weight values produced task learning algorithm described later next section shows problems fixed come back simulation again see work practice cpca algorithm results normalized weight values saw previous section tend selectivity dynamic range introducing correction factors way weights taking account sparse expected activity level sending layer way contrast weak strong weights correlations computing effective weight value sigmoidal function underlying linear weight value cpca computed changes affects basic underlying computation performed cpca simply effective further continuum standard cpca weights weights effects clearly understood note exploration next section clear effects correction factors better job verbal arguments provided here fully understand section continue exploration reading section lot sense automatically expected activity level layer represented variable compute net input projection see automatically cpca learning rule idea expected activity level reflects expected conditional probability receiving unit active input unit chosen random active correct fairly sparse activity levels lack correlation input receiver produce weights values expected activity level correct idea conditional probability indicates lack correlation larger values indicate positive correlation smaller values negative correlation procedure weights standard range turns best way accomplish renormalization simply increase bound upper weight increases form cpca update weight equation simple verify form equation equivalent form understood analogy membrane potential update equation excitatory conductance term drives weights towards reversal potential term inhibitory conductance drives weights towards reversal potential correct sparse activations excitatory reversal potential greater increase range weight values produced new maximum weight value purposes learning weights standard range resulting potential resolution value conditional probability happens practice linear relationship true underlying conditional probability equilibrium weight value turns obvious way expected activity levels input produces relationship weight actual conditional probability learning things appear correlated actually set correction factor following equation expected activity level same standard cpca smaller values produce relatively larger weights finally order convenient specify general way correction factor want introduce parameter simulator close actual starting expected activity level compute effective activity level follows compute renormalization occurs weights fit range renormalization amounts relatively simple modification cpca learning rule contrast enhancement requires significant changes level simpler contrast enhancement important contribution well worth additional complexity subsequent simulations show further important keep mind implementation here involves different weight values introducing contrast enhancement effects learning rule itself essentially amount form soft weight bounding implement contrast enhancement sigmoidal function provides obvious mechanism contrast enhancement function gain parameter determines gradual function enhance contrast weight values linear relationship weight values turn conditional probabilities computed cpca algorithm sigmoidal relationship mediated gain parameter biologically speaking amount sensitivity weight changes weight values middle range opposed plausible directly supported data far know note kind contrast enhancement weights equivalent effects standard gain parameter activation values changing activation gain unit sensitive differences total net input sending activations times weights contrast changing contrast enhancement weights affects weight value separately allows unit sensitive individual input level total input put way weight contrast enhancement unit sensitive detecting patterns inputs whereas activation contrast enhancement net response sensitive threshold increase quality signal coming unit simulator implement weight contrast enhancement introducing effective weight computed underlying linear weight following sigmoidal function weight gain parameter simulator controls extent contrast enhancement performed note function derived same form repeatedly text difficult gain parameter equation latter form see plot function standard parameter effective weight value computing net inputs units standard value simulator linear weight value simulator internal variable computing weight changes basically sure averaging events results appropriate conditional probability values difficult measure linear weight value biologically effective weight value measured result activating sending unit excitation produced receiving fact said possible computations effective weight value computationally simpler keep values effective weight value function underlying linear weight value showing contrast enhancement correlations middle values conditional probability represented linear weight value note function offset parameter add additional parameter function controls offset acts threshold useful higher threshold correlation order further enhance contrast different features present input offset parameter simulator introduced effective weight value equation follows values greater higher threshold underlying linear correlation values weight value results value shown important note point threshold dividing line contrast enhancement process values sigmoidal introduced important different conditional probability values relative contrast enhancement threshold renormalization parameter plays important role addition obvious importance parameter see following simulations think effective weight function bias learning algorithm favors extreme weight values controlled parameter bias requires particularly strong correlations strong weights controlled parameter see biases appropriate range different learning tasks finally way thinking contrast enhancement context effects soft weight bounding property cpca algorithm weights constrained range approach slowly difficult units develop highly selective representations produce strong activation input patterns weak activation contrast enhancement limitations advantages soft weight bounding cpca open already project view weights hidden unit going explore renormalization weights taking account expected activity level input layer line stimuli units active units input layer value set lets explore issue environment features zero correlation receiving unit see renormalization results weight values case control panel pick environment see contains horizontal lines presented equal probability line features represent zero correlation case occur receiving unit same probability expected activity level input layer words expect same level occurrence simply input units random overall activity level input network see due behavior cpca algorithm reflecting conditional probabilities weights end learning weights terms standard meaning conditional probabilities represents zero correlation conclude input units correlated receiving unit know correct sparse activity levels input now set note variable name means sending average activation parameter control panel value means now applying full correction average activity level sending input layer network again observe weights now correct value expressing lack correlation ability fully correct sparse sending activations useful want particular prior expectation individual input patterns represented hidden unit set appropriately level corresponds roughly prior expectation example know units relatively selective representations input features unit want set full correction input layer result larger weights features relatively weakly correlated compared expected level selectivity units expected represent number input features value issue now lets explore contrast enhancement function effective weights parameters control panel control gain offset function lets set result substantial contrast enhancement see shape resulting effective weights function button control panel press file selection comes see sigmoidal function plotted graph log window bottom screen try setting different values effect shape function order see effects learning lets baseline comparison see blob representation right lines bit strong left lines now set see right lines represented relatively strong weights contrast enhancement allows network represent reality distinct underlying left right categories features selective features important organizing self learning see now lets parameter encourage network pay attention strongest correlations input set see affects effective weight function back forth couple times able see difference set network change results compared case explain occurs find value central non overlapping non units right lines units lower left units upper right weights resulting weights accurately reflect correlations present single input pattern imagine representation useful cases alternative way accomplish effects value described units selective weak correlations high weight value set back set effect learned weight values compare parameter found previous question last question shows contrast enhancement big effect changing amount correlated activity necessary achieve value lower correlated inputs parameter large smaller values towards zero causing unit essentially ignore inputs interactions contrast enhancement renormalization play important role determining unit tends detect already times ability network hebbian learning inhibitory competition organize self internal model environment context conditional pca learning algorithm organization self amounts competition set receiving units way conditionalizing responses units unit active extent strongly activated current input pattern units happen weights unit sufficiently well tuned input pattern cpca learning algorithm causes tuning weights input units active receiving unit active effectively positive feedback system initial selectivity set input patterns learning algorithm producing greater selectivity positive feedback systems potential positive feedback saw bidirectional excitatory connectivity previous chapter organizing self learning case individual receiving units end representing number input features receiving units represent features important phenomenon happens learning causes units tuned subset input patterns unit ends representing set patterns causes unit likely activated ones example consider case unit represented right diagonal line explorations appropriate contrast enhancement parameters learning unit caused weights decrease left diagonal line increased right diagonal line unit likely respond left diagonal lines allow unit competition case resulting good representations types lines continue lines exploration set hidden units environment consisting horizontal vertical lines input retina start opening project see number windows lets focus network input projects hidden layer hidden units fully connected input random initial weights usual select view weights units viewing pattern weights hidden units primary network learns special grid log window displays weights hidden units see press control panel display weights grid log window lower right side screen scale larger grid same network grid elements smaller grid representing input units showing weights unit clicking hidden units network window able verify now lets see environment network press control panel bring window showing events representing different combinations vertical horizontal lines events unique combinations type line real correlations lines reliable correlations pixels particular line putting way line thought appearing number different randomly related lines pretty obvious computed correlations individual pixels images equally weakly correlated learning conditional particular type line order meaningful correlations see simply organize self interactions learning rule kwta inhibitory competition note lines present image network require least active hidden units input assuming unit representing particular line iconify environment window done examining patterns return viewing network window lets step bit processing network default network button turned turn clicking button upper left hand network window now locate process control panel weights grid log adjacent network window press button hit present single pattern network see event patterns containing lines input network pattern roughly active hidden units hidden layer based average kwta inhibition function parameter set see parameter control panel function allows variability actual activation level depending actual distribution excitation units hidden layer units active units fairly equally activated input pattern due random initial weights selective important effect weaker additional activations enable units bootstrap stronger activations gradual learning end reliably active conjunction particular input feature particular line case single press button process control panel want turn epochs different events environment learning network weights grid log updated epochs weights came clearly reflect lines present environment individual units developed selective representations correlations present individual lines random context lines happened result individual units developing initial weak selectivity function random weights caused inhibitory competition subset images contained elements selective turn conditional pca learning increase selectivity etc units initially selective multiple lines units better able represent lines competition back representing line dynamics inhibitory competition critical organizing self effect finally case representations multiple hidden units encode same line feature net result organizing self learning nice combinatorial distributed representation input pattern represented combination line features present obvious way represent inputs network representation complex organizing self learning procedure see representation action turn network back general units strongly activated input pattern extra activation reflecting fact lines coded multiple units flexibility based average kwta function allows observed flexibility network patterns same time providing sufficient competition force specialize main thing notice weights shown grid log units obviously selective units aka units reliably activated input feature experience learning typically important units organization self requires learning units stable correlational features hidden units increases large range initial random organization self process consequence need units necessary end presumably problem brain worse algorithm didnt work unless precisely right number units idea brain neurons actually needs units activated new features later presented network diagonal lines act kind looking weights informative measure well networks internal model matches underlying structure environment measure plotted graph log network window network learned shows results unique pattern statistic simulator number unique hidden unit activity patterns produced result network different types horizontal vertical lines presented separate testing process epoch learning goes tests network lines resulting hidden unit activity patterns kwta parameter set critical due flexibility based average kwta function number unique patterns measure line encoded least distinct hidden unit show unique pattern units encode lines good model environment result unique representation lines resulting measure lower extent statistic internal model produced network fully capture underlying line lines seen run model eventually produced perfect internal model according statistic well analysis weight patterns order better sense well network learns general run batch training runs starting different set random initial weights time press button repeatedly pressing begins new random setting random initial weights updates end training run graph log updated training runs text log window upper right screen show summary statistics average maximum minimum unique pattern statistic last column contains count number times perfect perfect runs now lets explore effects parameters control panel lets manipulate parameter affect contrast selectivity units weights set run network statistics number uniquely represented lines obtain ways final weight patterns shown weight grid log different default case explain findings related specific reference role selectivity organizing self learning set back change run batch statistics obtain case change weight patterns compared default case explain results terms effects function contrast enhancement mechanism again explain important organizing self learning now lets consider parameter controls amount renormalization weight values based expected activity level sending layer value parameter weights increase rapidly driven larger maximum value see value result smaller weight increases described smaller values appropriate want units selective representations larger values appropriate general categorical representations smaller value parameter help prevent units developing selective representations multiple lines default value parameter set back set case explain result terms weight patterns observed weight grid log training run compare general effect parameter parameter set back set sets initial random weight values mean value batch run network pay particular attention weights see units unit now line illustrates interesting details organizing self learning effects learning overall weights general cpca rule causes weights increase input units active decrease amount weight decrease large relative amount increase input patterns active units start relatively large initial weights unit active pattern likely later similar identical pattern pattern containing lines common previous pattern weights overall increased happens increase initial random weight values results units useful effect network sufficient numbers units ends based task learning see later finally lets manipulate learning rate parameter set back set increase learning rate effect network explain case hint representations integrate multiple patterns exercise feel dynamics organizing self learning importance contrast enhancement order cpca algorithm effective generally now extent parameters provide appropriate biases learning process benefit produce basic process organizing self learning explored characteristic number models including competitive learning networks primary advantage model weve here traditional approaches comes kwta form inhibitory competition allows arbitrary combinations multiple units represent input pattern obviously critical exploration able learn powerful distributed representations organizing self manner intermediate case single wta kwta based neighborhood networks organizing self learning applied neighborhood activities active point enables organization self representations networks studied developed seen extension earlier ideas role lateral connectivity interacts hebbian learning producing representations influential hebbian learning modeling organization self early visual system developed showed hebbian representation input correlations produce unit activities early visual system extent model linear shown parameter sensitive model useful potential benefits hebbian learning cover issues detail slightly different approach organization self based primarily activities receiving units cover space input patterns example type learning algorithm ensures unit active essentially same percentage time effective features similarly distributed environment good match number units hidden layer number features environment plausible assumptions real world human brain respectively useful place upper lower limits average activities units layer order prevent active done simulating effects accommodation sensitization accommodation effect difficult activate neurons active time sensitization effect causes neurons easily excited active recently saw effects point neuron activation function see later improve learning cases finally important organizing self learning models called models based idea recognition idea here top image based internal model world learn based difference generated actually advantage learning mechanism requires internal model fit precisely possible actual input patterns principle lead representations correlations units problem produce worse fit details specific patterns further models easily understood terms bayesian statistical framework likelihood term plays important role framework essentially model extent hypothesis internal model produced data actual perceptual input problems models problem encourage network capture information present input image goes important bias requires away lot information order represent general underlying structure example cpca algorithm away principal component correlation matrix subset patterns represented unit question addition biological reality process clear algorithms typically bottom top weights set weights time clear weight switching occur biologically alternative implementation units activity reflects difference states predicts expected states result activity appear true brain expect reduced activity expected states problematic models require processing neurons areas higher short kinds processing required models run ideas developed previous chapter processing brain viewed involving bidirectional propagation information performing multiple constraint satisfaction information relevant levels processing resulting activation state constraint satisfaction model including top bottom processing critical explaining cognitive phenomena later chapters processing principles model based learning biological data consistent hebbian learning relatively simple easy implement form think brain pretty good learning algorithm turns models performance advantages further think hebbian learning naturally implements number important valid biases described learning framework anyway developing internal model environment useful survival environment certain learning solve actual tasks useful now focus task learning neural networks simple general means solve task produce specific output pattern input pattern input context contingencies demands task output appropriate response reading text correct answer addition problem straightforward examples output input mappings learned see subtle ways tasks learned cpca hebbian learning rule developed model learning good learning solve tasks need learning algorithm perform important kinds learning lets begin seeing well simple output input mappings exploration based simplest form task learning set input units project single output unit task terms relationships patterns activation input units corresponding target values output unit type network called pattern associator objective associate patterns activity input output lets begin opening project output units receiving inputs input units set feedforward weights locate control panel press button select output input relationships learned task simply input units left output unit active units right output unit active relatively easy task learn left output unit develop strong weights units ignore ones right right output unit opposite note kwta inhibition output layer parameter network trained task simply input output units corresponding values events environment performing cpca hebbian learning resulting activations see action locate control panel upper right side network window continue press presented random order end epoch events see activations updated result testing phase run epoch training testing phase events presented network time output units clamped updated according weights input units clamped testing phase actual performance network task test results testing displayed grid log lower right hand side screen row represents events input pattern actual output activations shown right column squared error sse simply difference actual output activation testing target value clamped training sum output units actually computing thresholded sse absolute differences zero means unit activation correct side order zero error think units essentially binary activation value expressing likelihood underlying binary hypothesis true single training epoch output unit likely errors now turn button upper left hand side network window press button control panel see grid log update epoch showing pattern outputs individual sse errors graph log immediately network provides summary plot epochs sum thresholded sse measure events epoch shows referred learning curve network rapidly zero indicating network learned task training automatically network correct epochs row sure learned problem order see learned press button control panel plot final output activation patterns testing window turn network back see actual unit responding see network learned easy task turning left output patterns right next now lets take look weights output unit see exactly happened click network window left output unit see expected weights left units strong right units weak complementary pattern hold right output unit explain pattern weights cpca hebbian learning algorithm now lets try difficult task press hard task overlap amongst input patterns cases left output right output unit somehow figure task relevant input units set weights problem hebbian learning apparent concerned correlation conditional probability output input units learn sensitive inputs task relevant unless happens same output input correlations case easy task hard task complicated pattern overlap amongst different input patterns cases left output inputs correlated extent middle strongly correlated cases left output right overlap considerably last event containing highly correlated inputs network attention correlations tend respond last case shouldnt lets see happens network task button produce new set random starting weights viewing weights left output unit network window network learns see weights learned middle units highly correlated output unit expected training network getting right different runs produce slightly different results case middle events turning right output unit last turning left output producing weaker activation output units relatively equally excited looking weights right output unit show strongly represented correlation input unit pattern output responses reason weight right output unit stronger left output unit middle inputs different overall activity levels different input patterns difference affects renormalization correction cpca hebbian learning rule described renormalization set constant different events help network solve task level detail task network ever solve task report final end training run depending answer last question now difficult impossible hebbian learning solve task now turned algorithm solve task impossible single layer units learn mapping big deal turns pattern weights lead correct solution see algorithm better suited task learning conclude hebbian learning limited task learning works correlational structure task structure number subsequent simulations well finally experiment parameters control contrast enhancement cpca hebbian learning rule see playing important role networks behavior find combination contrast enhancement parameters default default reliably networks performance multiple runs problem hebbian learning rule task learning care network actually performing task representing correlations conditional probabilities sending receiving units evident explorations network learned correlational structure ignored fact producing wrong outputs find specific ways hebbian learning work specific tasks form learning moment develop based task learning algorithm principles central principle base task learning directly goal producing correct output activations order need kind measure tell close network producing correct outputs way measure weights obvious measure squared error sse statistic described want extend sum events resulting again target value confused event index actual output activation implicitly function time event zero outputs exactly match events environment training set larger values reflect worse performance goal task learning error measure serves objective function driven error learning objective learning illustration computing derivative sse respect weights shows configuration single input output neuron target activation input activation assuming linear activation rule means weight needs task right shows sse function weight weight smaller change sse increasing weight derivative negative adjust weight negative increase weight towards true weights larger standard direct way minimize function take derivative respect free parameters weights network adjust parameters according negative derivative sense derivative change function want direction changing function negative shows illustration simple case single input output units linear activation function input activation fixed output activation equal weight value target say weight starts smaller value increase gets conversely weight larger shows derivative sse weight values negative negative results expected weight increase derivative positive weights larger taking negative results weight mathematically take derivative sse function produce expression tell exactly sse line function changes individual weight values network taking negative resulting expression weight update learning rule called delta rule input stimulus unit activation known least mean time essentially same equation rule conditioning look form learning rule sense adjust weights reduce error basically weights change function local error individual output unit activation sending unit sending units active big error receive error example output unit active shouldnt negative weights input units active hand output unit active weights increase input units active next time units activation closer target value error reduced illustration credit process activity unit represented output unit active positive weights increase proportion activity sending units active sending units good same principle holds output unit active process weights proportion sending unit activations called credit appropriate name illustrated important computational property driven error learning rules similar level pca hebbian learning rules view representations rule delta driven error learning reflecting results multiple credit satisfaction mechanism credit process output input pattern integrated weight changes entire training set reflecting strongest correlations case hebbian learning weights here reflect strongest solutions task hand now see delta rule works show derived directly derivative squared sum error measure respect weights input units essential reader understand details mathematical details important thing understand effects terms credit explained issue fact weights appear directly clear enter computation appear order proceed need actually specify exactly computed function weights expression derivative depend function sense know exactly changing weight affect activations figure change weights produce target activation values things simple start linear activation function now chain rule take derivative respect appear error equation take derivative respect weights based linear activation function chain rule expression written follows thing term thing next term chain rule series terms case figure terms separately end taking term notice outputs events considering change weights particular output unit particular event learning rule local sense depends single output unit single pattern take next term linear activation function notice elements sum involve particular weight appropriate full derivative case linear activations follows derivative want negative learning rule ignore introduce factor error measure arbitrary learning rate constant anyway delta rule shown previously issue bias weights learn recall bias weights provide constant additional input neuron see proper bias weight values essential producing useful representations example allowing units represent weaker inputs turns isnt straightforward way train bias weights hebbian learning model correlational information correlation unit itself perfect perform driven error task learning delta rule algorithm put bias weights good standard way thinking train bias weights consider weights coming unit active bias weight change sense bias weight adjust try decrease error example extent unit active shouldnt bias weight change negative positive cause bias weight decrease causing unit active bias weight learns correct relatively constant errors caused unit generally active inactive problems rule delta based task learning mechanism larger deal later problem computing derivative point neuron activation function respect weights simple linear activation function problem delta rule natural range weight values take positive negative values magnitude biological psychological reality target output values turns reasonable solutions problems solution problem involves steps assume approximate point neuron function sigmoidal logistic function showed back reasonable thing see later able avoid approximation necessary time see different error function appropriate valued binary underlying representations case analysis point neuron activation function terms representing unit detected derivative logistic activation function ends mathematically need know problem comparison entropy cross squared sum error sse single output target value larger output new error function called entropy cross distance measure probability defined actual output activation target activation probability variables range target itself entropy variable defined log entropy cross entropy variables think variance variance squared error function zero actual activation equal target increasingly larger different squared error function treat entire range value large whereas values produce error way function takes account underlying binary true aspect variables see comparison sse define functions unit sigmoidal logistic activation function here useful break activation function net input term weighted activations sending units logistic activation function operates net input convenient write function intended sigmoidal character function now take derivative error function respect weight again chain rule time extend chain rule include separate step derivative activation function respect net input net input respect weight again well break component terms note here notation indicate derivative sigmoidal logistic function cases equation single variable simple derivative partial derivative respect multiple variables shouldnt trying figure derivative involved finally derivative net input term linear activation put terms see term derivative logistic function resulting exactly same delta rule implement version delta rule learning problem ignore fact weights naturally learning algorithm allow adapt additional problematic least reasons biologically implausible excitatory neurons weight take positive negative values function learning know cortical neurons excitatory inhibitory switch time inhibitory neurons directly implemented leabra effects simulated kwta activation function meaningful combine subsequent driven error algorithms cpca hebbian learning rule later unless weights produced driven error learning same natural imposed cpca recall cpca computes conditional probabilities weight values naturally range reasons following mechanism bounding driven error weights weight change computed driven error algorithm otherwise opposite otherwise same kind soft weight bounding cpca algorithm naturally weights approach slowly note equation same general form form cpca hebbian weight update rule equation updating membrane potential functions effect producing weight value series individual weight changes equal magnitude opposite sign corresponds well hebbian interpretation reflecting lack positive negative correlation similarly positive weight increases negative ones weight value increases decreases negative changes positive finally regarding problem simplest interpretation reality target value think activation state corresponds experience explicit signal external source actual observed outcome event world produced activation state thought response internal expectation outcome refer activation states different phases call expectation phase minus phase outcome target phase plus phase reasons clear later discuss issue depth now assume target values actually experienced network activation states rule delta learning involves taking difference target actual produced activations further need assume somehow forms information present form learning again discuss greater length now actually exactly equation equation described reduces equation conditions revisit pattern associator task open default initial parameters pick switch rule delta updating weights start training network need understand different activation phases work simulator sure network window locate control panel press now pressing button usual need increase resolution press button perform phase processing processing entire event locate parameter process control panel note button case lower value currently set change means press button result settling activation updating process associated phase processing now hit button see network actual activation produced response input pattern aka expectation response minus phase activation now hit again see target aka outcome plus phase activation tell target exactly note activations units easily produce activations typical net input values learning occurs phase activation now lets monitor weights clicking left output unit process control panel complete training task network learning task perform multiple variable relative hebbian case switch learning algorithms reflects critical driven error learning basically output unit performing task correctly learning effectively stops whatever weight values happened contrast hebbian learning weights reflect conditional probabilities task least results roughly same final weight values regardless initial random weights return issue later discuss benefits combination hebbian driven error learning now real test pick press see network learns task apparently difficulty delta rule performs learning function well network actually adapt weights specifically order solve task compare contrast weight values learned delta rule task learned hebbian rule explain delta rule weights solve problem hebbian ones sure include bias weights perform multiple delta rule case pressing weights describe aspects solution remain constant runs changes finally explain relatively general terms delta rule able learn task hebbian rule experience think delta rule powerful choose notice input unit active equally output active inactive difficult input unit end output input patterns same seems humans learn solve task fairly attention overall patterns activation lets see network press delta rule learn appears relatively simple task conclude powerful necessary clearly delta rule better hebbian learning task based learning limits interestingly limitation played large role early neural network models book contained mathematical limitations delta rule learning algorithms problems impossible case tried researchers neural networks simply powerful model human cognition field towards metaphor computer based models sufficiently apparent people limitation applies networks layers input output layer pattern associator models turns delta rule relatively directly extended generalized deal networks hidden layers input output layers results algorithm commonly called backpropagation learn impossible task previous example algorithm learn function uniquely input patterns output patterns hidden units critical advantage hidden units enable problems represented ways easier solve way problem represented plays critical role ability solve familiar insight problems essentially amount new useful representation answer apparent example try figure following sense words here chosen context sense able represent words context see perfect sense example comes simple turned representing part tree cognition based development appropriate representations input patterns think important issues study cognition importance developing learning mechanisms backpropagation produce representations series hidden layers return issue repeatedly text already discussed context transformations discussed previous chapter illustration basic processes standard backpropagation algorithm shows feedforward propagation activations layer same sigmoidal logistic activation function shows error backpropagation step challenge figure update weights input units hidden units already know adjust hidden output weights delta rule took time idea times backpropagation algorithm simple extension delta rule continues chain rule hidden units weights lets take standard case network input hidden output feedforward connectivity sigmoidal logistic activation functions see chain rule propagate error signals back output layer hidden layer finally weights input units hidden units see illustration assume moment units standard logistic activation function well write activation output unit hidden unit input stimulus unit net input unit unit function written equal note weights hidden units output units trained simple delta rule derived real problem training weights input hidden chain error function continue entropy cross error function weights input hidden layer follows lets break smaller lets look derivative terms take expression contribution hidden unit activation overall error term interesting tells change activation hidden unit affect resulting error value output layer imagine weights hidden unit output units way hidden unit influences outputs already expression similar chain rule delta rule steps difference delta rule took derivative net input output unit respect weights equal sending unit activation whereas here want take derivative same net input respect sending unit activation equal weight relevant term net input taking derivative terms versa vice expect overall derivative delta rule weight end activation computation interesting suggests hidden units compute contribution overall output error errors output units project errors strength units contribution output units return equation bit now continue chain backpropagation compute remaining actually similar already computed consisting derivative sigmoidal activation function itself derivative net input hidden unit respect weights know sending activation overall result entire chain negative adjust weights way minimize overall error layer network backpropagation delta rule output units weights hidden units weights illustration computation backpropagation learning standard feedforward layer network generic delta notation error values lets take moment understand properties backpropagation computation right way clear introducing new variable unit variable represents contribution unit towards overall error output layer derivative error respect activation unit hidden units computation derivative error respect net input unit output units recall due entropy cross error term derivative activation function respect net input here hidden units part comes equation terms new variables output layer turns same equation applied regardless number hidden layers network hidden layer takes weighted sum delta terms hidden output layer derivative activation function derivative activation function understood terms difference amount weight change actually going activation value unit unit sensitive middle range function weight change big difference activation derivative maximum conversely unit weight change relatively difference consistent derivative small learning rule effectively focus learning units think versus want focus chance generic delta backpropagation clear strong similarities feedforward propagation activation layer computing same basic function activations preceding layer backpropagation error layer computing same basic function errors subsequent higher layers clearly shows backpropagation error error essentially term algorithm provides illustration generic backpropagation terms write completely generic expression change weights terms sending unit error variable receiving unit interestingly looks simple hebbian learning rule product presynaptic sending postsynaptic receiving variables presynaptic term activation value fact postsynaptic term terms means isnt hebbian well see next section write equations completely terms activations come learning rule hebbian biologically plausible hebbian learning rule major problem backpropagation procedure biological reality error variable far certain necessary mechanisms network required equations finally note weight update rule bias weights according set sending activation value constant backpropagation procedure implemented biologically plausible way based important ideas algorithm developed ideas generalized restricted case algorithm resulting generalized algorithm generec providing complicated development ideas present key derived algorithm context generec algorithm end based task learning algorithm rest text illustration generec algorithm bidirectional symmetric connectivity shown minus phase external input provided input units network settles record resulting minus phase activation states plus phase external input target applied output units addition input units network again settles generec notion activation phases mentioned previously implementation delta rule network itself responsible setting activations output units expectation response minus phase environment responsible providing target outcome output activations plus phase generec operates allowing network symmetric bidirectional connectivity settle updating activations network minus phase settle again plus phase see illustration indicate phase plus variables indicate phase minus variables equations key idea way generec compute error terms activation values error signals seen examining way net input computed bidirectionally connected network standard layer network analyzed net input hidden units sum weighted activations input output units minus phase net input plus phase observe due bidirectional connectivity activation hidden unit reflects influence output activation states lets assume hidden unit somehow measure difference input signals received output units different phases difference related error output units seems bidirectional activation propagation hidden unit able measure output error see works mathematically lets net input terms different phases perform simple recognize last form equation similar expression derivative error respect hidden unit activations shows key insight algorithm difference net input terms compute error contribution hidden unit putting way start looks difference net input terms compare equations important subtle difference apparent net inputs equation terms feedback weights output units hidden units whereas equation terms feedforward weights hidden units output units order hidden unit accurately compute sending contribution output error weights hidden units output units based receiving input output units weights output units back hidden units assume weights same value recall needed symmetric weights order understand constraint satisfaction properties networks well discuss biological assumption now need least insight working algorithm difference activation states difference net input states times derivative activation function key insight algorithm compute error terms activation values hidden units net inputs greatly simplifies things see works lets equation variable hidden unit key insight algorithm term difference net inputs hidden units output units phases recall expressed derivative activation function turns resulting product difference sigmoidal activation values computed net inputs difference hidden units activation values equivalent difference net inputs times activation function net input values long linear approximation activation function reasonably actual net input units inputs coming output layer inputs relatively constant phases helps ensure phase differences relatively small linear approximation reasonable illustrated clear differences axis equal differences axis times function simplification differences activation states major advantage addition simpler eliminate need explicitly compute derivative activation function derivative implicitly computed difference activation states big advantage allows based biologically point neuron activation function kwta inhibition entire layer units layer derivative difficult compute computing resulting complicated derivative need imagine neuron itself compute derivative weight updates computed generec algorithm terms hidden units computed difference activation phases nice result learning rule same units network essentially delta rule see generic learning rule generic receiving unit activations sending unit activation backpropagation learning rule applies units regardless hidden layers network activation differences continue reflect appropriate backpropagation error exercise interested reader show means locally available activation states pre postsynaptic units perform driven error learning need biologically error backpropagation mechanism different normal propagation activation network rule bias weights summary difference phases activation states indication units contribution overall error signal interesting thing bidirectional connectivity ends naturally signal network needs done difference drive learning activation states local space sequential time synapse weight changes occur see weight changes happen biological synapse important consequence algorithm error signal occurring network drive learning enables different sources error signals addition form learning bidirectional connectivity known exist cortex actually requires connectivity order drive learning effectively finally noted generec approximation actual backpropagation procedure bidirectional network potentially complex settling dynamics propagation phases activation values separately difference generec same difference itself backpropagation approximation holds well deep networks performing complicated learning tasks done problem basic generec learning rule shown symmetric weight changes computed unit unit same computed unit unit units rule end symmetry weights learning rule work properly place way updating weights known method average minus plus phase activation sending unit method method symmetry sum ideas change weights combined result following interestingly additional generec algorithm equivalent called hebbian learning algorithm chl aka mean field learning algorithm locally available activation variables perform driven error learning connected networks algorithm derived originally networks activation states described distribution context chl amounts reducing distance probability arise phases settling network algorithm extended case restricted cases derived distribution continuous energy function require problematic assumptions conclude chl networks chl directly backpropagation algorithm via generec basis ability learn difficult problems further generic form chl aka generec remaining problems largely learning rule context kwta inhibition function conjunction hebbian learning rule showed problems chl effects purely driven error learning bidirectionally connected network well explore issues short providing further biological psychological potential reality generec algorithm human cortex least done mathematical side things summarize adjust weights network subject soft weight bounding procedure described previously generec approximate error backpropagation locally available activation variables fact variables available locally plausible learning rule real neurons based activation signals opposed error variables increases relatively straightforward map unit activation onto neural variables averaged time membrane potential spiking rate see main features generec algorithm potentially problematic biological perspective weight symmetry plus minus phase activation states ability activation states influence synaptic modification according learning rule recall generec requires symmetric weights order units compute sending error contribution based receive back units ways biological weight symmetry generec addressed show exact symmetry critical proper functioning algorithm rough form symmetry required biology show least rough form symmetry actually present cortex data consistent arguments summarized here order point noted symmetry learning algorithm chl combined weight automatically lead symmetric weights start way assumes units connected place difficult case connection chl learning algorithm algorithm effective connectivity input non units possible connections ways error signal information obtained hidden unit obtain error signal directly output units connections hidden units note absence connection different presence connection symmetric non weight value form found problematic analysis case subset error information available latter case result specifically wrong information due influence symmetric non weight due automatic property chl latter case problem terms biological evidence symmetric connectivity indication cortex least roughly connected level anatomical cortex areas visual cortex connected area projects area area receives projection area level cortical columns prefrontal cortex connectivity symmetric interconnected neuron received projections neurons neurons detailed level individual neuron symmetric connectivity difficult assess empirically least evidence exist further evidence least rough symmetry detailed symmetry critical demonstrated chl connectivity long way information subset symmetric connections via neurons same area based phase activations central generec algorithm aspect driven error learning cortex signal come emphasized generec signal phase plus activation state suggests signal state experience network state actual outcome previous conditions minus phase thought expectation outcome conditions example words sentence expectation develop word likely come next state neurons upon generating expectation minus phase experience reading actual word comes next subsequent demonstrate generated state activation serves plus phase idea brain constantly generating subsequent events subsequent driven error learning suggested psychological interpretation backpropagation learning procedure particularly generec version backpropagation activation states requires additional mechanisms providing specific signals effects experience neural activation states via standard activation propagation mechanisms different forms error signals simple layer network right layer figure represents state adjacent layer left subsequent time step additional layer explicit signal output implicit based expectation error signal input triggers subsequent expectation experienced implicit expectation based motor output consequences motor action expected experienced implicit single input layer note here layers different input representing different states time provides kinds based expectation signals arise conditions shows case people typically assume think driven error learning signal explicitly provided word explicitly shows similar kinds error signals arise result implicit expectation word pronounced followed actual experience word pronounced ones reading show different implicit generated including outcome motor output actual outcome actual input received addition ways error signals based outcome expectation differences evidence electrical activity behavioral tasks cortical activation states reflect sensitive example studied widely going positive occurs stimulus onset considered measure subjective determined preceding experience short long term terms showed determined amount prior resolved processing event nature consistent idea represents plus phase activation following relatively short time activation minus phase specific properties itself due specialized neural mechanisms presence suggests possibility neurons neocortex experience states activation relatively rapid corresponding expectation corresponding outcome weight change according chl cpca hebbian learning rules conditions minus plus phase activation values cpca care phases assumed take place plus phase plausible biological mechanism produces combination rules shown column suggested minus plus phase activations follow rapid remains shown activation states influence synaptic modification manner largely consistent chl version generec equation discussed capable implementing critical aspect chl recall already showed biological mechanisms consistent cpca hebbian learning rule extent chl inconsistent cpca possibly argue biology argue biology consistent combination cpca chl shows chl cpca learning rules differ direction weight change different values minus plus phase sending unit receiving unit note cpca based phase variables further specify occurs plus phase see learning rules row table row rules differ predict different weight change rule predicts weight change predicts weight change combination rules ends further see combination driven error hebbian associative learning generally beneficial solving different kinds tasks already shown cells accounted biology cpca hebbian learning remains shown lower hand left cell table chl learning rule predicts weight decrease ltd occur according biology refer cell error correction case occurs synaptic larger minus phase plus phase strong expectation associated units minus phase actually experienced outcome plus phase important contribution driven error learning enables network correct expectation output otherwise hebbian learning capable right things shown table explain error correction case relationship calcium ion concentration direction synaptic modification proposed discussed shown idea synaptic modification level calcium higher high threshold leads ltp level lower high threshold lower threshold leads ltd basic idea here minus phase synaptic activity followed similar greater levels plus phase synaptic activity lead level threshold threshold resulting ltd mechanism plausible consistent known data directly tested further requires kind additional mechanism indicate activations considered plus phase seems likely phase plus signal produced same kinds dopamine mechanisms described modeled discuss brain dopamine systems apparently fire whenever expectation outcome specifically case reward studied general known dopamine efficacy ltp appropriate phase plus learn now signal summary available data consistent directly supporting biological mechanism enable driven error task learning now lets put theory work see generec scale small task learning problems same problems pattern associator case time extra hidden layer units inputs outputs theory enable network solve impossible task start opening project associator major exception introduced hidden layer units increased learning rate takes time solve impossible problem lets start network problem select choose default learning rule set now press network displays sse error measure epochs training testing grid log lower right now updated epochs training shows states hidden units training network stops automatically gets entire training set correct epochs row note correct solutions happen due noisy behavior network learning shape learning curve reason noisy behavior relatively small change weight lead large overall changes networks behavior due bidirectional activation dynamics produces range different responses input patterns sensitivity network property networks networks bidirectional connectivity typical feedforward networks feedforward backpropagation network learning same task learning curve people nature learning networks share backpropagation find benefits bidirectional connectivity dynamics far learning curve further turns larger networks exhibit learning sensitive small weight changes epochs take network learn press times order sense fast learns general provide general results rough average learning done explain hidden units represent enable network solve impossible task weight values network display explain terms hidden unit activated sure multiple runs extract general nature networks solution data try explain general terms hidden units let networks solve difficult problems otherwise solved directly adding hidden units enable network solve problem button select couple times observe complete hebbian learning work well hidden units simple tasks see network learn couple epochs added end sure now select fast generec learn task compared hebbian rule sure run times good sample try explain occurs last exercise important hebbian learning faster reliable driven error learning reasons explored greater depth clearly interacts nature task tasks network typically going easy sense hebbian driven error learning see fact combination types learning works best developed explored different kinds learning model task obvious question arises relationship forms learning think part cortex performing model learning task learning extent issue addressed field assume case people assume sensory processing proceed largely entirely basis model learning higher output oriented areas task learning general division important consider alternative view suggests task model learning performed integrated fashion areas cortex good reason considering idea biological mechanism synaptic modification discussed suggests case number functional computational level reasons think model task learning work well generally levels processing benefit representing relevant structure world model learning developing representations useful solving tasks task learning example seem reasonable relatively early areas visual system develop representations relevant task distinctions important survival accurately represent subtle perceptual features correspond different states imagine based task learning provide necessary emphasis distinctions collapsing distinctions task relevant different trees seems likely left own model learning incapable determining subset huge number reliable distinctions actually represented case statistical structure clusters correlations tree features stronger likely represented model learning features case pretty shape form argument well model learning generally facilitate task learning cases solving tasks requires developing representations underlying elements task likely things space time appear correlational structure model learning representing further likely correlations particular input patterns particular output patterns meaningful task learning model learning enhance representation correlations model learning enable large deep hidden layers networks learn rapidly reliably critical typically hidden layers cortex perception motor output saw previous chapter previous exercise demonstrated easy task hebbian learning generally faster reliable learning important speed local nature hebbian learning rely driven error learning possibly weak error signals order weights learn appropriately note generec driven error learning algorithm locally error signals drive learning communicated backwards network output layer reliable truly local correlations hebbian learning represents saw exercises hebbian model learning produces reliable weight patterns random initial weight values represents correlations conditional probabilities available activity patterns important degrees network otherwise case learning actually constrained hebbian model learning saw generally capable learning tasks view combination types learning learning process primarily task learning task learning constrained model learning levels processing respect view model learning particularly form term commonly learning describe additional biases introduced further constrain otherwise types learning common example weight small portion weight value weights updated network weights reliably contributing solution otherwise weights zero finally note hebbian learning addition driven error learning inhibitory competition kwta functions represents important additional constraint learning process substantially performance network cases inhibitory competition appears significant overall contribution additional hebbian learning inhibitory competition form long model learning context task learning context error difficult compute unless generec algorithm computes implicitly form competition kwta allows distributed representations develop essential interesting tasks generally architecture gating units soft wta competition aka output contribution corresponding problem context inhibitory competition task learning situation framework illustrated here soft wta take competition takes place special group units called gating units provide modulation outputs corresponding groups units specialize solving different parts overall task group distributed representations same limitations wta algorithms operate individual units apply here well example extreme nature wta competition network active time limits extent different solve problems algorithm individual units powerful same limited overall dynamics wta systems contrast think effect kwta algorithm learning task context better kind fine specialization different units different aspects task allows powerful distributed representations implementation combined model task learning straightforward amounts simply adding weight changes computed hebbian model learning computed driven error task learning additional parameter simulator controls relative proportion types learning according following function learning components cpca hebbian rule chl generec driven error rule respectively parameter typically smaller values larger relatively small amount hebbian learning needed primarily error signals typically smaller magnitude bit training hebbian learning constantly same kind pressure learning error signals change depending network getting right wrong consistency hebbian learning larger effective impact well important note parameter useful means relative importance hebbian versus driven error learning different simulations possible single value parameter modeling different areas cortex hand different areas differ parameter way genetic biological biases influence development specialization different areas see note hebbian rule computed phase plus activation states sense computationally want move towards activation states sense learning statistical structure biologically assuming learning actually plus phase see benefit combining task model learning bidirectionally connected networks comes generalization domain central cognition specifically generalization viewed system representations capture underlying structure regularities environment simple example know count generalize sequence produce further instances sequence long structure systematic mapping onto digits enables produce understand domain numbers clearly way thinking generalization central understanding humans exhibit rule systematic behavior purely driven error task based learning typically highly constrained actual tasks saw case delta rule pattern associator task large range weight values network solve task true cortex neurons number possible combinations weight values truly likely task based learning capture essential structure environment weights reflect large contribution initially random values seen hebbian model based learning better job extracting structure constrained task feedforward driven error networks typically generalize well due ability networks treat novel input graded based similarity novel input known input patterns hidden unit provides robust graded estimate similar input pattern trained examples representing same level exists output hidden mapping well further important hidden units participate representation input residual hidden units representation averaged performed output units hidden units illustration bidirectionally connected interactive networks small initial differences settling resulting large final differences known effect shown initial points dimensional state space feedforward network processing takes place allow differences significantly contrast interactive network requires updating allowing differences greatly problem constrained nature weights purely driven error network show introduce bidirectional connectivity know computationally important feature cortex essential generec algorithm bidirectional connectivity complex settling dynamics lead small differences input patterns opposite graded feedforward network treating novel pattern familiar patterns closely interactive network likely treat resulting generalization performance illustrates phenomenon understood terms effect effect name idea interactive dynamics lead later small initial differences considerably complex dynamic systems extent bidirectionally connected network operates activation space processing similar inputs network settles time interactive networks exhibit dynamics similar inputs result roughly same final activity state thought dynamics theory lead better generalization effect sensitivity behavior same behavior expressed network depends nature weights define relative activation space settling bidirectional network sensitive noisy weights lead activation space feedforward constrained nature weights driven error network result worse generalization case suggests way solve problem eliminate weights exactly hebbian model learning produces reliable final weight values seen addition inhibitory competition present cortex model causes individual hidden units take greater representing specific input patterns causes weights closely correspond important distinctions different inputs further inhibition effect activation dynamics network see reduces effects explore simple example effects combined task model learning generalization simple oriented lines environment explored model learning section begin open project notice network now output layer output units corresponds different vertical horizontal line task learned network simple activate appropriate output units combination lines present input layer task simple provide good demonstration pure hebbian learning actually capable learning task time later next section task provides particularly clear relevant demonstration benefits adding hebbian learning otherwise purely driven error learning control panel contains parameter learning compared driven error learning see main parameter order compare purely hebbian model learning purely driven error task learning combination turn learning bias weights pure hebbian learning well select purely driven error task learning see network trained lets step events sure network window locate control panel right next network window event press again see plus phase output units reflect lines input position left bottom press button process control panel turn network network graph log updated epoch training error statistic epochs important test statistics sse training error statistic count number events error again threshold unit unit activation right side event counted measure plotted red line graph log test statistics same unique pattern statistic see measures extent hidden units represent lines plotted yellow statistic plotted green measures generalization performance network previous cases network trained total patterns remaining testing generalization lines presented subset lines training set possible recognize produce correct corresponding output unit novel context lines testing set network recognize underlying regularity environment line line line regardless context appears generalize well produce relatively number errors testing items green line plots number testing events network gets wrong smaller value better generalization performance addition graph log weight grid log updated epochs showing weights change learning well purely driven error network generalization task final generalization error epochs training explain performance terms unique pattern statistic weight patterns hidden units reference constrained nature purely driven error learning order determine particular result run batch training runs press button overall control panel see new text log show place graph log present summary statistics training runs addition bring text log weight grid log shows final results training run report summary statistics batch run indicate earlier generally applicable now lets see improve performance adding hebbian learning overall control panel select see value control panel now now control panel bring back graph log training run run additional hebbian learning change results compared purely driven error learning report results batch run explain results terms weight patterns unique pattern statistic general effects hebbian learning representing correlational structure input finally lets see well pure hebbian learning task select changes notice network learns rapidly depending goes network perfect performance task itself generalization test runs network fails learn perfectly due hebbian learning correct errors gets better find case rapidly press number epochs press again task obvious correlational structure well suited hebbian learning algorithm clear hebbian learning helps here network reliable driven error learning see next task hebbian learning helps correlational structure particularly obvious pure hebbian learning completely incapable learning apply range different generalization tests critical benefits combined model task learning training deep networks hidden layers explained previously additional hidden layers enable representation problems ways easier solve clearly true cortex visual system example original retinal represented huge number different ways build upon hidden layers revisit issue develop model visual object recognition multiple hidden layers examined issue example problem benefits multiple hidden layers known family trees problem task network learns family relationships capable relationship instances trained representing intermediate hidden layer individuals family represent specifically functional similarities individuals enter similar relationships represented similarly happens deep network example difficult purely driven error network learn learning time actually decreases network single hidden layer original deep network generalization completely finally depth scaling driven error learning worse bidirectionally connected networks training times long feedforward case times longer combination hebbian driven error learning illustration analogy driven error learning deep networks difficult hebbian organizing self model learning adding layer learn useful representations constraints inhibitory competition viewed restricting flexibility motion useful analogy understanding driven error learning work well deep networks bidirectional ones easier balance placed moving base back forth direct effect single effects indirect higher bottom increasingly effects higher nature effects depends position lower similarly error signals deep network increasingly indirect effects layers further away training signal output layer network nature effects depends representations developed layers output signal increased non associated bidirectional networks problem gets worse way problem easier internal greater self least partially balance themselves idea model learning provide exactly kind self learning local produces potentially useful representations absence error signals slightly abstract level combined task model learning system generally constrained purely driven error learning algorithm degrees adapt learning thought range motion easier balance explore ideas following simulation family trees problem subsequent chapters family tree structure learned family trees task english indicates now lets explore case learning deep network same family trees task structure environment shown task network trained produce correct name response presented activating name units input layer conjunction units input layer training network produce correct unit activation output layer open project notice network input layers output layer bottom network layers localist representations different people different relationships means similarity input patterns people localist representations richer distributed representations able support based similarity generalization performance see finally central layer responsible performing mapping representations rise correct locate control panel press training events help understand task presented network ahead scroll events list see different events presented network click look particularly interesting see represented now lets see works network itself locate control panel right next network press button followed button activations network display reflect minus phase state event press network train events rapidly network graph lower right displays error count statistic training red average number cycles took network settle orange learning curves family trees task showing combination hebbian model driven error task learning leabra results faster learning deep networks compared pure driven error learning backpropagation network network train epochs initial default parameters take press network learning curve load network network window select menu upper left epochs learn indicated file name graph log select menu choose epochs total train network completion network window name want come back later epochs took network learn problem indicates relatively rapid learning deep network example shows comparison typical learning curve leabra algorithm weve versus standard backpropagation network took epochs learn required large learning rate compared standard leabra network interested here raw learning speed own fact additional biases constraints imposed combining model task learning obviously learning networks hidden layers deep networks order see contribution hebbian learning run network overall control panel select network again want train load trained pre network log time select train network way different name hebbian learning clearly learning deep networks takes longer epochs case compared learn repeated runs networks different parameters effect further compare pure driven error network backpropagation network shown connected bidirectionally driven error chl network kwta activation constraints takes epochs learn clear kwta activation constraints playing important role well now lets see pure hebbian learning task select run network epochs isnt going improve load network trained epochs selecting log file hebbian model learning useful driven error learning simply capable learning tasks own order compare cases load identify lines based epoch end interesting note orange line average settling cycles fairly well correlated training error combined error network achieves significant cycles note pure case starts settling epochs learning turns contribute generalization performance networks see later simulations training patterns containing elements item question required achieve good generalization whereas individual appear times training set here notice general shape learning curve sse epochs compared leabra network ran pay special attention epochs learning primary differences cases network inhibitory competition via kwta function possible importance learning based results note network larger learning rate now compare sse learning curves red lines start different case suggest role hebbian learning hint error signals smaller network learned cluster plot hidden unit representations prior learning learning combined hebbian driven error learning trained network corresponding different clusters organized generally according generation finally examine representations network performing cluster analysis hidden units lets comparison initial clusters network learned task press process control panel weights press tests patterns cluster window appear load trained network default again results look note ways people appear related think initial plot final plot sensible structure terms overall difference coming clusters individuals generation overall clusters far considered relatively discrete kinds tasks output response expectation etc depends input pattern world real tasks extend time obvious example language meaning currently reading depends sequence words sentence language words themselves constructed sequence distinct sound patterns phonemes examples including tasks driving work seem sequential tasks exception consider here general categories temporal sequential temporally delayed continuous sequential case sequence discrete events structure aka grammar sequence learned temporally delayed case delay event production output consequences example corresponding smoke fire etc ones slow coming benefits degree here issue learning particular sequence events determining relationships set possible events delayed obviously mutually cases issues apply task important type learning applied temporally delayed problems called reinforcement learning based idea temporally delayed reinforcement backwards time update association earlier states likelihood causing subsequent reinforcement finally case focus timing continuous case relevant information detailed temporal evolution best described continuous system discrete set events obviously important motor control perceptual tasks discrete tasks detailed timing matter meaning sentence depend significantly rate read case possible sequence taking set regular long order capture relevant information difference sequence important processing continuous depends critically detailed temporal response characteristics neurons biological parameters weights likely critical further available mechanisms learning continuous far biologically plausible tend work well relatively simple tasks learning sequential temporally delayed tasks challenging area particular neural network models reason adding time steps output input contingencies lot adding extra hidden layers network way learning time backpropagation additional time step equivalent adding new hidden layer input output layers advantages model learning biases specifically hebbian learning inhibitory competition leabra algorithm weve learning deep networks expect useful learning temporally extended tasks example explore indicates likely case follow interesting complex cognitively relevant tasks explore context representations deal learning sequential tasks move reinforcement learning temporally delayed learning central problem learning sequential tasks developing useful context representations capture critical information previous events needed produce appropriate output interpretation later point time simplest case called case prior context necessary immediately previous time step case particularly convenient problems information away arise context include information prior time steps interesting note non environment context representation prior state representation include sufficiently rich amount information contain necessary contingencies case context representation contains copy prior states simple recurrent network srn context layer copy hidden layer activations previous time step neural network model incorporates style context representation developed known simple recurrent network srn called network context layer network input layer activity state set copy prior hidden version output version unit activity state see hidden state based context layer network considerable flexibility context representations learning representations hidden layer adopt basic srn idea main way sequential tasks later chapter simple limited context srn introducing additional mechanisms introduce greater flexibility represent context exploring basic srn couple issues addressed nature context layer representations updating further computational perspective need examine biological representations explore simulation basic srn standard srn context layer copy hidden layer computationally convenient necessary basic function performed layer context layer information transformation hidden layer set weights going context hidden layer adapt fixed slowly updating transformation hidden layer representations context layer important possible biological implementation context discussed essential context layer copy hidden layer essential context layer updated controlled manner example alternative idea implement context layer seem easier imagine biology implementing kinds network dynamics described additional layer free context units presumably connected amongst themselves enable sustained activation time somehow maintain information prior states special operation context units communicate hidden layer via standard bidirectional connections couple problems situation basic tradeoff context units information prior hidden state hidden units settle new state new input updating new state hidden units units need stable generic activation functions well related problem driven error learning procedure generec take account activation units settle learning based final activation states reasons simple context representation work simulations easily demonstrate simulations simply operation update context representations equation update context unit parameter determines extent context unit gets updated new input typically set determines reflects previous state basic srn constants typically additional gating mechanism place control special set units network simulator following steps need taken context layer needs set units projection projection hidden layer context layer needs tells context unit hidden unit activation standard full projection going opposite direction finally needs completely network start trial context layer previous hidden context layer activations done setting parameter default ensure network starts settling state context state variables processing preserved state state obvious brain area context representation frontal cortex discuss greater detail frontal cortex frontal pre cortex pfc seems involved planning extended temporally behaviors example people frontal lesions incapable sequence behaviors tasks perform individual step perfectly well appear specific deficit steps addition frontal cortex appears important maintaining representations time discuss argued demonstrated neural network models internal maintenance system important called context example example ambiguous words require kind context meaning writing implement argued pfc responsible maintaining necessary internal context words context established information presented earlier text further showed people pfc functioning context appearing immediately ambiguous word context appearing previous sentence findings internal context maintenance points directly srn context layer provides context necessary sequential temporally extended tasks further data show context representations produce sequential behavior sequentially presented information language see model addition behavioral data frontal cortex lot evidence neurons brain area exhibit sustained firing relevant task delays sequences frontal cortex receives projects areas posterior frontal non cortex connectivity produce appropriate context representations context representations affect ongoing processing posterior system considered hidden layer finally likely basal ganglia play similar role frontal cortex lower levels temporal extent relevant motor control perception rapidly changing stimuli speech sounds basal ganglia neurons firing patterns pfc context representations lesions areas produce motor deficits argue possible basal ganglia interact frontal cortex via thalamus implement selection mechanism described aspects context update point time good example task srn works reber grammar task modeled psychologists implicit learning explored people press buttons corresponding letters appear sequentially screen subject letters followed regular grammar subjects pressing buttons faster faster sequences letters significantly faster sequences follow grammar shown evidence learning grammar implicitly explicitly knowledge grammar asked simple state grammar reber string grammar produced starting start generating letter link followed chosen random probability string ends end reached example string grammar shown figure state fsa state grammar string letters letter corresponding link takes sequence produced follows starts start letter equal probability next chosen letter process going generating letter link continues end corresponding letter reached connectivity regularities present grammar srn learn reber grammar training network predict next letter sequence output network prior input link fsa unique letter task relatively easy input uniquely identify location fsa different same letter kind internal context necessary keep based prior history grammar context layer srn next letter actually chosen random possible best network activate pick random backpropagation network pick items random produces blend possible outputs see detail leabra network pick multiple possible output patterns essentially random take advantage simulation start reber grammar simulation directory see network log windows addition process control overall control panel usual begin exploring network click observe connectivity note particular context layer units single receiving weight hidden units context units connection determine hidden unit update weight random value reasons otherwise network fully connected notice simply display purposes shows possible valid outputs compared actual output lets view activations see trials learning click sure locate process control panel right minus phase beginning sequence pass fsa grammar starts letter context units network produce random expectation letters coming next note noise unit activations helps pick unit possible ones random again see plus phase see possible subsequent letters strongly activated unit indicates letter actually came next sequence network ever learns possible subsequent letters trial chosen random learn possible outputs integrating experience different trials things challenging task learn interesting aspect task network done good possibly roughly errors ends discrete output come next right time reason cause problems learning introduced systematic error signal constantly increase decrease bias weights problem unit active inactive overall net error zero note allowed units active case units active introduce net negative error large magnitude negative bias weights eventually activation output units possible network pick output random allowing somehow means network actual response time actually case hidden layer representation remains essentially same outputs reflecting identity change actual output presented plus phase higher level internal representation possible outputs level lower output representation randomly situation important later consider networks represent multiple items see further discussion order monitor networks performance learning need error statistic zero network learned task perfectly case standard sse new statistic error output unit possible outputs shown layer labeled log displays now continue minus phase next event sequence see now units updated copy prior hidden unit activations verify click show plus phase activations previous event point continue reset sequence want network runs special type environment called dynamically new sequences events epoch whole bunch training examples underlying fsa advance line implements reber grammar fsa takes train network time load trained pre network training log file load network network window select load log file select network take epochs learn problem point gets zero errors epoch range random networks ran trained pre network took epochs zero trained longer epochs total point row representations robust noise network errors extra training epochs amounts different sequences epochs amounts sequences set sequences epochs case faster backpropagation network took sequences epochs scheme able train backpropagation networks larger hidden layers units learn epochs evidence advantage additional constraints model learning inhibitory competition task leabra networks generally learned faster backpropagation networks required larger learning rate now test trained network see solved problem see well letter locate testing process control panel test network sequence letters results shown grid log right note network display updated cycle see possible outputs network producing correct outputs indicated column fact pattern matches pattern due noise better understand hidden unit representations need sequence reasonable length events longer sequences fsa due selecting tell representation individual total number events sequence events counted again find sequence events need find sequences epoch running sequence events press bring cluster plot hidden unit states event interpret cluster plot clusters events zero distance terms hidden states letter labels now back process control panel change produces random sequence letters obviously network capable letter come next lots errors network detector determine string fits grammar sense network fsa structure itself own representations context layer srn provides means immediately preceding context information cases need able learn temporal contingencies time steps specifically need able solve temporal credit problem recall discussion driven error learning credit problem units responsible current error signal temporal credit problem similar events past responsible subsequent outcome see temporal credit problem solved similar way structural form credit earlier context driven error learning based time form driven error learning primary means solving temporal credit problem temporal differences learning algorithm developed based similar earlier ideas model phenomenon reinforcement learning reinforcement learning based idea relatively global reinforcement signals reward drive learning enhance reward avoid kind learning goes conditioning form learning solve temporal credit problem closely related psychological biological relevant phenomena fact recently shown detailed properties algorithm close relationship properties subcortical brain areas start discussion biology reinforcement learning standard algorithm show notion activation phases generec algorithm implement version leabra algorithm relationship standard driven error learning apparent explore simulation learning action primary brain areas appear specialized reinforcement learning midbrain ventral area substantial cortical subcortical areas control firing neurons neurons midbrain areas project neurotransmitter dopamine widely frontal cortex basal ganglia action dopamine likely learning areas things areas provide relatively global learning signal brain areas relevant planning motor control see properties firing neurons consistent temporal differences learning rule midbrain play role reinforcement signal brain areas required control firing signal see key idea reinforcement learning computing future reward complex task performed areas frontal cortex basal ganglia project control midbrain evidence basal ganglia neurons representing reward provided neural summarized evidence areas frontal cortex related structures involved patients lesions area show ability predict future comes control behavior based control distribution midbrain signals basal ganglia frontal cortex shows basal ganglia control firing sends dopamine back entire basal ganglia shows parts frontal cortex control ventral area sends dopamine back entire frontal cortex shows relationship controlling areas midbrain areas dopamine signal case basal ganglia system fairly well established areas called control constitute distinct subset basal ganglia dopamine signal coming affects basal ganglia signal controlled specialized subset areas notion distinct system essential aspect learning framework called established well plausible frontal cortex exist well certain areas play role controlling dopamine signals entire frontal cortex finally data firing properties neurons documented simple conditioning tasks particularly shows start conditioning task neurons fire reward presented experiences repeated trials reliably predicts reward neurons start fire onset seem fire whenever reward reliably early learning reward actually presented learning comes note neurons fire fire reward itself actually pattern firing essential computation performed algorithm see note firing predicts predicts reward called order conditioning neurons learn fire onset new subsequent actual reward now see biological properties discussed provide good fit temporal differences algorithm developed basic framework algorithm reinforcement learning algorithms produce actions environment environment produces upon delayed effects actions goal naturally produce actions result maximum total amount reward come long ways future typically interested future expressed mathematically factor determines ignore future reward obtained time time considered relative event beginning training trial expectation repeated trials note power future time increments gets smaller smaller times further future unless plays role squared sum error sse entropy cross error driven error learning algorithms objective learning called objective function particular function called value function term thought expressing value things point time whereas error driven learning goal minimize objective function goal here rapidly apparent function going difficult value point time depends happens future same issue temporally delayed weve discussed now see showing objective function approach taken algorithm problem divide specifically problem basic components component learns estimate value current point time based currently available information component actions take components map onto components basal ganglia frontal cortex previous section job resort value function different alternative actions order select action perform next right reward receive compared going left todo clear focus sensory estimate value call value distinguish actual value needs learn sensory predictive reward sounds happens conditioning obviously ever reward actually receives propagate reward information backwards time point reliably sensory cue point time next point time adjust current point time words looks ahead time step updates estimate predict ahead look value initially learn predict reward immediately time step reward happens next time able predict reward predict backwards time reward note propagation takes place repeated trials trial error backpropagation procedure error way network point error received means require remember information point reward required order propagate reward information back time point reward see happens mathematically start noting written terms follows same relationship hold estimates now define error tell update current estimate terms ahead look estimate next point time computing difference represented value estimate according current estimate note expected value notation changes slowly time compute expected value hebbian learning rule note equation based notion future reward consistent time time time error signal measure residual said learning able temporal delays notion consistency computes total expected future reward based current stimuli learns weights minimize difference estimate value based time step ahead look last thing specify exactly computed directly external stimuli error signal adapt estimates expect neural network computes based weights representations stimuli potentially processed hidden layers see illustration error train weights network computes treating same error signal squared sum error entropy cross error extent stimulus environment reliably produce correct value network learn function error reliable stimulus exists reward remain stages learning simple conditioning experiment showing error function time shows initial trial reward shows estimate reward gets earlier earlier shows final trial onset completely predicts reward now lets see learning works practice same kind simple conditioning experiment shown case reward fixed time figure shows function time reward coming starting shows happens trial learning large error reward occurs completely refer means note set case weights produce increase value larger next time say effects reduce value next time propagate reward backwards time step equation time includes shows propagation occurs way back finally shows final state network learned propagate further back predictive stimulus earlier time network occurs reward follows provides nice fit neural data shown computing error well explore example further finally need specify learning error signal easily train network produce actions increase total expected reward see lets imagine network produced action time action leads directly reward leads previously increase future positive adjust weights network similar way network increase likelihood action produced again similar possible action time step led greater reward produce larger weight changes weaker reward clear error signal provides useful means training system itself reflected biology fact dopamine signal represents error projects areas control dopamine signal itself areas considered network see noted different learning exist general category reinforcement learning algorithms generally mathematical analysis performed showing algorithm converge correct result particularly important called basically averaged time activation value learning activation value role analogous hysteresis parameter srn context units value parameter represented form parameter case implicitly considering included activations able activation differences implement driven error learning relatively straightforward learning looking consider value represent minus phase activation output unit represents plus phase activation difference phase values error error generec based phase driven error scheme weights rest network automatically updated reduce error standard generec learning phases activation completely introduce learning overall leabra framework issues need addressed regarding way phase values actually computed lets consider happens network experiences reward case plus phase activation equal reward value plus additional expected reward current reward simpler consider additional term zero unit reward value plus phase fact reinforcement learning entire network reset reward achieved new trial referred reward reward assumption reward value plus phase well see moment minus phase now lets consider happens isnt reward obvious thing here allow output unit settle value minus phase representing current future reward state problem plus phase explicit reward future opposite todo provide high level settle plus phase obtain set minus phase value next time step previous plus phase value assuming zero unless plus phase clamped value plus phase exactly detail parameter well consider moment next set plus minus phases current time now set value computed plus phase estimate next time steps value network actually value expected future reward next time step current relationship time steps based phase algorithm plus phase unit predicts value function next time step consider unit phase ahead stimulus represented forward units relative stimuli think plus phase actually minus phase next time step learning based phase difference stimulus activations previous time step performed context standard generec driven error learning based phase scheme learning effectively offset phase illustrated output unit settling plus phase time computes expected value next time step experiences actual outcome form reward figure clear coherent think plus phase unit minus phase next time step consistent idea minus phase next minus phase actually plus phase time step unit clamped value computed unit computing based stimulus activities time error updating actually computed next time step weight computation sending activations time error minus plus phase difference time shown important note time based phase implementation aspect algorithm requires future states adapt prior estimates contingencies time step allows network propagate information future back time implementation explained here biologically plausible combining context representations srn model explain further error signal control context representations updated context representations simplifies issues time discussed version algorithm actually think brain implementing biologically plausible otherwise seem finally note implement parameter set minus phase activation reflects idea actual value computed prior plus phase actually activation computed settling typically value simplifies things considerably appropriate isnt problem future states point receive next reward well discuss effective greater delayed achieved simultaneously performing learning multiple time scales explore learning rule based phase implementation described simple conditioning task discussed network learn stimulus reliably predicts reward stimulus reliably predicts need algorithm context nature stimulus representations network recall said delta rule aka rule provides good model conditioning needed issue timing timing stimulus relative response fact rule equivalent delta rule happens time step match sensitive timing relationship importantly purposes modeling timing provides particularly clear simple demonstration basic properties learning problem simple demonstration involves representation timing basically stimulus representation distinct unit stimulus point time unique units weights learn representation complete serial proposed see exactly works look model imagine brain involves error signal control gating stimulus information active memory context representation srn stimulus identity setting set neurons range different activation different time periods subset timing units active point reward learning come represent expected delay stimulus onset reinforcement well explore type mechanism further simpler time open project lets start examining network input layer contains units represent different stimulus columns represent points time single unit receives weights inputs lets see works action locate process control panel happen stimulus reward present monitor time steps value displayed bottom network view time step sequence events simulator continue steps input activation represents fact stimulus stimulus row came continue active time steps notice stimulus unit activated reflects fact reward received phase plus activation unit clamped reward value here now lets see reward weights zero begin click unit notice weights increased unit representing stimulus last position went reward caused unit minus phase plus phase updated weights based sending activations previous time step discussed previous section seen graph log minus plus phase difference unit function time step clearly shows continue goes back maintaining reward active end entire sequence now switch back again recall weight unit increased activation unit expected due thresholded nature units click continue trial shown bottom network time step due weight changes previous trials weight unit now strong activate threshold look graph log see now positive time step network now reward time step earlier effects click note weight previous time step now increased result positive lead trials reward earlier now reduced magnitude now click back network let process play process control panel see forward weights graph log ultimately resulting activation unit stimulus comes same process shown represents algorithm point standard phenomena conditioning explored model well look order conditioning occurs stimulus longer predictive reward ability predict reward appropriate simulate simply turning reward appears locate overall control panel contains parameters determine nature stimulus input reward controls look field controls timing stimulus representing time stimulus comes long fields provide variance points zero explore own later reward known conditioning stimulus view timing parameters stimulus mechanism control probability coming want manipulate control probabilities contained field button press select indicates stimulus presented now graph log trial happened point reward occur explain happened equations network describe occurs next terms error signals plotted graph log explain network done learning again stimulus expectation reward thing procedure weights reduced back zero reduced bring unit threshold effects threshold applicable real brain appears unit constantly active low level additional inputs driving resting potential threshold effectively closer simulation expect weights reduced bring unit threshold behavior suggest complete least situations kind threshold effect work now lets explore order conditioning network stimulus association press onset stimulus clearly driving expectation reward note happens faster due now turn stimulus starts time steps see field control panel selecting already trial back look weights essentially stimulus acts reward positive allows stimulus learn predict stimulus stimulus point feel free explore parameters available see network responds note change parameters sure press button order new environment based new parameters final regarding limitations representation example learn conditioning order saw properly configured representation allow learned question time zero right point trial properly finally requires stimulus manipulate last problem points important issue algorithm learn temporal requires stimulus representation support see problems resolved allowing system control updating context representations form ltp mechanism cortex based nmda receptor allows calcium ions enter synapse response conjunction presynaptic neural activity excitatory neurotransmitter glutamate postsynaptic activity sufficiently excited membrane potential calcium synapse triggers complex sequence chemical events ultimately results modification synaptic efficacy weight available data suggests pre synaptic neurons strongly active weight increases ltp due relatively high concentration calcium weaker activity results weight decrease ltd due lower concentration calcium called associative hebbian learning important model learning addition important way calcium low concentration lead ltd transient neural activity turns important biological mechanism driven error learning taken biology ltp suggests combination hebbian driven error learning mechanisms work consistent computational advantages combination goal model learning develop internal models information present world difficult relatively limited relevant information via senses possibly represent information experience appropriate biases world order organize experiences simple parsimonious models critical successful model learning strong bias towards representing correlations appropriate reflect reliable stable features world parsimonious representation correlations involves extracting principal components features dimensions correlations simple form hebbian learning perform principal components analysis pca modified fully useful importantly individual units represent principal components subset input patterns implemented network property inhibitory competition described previous chapter results distributed representations informative principal features features sub input unit effectively representing strongest principal component input items best adapted specialized represent biological processes accommodation sensitization important different types input patterns well represented appropriately specialized units hebbian learning number ability learn produce particular output patterns function particular input patterns seen simple pattern associator task error target output pattern actual output pattern drive learning networks learn successfully idea delta rule error weights function derivative error order learn complicated output input relationships functions networks least intermediate hidden layer necessary networks require generalized form delta rule error signals chain rule propagate error derivative back multiple layers units evidence cortical neurons communicate error signals directly possible difference activation states phases compute essentially same error derivative activation phases corresponds expectation production particular output pattern corresponds experience actual outcome long network bidirectional weights difference activation states units network reduce error layer idea generec algorithm seen generalization earlier algorithm order biological synapses compute weight changes necessary algorithm perform ltd expectation state unit outcome state ltp otherwise assuming relatively rapid expectation outcome activation phases expect ltd transient expectation consistent biology ltd see expect ltp sustained activity expectation outcome phases according purely driven error learning rule biological mechanism suggests ltp occur case biology resolved assumes hebbian driven error learning taking place sound functional reasons believe hebbian model learning driven error task learning taking place cortex see later chapters types learning required account full range cognitive phenomena considered further model learning provide important constraints biases development representations otherwise purely task learning context understood terms variance bias well known phenomenon statistics result representations encode important statistical features activity patterns same play role solving particular tasks network perform viewed way task learning seen important way conditionalizing model learning combination appears produce better performance cases compared simply shaped task demands better performance based pure model learning relevant solving tasks sequence temporally delayed learning require proper development maintenance updating context representations specify location sequence appropriate factor subsequent outcome biological mechanisms facilitate forms learning including connectivity cortex basal ganglia thalamus modulation learning neurotransmitter dopamine specialized aspects frontal pre cortex pfc computationally mathematical framework reinforcement learning useful understanding type learning work provides possible means understanding important roles reward motivation learning computational mechanism active gating useful controlling maintenance updating context representations dynamic manner required complex tasks biological evidence gating mechanisms operating pfc locations book last chapters competitive learning lots part text presents models cognitive processing taking place cortex brain areas building upon principles basic mechanisms developed part text chapter overview general function large scale organization brain areas presented focus functional computational observed different brain areas emphasize specialization understood common principles hold areas goal chapter provide useful coherent framework specific models subsequent chapters related framework supported existing data reflects common thought number researchers years remains certain aspects suggest reader view follows broad framework take established fact central organizing principle framework notion achieved achieving achieving tradeoff different functional tradeoffs identified specialization understood means separately otherwise system try achieve turns mechanisms principles developed part play important role tradeoffs help explain aspects large scale organization brain begin brief summary general functional computational principles underlie cognitive models provide brief overview different anatomical areas cortex relevant aspects subcortical describe functional organization areas terms following specialized systems posterior perceptual motor cortex pmc frontal pre cortex pfc hippocampus related structures hcmp organization based underlying tradeoffs rate learning nature resulting representations cortex pmc pfc slow integrative integrating instances hcmp fast keeping instances separate ability maintain representations active state delays face interference ongoing processing active maintenance pfc appears specialized active maintenance plays important role active memory controlled processing executive control cognition taken common underlying mechanisms broad provide framework cognitive architecture different cognitive phenomena explained terms interactions specialized systems addition common principles mechanisms applicable areas finally end chapter address number general problems arise framework divide discussion general properties cognition structural dynamic aspects structural aspects describe ways information processing system determined overall patterns connectivity relationships representations different levels stages processing dynamic aspects describe nature processing time determined activation flows processing levels achieves useful overall outcome number reasons detailed follows thinking processing information cortex generally levels hierarchical fashion different specialized pathways operate emphasize different aspects overall sensory input motor output intermediate processing number benefits rich interconnectivity different pathways number different levels think completely distinct parallel pathways highly interconnected clear processing information embedded same underlying neural distributed potentially wide range different processing pathways number important consequences elaborated upon subsequent sections begin considering basic building model cognitive architecture detector neuron presented here individual neurons viewed relatively stable representations detect difficult define complex set conditions inputs saw taken layer detectors perform transformation input patterns emphasizes distinctions patterns saw transformations shaped learning represent important structural statistical properties environment enable tasks solved cognition viewed hierarchical structure see sequences layers transformations operating sensory inputs ultimately producing motor outputs responses useful internal states provide interpretation environment important subsequent behavior discussed sensory input contains large low quality information highly processed sensible responses interpretations example sensory signals viewing same object different directly common overlapping activations typically sense interpret object same potentially elaborate sequence transformations sensory input stages detectors neurons distinctions collapsing form useful generally abstract internal representations invariant differences input selective sensitive same process performing transformations emphasize dimensions aspects collapse operates levels processing example representations underlie words emphasize features properties define word collapsing irrelevant ones example relevant feature physical size irrelevant whereas notion reality central affect generic hierarchical system showing specialized pathways streams processing pathway connectivity important consequence hierarchical structure existence specialized distinct processing pathways streams necessary layer processing hierarchy requires specific types transformations performed previous layers accomplish particular job addition typically potential transformations input irrelevant transformation sense group relevant transformations coherent stream continue example visual object recognition layer processing visual representations letters digits appropriate categorical representations digit needs processing stream provide already invariant respect changes location image retina spatially invariant visual information transformation need possible digit location retina lead generalization knowledge novel retinal locations transformation necessarily require information useful described sense similar transformations visual stimuli overall visual form object processing stream appears cortex done same types specialization seen operating types hierarchical structures example functions typically divided different specialized processing streams details division typically details lot division level higher division developed summarizing level lower details levels analysis completely separate hierarchical structure effective flexible different specialized processing streams levels now connections lower levels mutually constrain processing different pathways better deal partial noisy novel particularly complex stimuli example explore idea visual form pathway interacts levels spatial processing pathway important result attention focused spatial scales depending confusions arise visual form pathway confusions level occur things way top hierarchy system deal rapidly appropriate level detail further general hierarchical structure comes higher levels processing pathway areas likely receive input different pathways own whole hierarchical notion here see better association areas associate level higher association areas collection perform constraint satisfaction processing described principles converge idea knowledge associated item distributed widely number different brain areas specialized processing pathways level higher association areas illustrates version general idea proposed ones representation item distributed different specialized systems semantic memory system evidence generally distributed model notion distributed representation large scale similar notion fine distributed representations discussed terms processing taking place layer pathway again kind nature brain scale small structure larger scale widely distributed view knowledge computer metaphor tend favor idea single representation knowledge item features stored convenient location bias towards assuming representation leads problems discussed last section chapter show distributed model problems important properties general structure knowledge processing processing transformations performed specialized dedicated systems neurons transformations reflect experience knowledge via learning transformations applied situation determined directly specific stimulus activation pattern processed easy system treat different stimuli specific properties consequences saw specialization important layers integrated elaborate processing stages subsequent stages come depend particular types transformations input layers turn reliably provide specific transformations subsequent layers representations enables rich set specific content associations built time connectivity different representations compare situation standard serial computer processing data knowledge explicitly separated processing typically operates whatever data arguments function advantage system relatively flexible function need written wide range different situations ability arbitrary variable binding case arguments function important flexibility well return later obvious advantages flexibility difficult treat different stimuli specific properties consequences resort sequences constructs elaborate representational structures clear exactly different stimulus compared basic tradeoff here specificity knowledge hand flexibility appears brain latter interesting view precisely deal type specific content real world knowledge led traditional computer metaphor based models human cognition seems case getting details right practical differences trees important world kind flexibility provided arbitrary variable binding demonstration point domain language provided following set due time specific world real knowledge necessary produce different interpretations words see flexibility favor specificity causes number problems resolved different ways discussed dynamic level view cognition result activation propagation interconnected hierarchical structure bidirectionally connected processing layers described via multiple constraint satisfaction property bidirectionally connected networks described network tend produce activation state response interpretation constraints imposed upon inputs learned weights stimuli familiar ones result straightforward rapid settling network relatively optimal state response stimulus require extended appropriate activity state case resulting activity state typically same time same stimulus presented due number factors learning sensitization importantly due influence maintained activation states prior processing provide additional constraints settling process internal states aka internal context dynamically interpretation response stimuli order provide coherent consistent set responses time simple example note following produce different interpretations word words preceding different internal context representations subsequent interpretation addition multiple constraint satisfaction amplification dynamics described pattern completion bootstrapping mutual top support etc play important role processing role mutual support providing form active memory elaborated further following section plays particularly important cognitive role relatively simple mechanistic basis particularly important aspect contribution inhibition described cognition form attention following section consequence bidirectional excitatory connectivity allows different representations mutually support important enables representations remain active absence derived excitation viewing stimulus rely internal excitation mutually supporting representations remain active time viewed form memory enables information time well call active memory based weight memory results changing weights new information see distinction active memory typically lasting long based weight memory active neurons ongoing processing active memory distinct advantage directly influence ongoing processing areas providing internal context described whereas based weight memories directly affect units weights see limits mutual support providing active memories active maintenance mechanisms needed robust flexible active memory system mutual support provides basic underlying mechanism main form active memory brain areas lack active maintenance mechanisms inhibition operating levels processing cortex discussed natural limitation amount activation number things simultaneously represented rise phenomenon attention aspects sensory input internal context ignored favor attention cognitive psychologists view attention discrete separable mechanism view emergent property constraint satisfaction limits inhibition similar described external environment internal context determine ignored further levels processing constrained inhibition levels mutually influence varying degrees via bidirectional connectivity attentional effects arise level abstraction important consequences processing levels functional level attention critical level processing representations different levels processing focused same underlying thing set things important solving problems discussed integrated basic principles mechanisms developed part book relatively coherent general framework thinking structural dynamic properties cognition now turn functional different brain areas relatively large scale analysis emphasis cognitively relevant brain areas including cortex parts system brain areas cortical associated functions human cortex generally organized contain number related specialized processing pathways level higher association areas general nature specialized functions described illustrated well following standard terminology describing locations aka upper aka ventral lower posterior aka towards back aka towards lateral towards towards middle lobe specialized visual processing area central posterior region receiving main visual inputs thalamus areas performing higher levels transformations well discuss areas detail lobe specialized functions including primary higher level perception regions level higher visual form object representations posterior regions see language processing lateral regions see term longer representations events regions regions hippocampus play important role rapid learning arbitrary information see obviously important relationship language speech perception contribute lobe specialized functions including spatial processing representing things located space see task specific perceptual processing organizing tuning reaching via visual perception regions primary level higher processing areas spatial nature consistent emphasis spatial processing areas temporal lobe important language left lobe specialized maintaining representations active state active maintenance executive control processing areas posterior regions contain primary level higher motor output representations consistent executive nature lobe group subcortical brain areas surface cortex called system consists principally hippocampus cortex contains thalamus areas mutually interconnected originally thought process information useful group system now known different individual roles better understood terms relationship cortex example hippocampus temporal cortex important rapidly learning new information different types interconnected temporal cortical association areas viewed top cortical hierarchy virtue connectivity located frontal cortex appears important tasks frontal cortex specialized including motor control action selection areas thought specialized cortical areas areas thought primarily specialized processing cognitive roles components system well documented thalamus subcortical area specialized provides sensory input cortex plays role attention provides source feedback basal ganglia group subcortical areas appear important sequence learning motor control see subcortical area important motor control timing representations cognitive role well finally midbrain small groups cells play important role state cortex relatively global fashion see systems control cortex play important role cortical regulation self controlled processing now attempt provide principled framework understanding different characteristics brain areas described based functional tradeoffs characterized systems brain areas tradeoff rate learning way interacts knowledge representations ability perform active maintenance delays face stimuli interacts ability graded distributed representations interconnectivity brain systems follows system posterior perceptual motor cortex pmc consists occipital temporal parietal motor sensory areas cortex areas directly responsible sensory inputs producing motor outputs higher level association areas serve integrate activities system frontal pre cortex pfc large pmc humans considerably smaller lower consisting frontal cortical regions forward motor cortex pfc appears specialized active maintenance information time particularly useful controlled processing responses mediated specific task constraints simply automatic responses stimuli system hippocampus related structures hcmp appears play critical role rapid novel information effect learning rate ability weight represent underlying conditional probability input unit activity output unit activity cpca hebbian learning objective conditional probability reflected case weight training example binary impossible represent overall probability apparent number individual examples functional analysis begins assuming cortical systems pmc pfc learning mechanisms described order develop representations important underlying structural statistical characteristics world effectively process perceptual inputs produce systematic useful motor outputs learning cortex necessarily slow order integrate individual experiences extract general underlying regularities environment shows simple example point taken exploration described slow learning rate enables weight converge actual underlying conditional probability event occurring environment somehow world provide underlying statistical regularities experience able away faster learning rate experience typically small noisy overall picture slow learning order blend note virtue integrating previous ones unique details specific remaining survival world demands rapid learning occur specific arbitrary information important saw family enter last notice rapid form learning memories individual separate integrating confused learning fast slow integrating basic tradeoff demands slow integration hand rapid seems brain resolved tradeoff allowing cortex learn slowly integrate experiences hippocampus provides strengths cortex complementary rapid learning system idea consistent large amount data people hippocampal lesions example known large temporal cortex including hippocampus removed prevent subsequently unable learn new information people events occurred etc able learn number relatively complex motor perceptual tasks showed normal term long priming effects recently read words example subsequently read rapidly preserved learning explained terms results slow integrative cortical learning mechanism ideas detail different type tradeoff understand difference pmc pfc time involving active memory maintenance information learning rate described active memories supported bidirectional connectivity important memory processing brain areas active memories conjunction highly overlapping distributed representations problems occur specifically distributed representations rely critically input order select appropriate subset distributed components activated absence selection information maintained system itself pattern overlap representations cause activation spread result original activity pattern illustration active maintenance via bidirectional excitation distributed representations value excitatory weights enable appropriate subset features maintained activating representations independent maintenance problem semantic features maintain level higher items terminal problem illustrated distributed representation encode different items terminal share total features monitor recall explored example inhibition needed prevent spread excitation overlapping distributed representations inhibition distributed representations active memories original input stimulus representations activity pattern difficult maintain unique subset features active activating activation spread unit via connections necessary maintain constructed avoid problem particular forms linear non inhibition weight modulation robust noise require further control mechanisms alternative shown panel figure independent isolated representations maintain themselves isolated representations lack distributed pattern overlap important representing similarity enabling based similarity generalization novel inputs easily maintain information active state similar representations explore phenomenon active memory highly overlapping distributed representations important again suggests tradeoff seems clear pmc areas require overlapping distributed representations order represent process complex visual perceptual inputs contrast large amount data supporting notion pfc uniquely capable robust active maintenance sense analysis pfc independent isolated representations isolated character representations function interconnectivity easily measured firing properties neurons solid evidence issue exist evidence discrete isolated connectivity pfc described data obtained idea remains basic point pfc perform robust active maintenance way pmc relatively well established finally particularly important additional advantage isolated representations pfc allow arbitrary combinations representations activated interference due kind combinatorial flexibility likely important factor achieving powerful solving problem executive pfc further insight specialization pfc active memories comes basic tradeoff ability maintain active memories face interference ongoing processing update rapidly new information needs maintained extent units simple bidirectional excitatory connectivity active memory interference bidirectional excitatory connections stronger prevents rapidly new representations activated existing ones updated inputs conversely weaker excitatory connectivity units sensitive inputs capable rapid updating enable sustained face interference tradeoff think pfc taken advantage midbrain systems provide gating mechanism controlling maintenance opened pfc representations sensitive inputs capable rapid updating closed pfc representations interference idea developed further diagram key properties principal brain areas active representations shown highly overlapping distributed representations overlapping non isolated separated weights active units shown solid lines active non units active representations shown thought feature values separate dimensions modalities pmc representations distributed embedded specialized specific processing areas pfc representations isolated combinatorial separate active units representing feature value areas pfc units capable robust maintenance self indicated recurrent connections hcmp representations sparse separated distributed conjunctive single representation active time corresponding conjunction active features central characteristics different brain systems described summarized form basic picture pmc learns slowly form integrative distributed overlapping representations interconnectivity exhibit short term active memory relatively easily new stimuli pfc learns slowly isolated representations dynamic regulation mechanisms enable maintain active memories longer delays face new stimuli finally hcmp learns rapidly form separated representations minimize interference similar memories taken systems basic principles earlier constitute description cognitive architecture brain aspects architecture general character well supported available data cognitive architecture described provides framework explaining cognitive phenomena explored detail subsequent chapters useful point elaborate issue cognitive control controlled processing ability adapt behavior different task demands generally act world issue obviously central understanding human cognition represents major problem computational models understand mechanistic basis cognition generally issues greater detail provide brief summary basic ideas here help models help elaborate important aspects general cognitive architecture controlled processing described contrast automatic processing thought involve limited capacity attentional system theories actually continuum controlled automatic processing view controlled processing graded defined extent certain dimensions task performance required particular controlled processing ability adapt behavior demands particular tasks processing relevant task information sources competing information relevant task behavior otherwise responses ways hcmp pfc contribute automatic versus processing controlled distinction bias provided pfc perform sustained processing facilitate processing learned weakly relatively tasks serve processing different areas binding provided hcmp rapidly learn information necessary perform novel tasks processing controlled processing involve contributions automatic processing performed independent aspects controlled processing terms hcmp binding pfc biasing illustrates central ideas account based directly pfc hcmp extent controlled processing task defined extent following conditions exist sustained learned weakly relatively processing required novel information rapidly stored extent automatic processing defined relative absence factors consequence robust dynamic active memory properties pfc relatively unique position provide sustained top activation bias processing posterior cortex brain systems dynamically update active representations according relevant task pfc bias processing rest system appropriate sustained pfc activation enable sustained focus extended temporally tasks additional top activation pfc facilitate learned weakly processing pmc otherwise dominated well forms processing task highly reading word relatively color pfc biasing processing different pmc areas different areas mutually constrain influence via located pfc representations hcmp hand contributes ability learn novel information rapidly interference typically binding familiar elements novel combinations relevant particular task particular combinations stimuli specific task intermediate states problem solutions etc think combination factors bias binding account distinction controlled automatic processing essentially automatic processing occurs via activation propagation pmc connectivity controlled processing activation propagation introducing dynamic relevant task constraints pfc hcmp appear account places large control pfc true control complex satisfaction constraint process interactions relevant brain areas likely specialized learning mechanisms associated dynamic regulation systems control gating updating pfc representations play important role point evaluate ideas implemented models exhibit relevant behavioral properties models described subsequent chapters constitute important building blocks add overall resulting system context issues cognitive control worth noting different systems cognitive architecture differ extent representations influential wide range areas opposed embedded specific processing areas influential view difference principally relative position overall hierarchy representations described hierarchy effectively defined far removed area direct sensory input motor output shows further removed areas located respect overall network connectivity resulting greater terms pfc hcmp top hierarchy influential general areas pmc essential emphasize view graded continuum distinction further pmc rich lateral connectivity areas same general level abstraction least levels sensory motor processing diagram purely hierarchical connectivity separate processing streams showing higher levels hierarchy virtue located overall network explicit equivalent standard view imagine actual connectivity purely hierarchical effect nonetheless relevant relate issues psychological distinctions explicit versus implicit versus view experience reflecting results global constraint satisfaction processing brain areas representations influential process greater general means highly influential areas pfc hcmp tend experience embedded areas pmc areas clearly associated explicit processing pmc subcortical areas associated implicit processing distinction add important big interactive system distinctions considered continuum notion assumed brain composed separable computations pass results next address number general problems arise general functional principles described people tend neural networks exhibit kind problem somehow models cognition example found case catastrophic interference found generic neural networks interference simple sequential list learning humans led conclude neural networks good models cognition showed actually tells important way brain works helps sense different kinds memory systems cortex hippocampus described important emphasize cases problems actually reflect documented limitations human cognition taking kind analysis approach argue human cognition perfect suggest reflects number tradeoffs fact neural network models seem provide useful insight nature human cognitive limitations real strength approach following problems lack flexibility resulting dedicated specialized representations indicated earlier tradeoff specialization flexibility dimension appears brain generally knowledge benefits specialized representations challenge problems understand measure flexibility emerge context system specialized dependent knowledge representations note largely examples based visual perception issues generalize aspects cognition multiple items illustration binding problem encoding terms separate features leads multiple items present input here red circle green square present input same representation activated green circle red square system know present illustrated representation commonly problems neural networks known binding problem arises whenever different features stimulus represented different underlying representations distributed representations multiple items need represented lets imagine set units encode color information red green blue set encode shape information circle square binding problem arises present red circle green square system know circle red square green way see illustration words system bind separate features applying same object see attention provides way problem place able restrict visual input image produced object time whatever activation results processing input automatically bound same object solution binding problem representations encode combinations input features color shape achieve greater representing multiple combinations show features objects feature separate representation shown columns show responses set representations encode separate features combination different conjunctions shown top column last red green blue square circle way binding problem represent feature information separate manner representations encode conjunctions different feature values red circle obviously impossible represent possible combinations way units required realistic numbers features individual units represent combinations conjunctions separate feature representations overall pattern activity units uniquely combination binding features present input shows works single additional conjunctions combination unit required additional unit responds red circle green square blue cases network otherwise confused basis separate features total number units required separate features conjunction unit units needed encode possible conjunctions scale problem advantages apparent example shapes conjunctive units required needed features plus conjunctions combinations scheme described here feature units conjunctive units different type solution binding problem suggested dynamics processing different feature elements idea type activations features same object encode binding information available evidence observed firing actually binding important observed likely natural consequence simple activation propagation spiking neurons brain neurons communicating tend drive spike roughly same time central problem brain achieve necessary kinds processing presumably taking place bound representations substantially addressed feature transient dynamic systems elaborate system dynamic processing established process binding information example unique consequence set associations red apply green blue information associated representation exists relatively temporal problem arise unique pattern activity set dedicated representational units case combination conjunctive representations scheme described finally essential cases people fail solve binding problem successfully provide important underlying representations involved example well known combinations visual features requires slow serial processing found fast parallel speed details complicated phenomenon consistent idea sequential mediated attention binding needed cases presumably sufficiently conjunctive otherwise distinctive representations relevant combinations extreme case binding problem occurs multiple instances same item clearly possible distinguish instances basis centered object features color shape least ways actual number items present accurately represented sequential application attentional mechanism items turn combined kind mechanism result appropriate representation situation unless items presented same location spatial location representations case item present case multiple instances same item important take account elements representations considering problem problem related representation multiple items arises whenever compare different representations attention share time same representational space different items work need actually compare items represented same time idea natural network dynamics inhibition pattern overlap result representation items common basis comparison recall exploration presenting multiple digits same network possible compare visible stimulus stored representation stimulus well visible stimulus activity pattern fits pattern weight values encode stored stimulus speed network settles see resulting activation state measures assess general fit stimulus stored assuming stored pattern obviously closest visible measure tell close idea posterior pmc frontal pfc representations comparison pfc thought redundant representations content pmc maintain information active form hold representation items represented posterior system viewing item comparison based overlap kind measure indicates consistent representations think large number different levels representations produced item cortex modalities different levels abstraction possible same idea implemented activating different pmc representations different items mutual consistency idea pfc somehow involved consistent existence same capacity limitations relative active memory number possible solutions problem aware implemented models demonstrate actual ideas emphasized notion structural hierarchy representations increasing levels abstraction forms hierarchical structure world need represented systematic fashion issue comes things components themselves component example face composed components eyes nose same time component body imagine spatially invariant representations visual aspects objects invariant representation nose structural relationship information parts components relative position nose face invariant representations invariant respect context object appears separate relationship representations somehow bind invariant nose representation containing face representation implemented neurons general solution problem imagine invariant representations active higher level system long representations share same units level lower visual feature representations encode structural information via same kind limited style conjunctive representations suggested solution binding problem well explore role kinds representations object recognition shared representation nose face invariant nose extent build same level lower features face representation conjunctive information relationship nose rest face invariant nose representation away context likely sequential attentional mechanisms focus different aspects objects state visual system focusing face whole different focusing nose higher levels system maintain invariant representations integrate sequential attentional lower levels attention way emphasizes information necessary process current focus attention focus nose emphasizes level lower features encode face encode nose job away context easier happens example face recognition particularly interesting issues appears representation face sum representations individual parts appear encoded objects challenge dedicated specialized representations problem same type processing different type processing needs performed middle processing step obtain result needed processing proceed serial computer current set state variables simply onto stored memory possibly same called appropriate arguments easy data processing separable particularly case same type processing performed different data previous data issue arises processing embedded bit imagine overall processing sentence composed number process turns tell people easily indicating limitation exactly type processing order model limited appear exist human cognition specialized systems rapid learning hippocampus active memory frontal pre cortex begin address issues aspect human cognition remains developed dedicated specialized representations appropriately recognize novel inputs generalization produce novel outputs problem major limitation neural network models cognition important point keep mind people actually particularly good knowledge learned context novel context clearly capable significant amount generalization important means achieving generalization learning associations appropriate level abstraction example lets imagine learns consequences visual image corresponding run away actual visual image retina clear learning took place level lower based representation generalize well subsequent situations image appear novel retinal location learning took place abstract spatially invariant representation images location retina trigger appropriate response same argument applies different levels representation system learn same similar representation instances appropriate generalize generalization problem think cortex organized according rough increasingly abstract representations abstraction solution generalization plausible further likely learning automatically tend form associations right level abstraction example invariant representation predictive correlated things associated whereas level lower representations learning task model sensitive automatically associations finally way thinking generalization terms distributed representations capture similarity structure novel item previously learned items here novel item represented terms combination known distributed features learned previously associations features provide basis correct responding novel item neurons naturally perform weighted average feature associations produce response reflects appropriate balance influences individual features see examples chapters follow interesting note problems caused set assumptions regarding nature representations involved example binding problem problem long representations generalization fine long knowledge encoded proper level abstract representations important question representational assumptions rise problem place finally emphasize solid cognitive architecture result powerful human cognition say properties sufficient simulate wide range cognitive phenomena challenging issues remain solved full scope human cognition understood mechanistic ultimately biological level chapter explore models visual perception level cortical representations abstract level high spatially invariant object representations demonstrate attentional effects emerge levels representation interactions different specialized processing streams focus studied upon senses similar principles likely apply senses discussed perceptual processing sequence transformations emphasize aspects perceptual input collapsing important learn visual system subjective experience constructed wide processing areas specialized extent particular aspect dimension visual world shape color motion depth location general structural principles specialized hierarchical processing streams discussed relevant here focus primarily pathway emphasizes object identity information aspects input irrelevant identity information examine role pathway emphasizes spatial location information interacts object pathway producing complex pattern spatial based object attention principles derived exploring subset phenomena apply generally aspects visual processing biology visual system obviously begins eyes completely clear ends increasingly abstract levels visual processing gradually cognitive processing considerable evidence visual brain areas activated person thinking semantic properties see section cover brain areas widely visual processing areas starting retina going next lateral lgn thalamus primary higher visual cortex occipital lobe etc continuing parietal temporal amount known visual system particularly early stages processing able provide sketch main findings here objective provide sufficient orientation empirical models described chapter focused presentation aspects visual processing particularly important object recognition aspects important processing motion color surface depth etc visual properties play role object recognition people good recognizing objects simple line indicates basic shape form information least sufficient object recognition addition representing basic form object general shape key properties necessary object recognition spatial invariance object recognized spatial location size etc models explore location size invariance best documented terms underlying neural representations well see visual system seems gradually build invariance subsequent stages processing retina light provides relatively highly processed signal cortical visual areas important understand nature processing models cases performing appropriate form processing pre actual images directly providing input representations roughly capture effects retinal processing general retinal processing performs contrast enhancement emphasizes places visual signal changes space tends enhance edges objects coding absolute values regions relatively constant beneficial result processing greatly representation visual scene pixel informative edges represented complex responsible retinal processing starts sensitive light cells selective color turn electrical signals subsequent stages retinal processing combine electrical signals local regions perform types different regions specialized specific retina mechanisms provide crucial contrast enhancement effect finally neurons provide output signals retina thalamus described next section center receptive fields computed retina bottom shows dimensional picture receptive field central surround regions upper show center dimensional receptive fields showing broad surround field central field combined form overall receptive field computed retina involve surround center receptive field receptive field neuron generally refers spatial distribution retina inputs light affect firing neuron later term generally refer set inputs activate neuron picture surround center target central region surrounding shows main categories surround center receptive fields center neuron active central portion receptive field surrounding portion center neuron active center surround illustrated figure center constructed tuned surround region tuned center region versa vice center individual tuning functions modeled gaussian normal shaped distribution function resulting surround center field called difference consider happen region light covered entire receptive field center neuron excitation inhibition leaving net effect light excitatory center center neuron inhibitory surround net excitation conversely light inhibitory center center neuron compared excitatory surround excited receptive field properties lead effect mentioned retinal output neurons cells fire change levels receptive fields constant focused coding retinal neurons center coding different color well different categories retinal cells identified meaning big meaning small generally speaking cells receptive fields color selectivity better motion selectivity better contrast sensitivity low light small differences compared cells high resolution small receptive fields better color sensitivity think cells uniquely contributing form processing cells motion processing correct types participate kinds processing participate varying degrees see subsequent sections basic retinal receptive field properties provide useful building blocks subsequent processing areas thalamus generally brain different sensory signals cortex places visual thalamus called lateral lgn information retina visual cortex brain area information resources retinal outputs directly visual cortex place increasingly people finding thalamus responsible number important forms processing existence fact remains basic center coding information retina relatively intact visual cortex complexity structural information visual signal reflects structure visual scene thalamus appears primarily concerned dynamic aspects information example good evidence certain neural lgn responsible temporal tuning properties neurons play important role motion perception dynamic aspect processing attention function comes large number projections back visual cortex back thalamus according estimates going forward projections factor projections back generally thought play role controlling attentional processing performed thalamus aspects regions visual scene dynamically focused thalamus visual input relatively structure uniquely suited implementing kind competitive activation dynamics entire visual scene result attentional effects similar argument regarding attentional competition modalities sensory modalities thalamus potentially further thalamus thought important levels versus controlling sensory input cortex explore models visual attention chapter models based mediated attentional processing attention similar principles likely apply focusing motion processing means models capture contributions lgn take advantage organization lgn organize center inputs models lgn different layers center cells different layers inputs coming eyes depth information reflected models next major processing area visual stream primary visual cortex known area back occipital lobe information subsequently processed subcortical brain areas well area builds input producing richer set representations provide basis subsequent cortical processing focus representations capture complex useful aspects visual form string center receptive fields represent edge surface goes kind edge represented dimensional shows lower values upper left higher values lower right considering center coding scheme provided retina lgn expect world points light combine basic receptive field elements represent basic building blocks visual form edges edge simply roughly linear separation region relative light showed neurons called simple cells encode oriented edges bars light proposed explain edge detectors constructed set lgn surround center neurons recent evidence consistent model edge detectors sense edges provide relatively way representing form object assumption region edges relatively visual system capture form information suggest visual system types neurons encode things color surface coding neurons summarizing surface properties region space well focus primarily visual form information encoded edge detector neurons imagine different kinds edges visual world edges differ orientation size aka spatial frequency low frequency means large high frequency means small position going light light light light light edge detectors exhibit sensitivity tuning different values different properties edge detector respond edge particular orientation size position responses properties optimal tuning example coarse coding visual properties different types edge detectors surface coding neurons dimensional visual cortex according topographic organization level neurons roughly organized according retinal position encode thought dimensional map organized according retinal space map number neurons encoding positions sensitive resolution high area center visual field scale large map surface coding neurons oriented edge detectors separated surface neurons structure called blob region edge detectors found region edge detectors appear organized according orientation neighboring neurons encode similar orientations relatively orientation found moving direction interesting occurrence orientations represented circle neurons middle circle orientation clearly coded extensively studied form topographic arrangement columns neurons respond input columns important depth coding present remains possible well see simulation key properties edge detector neurons topographic organization emerge cpca hebbian learning kwta activation function neighborhood interactions neurons diagram visual system showing pathway ventral pathway solid lines indicate full visual field projections whereas lines visual field adapted contains further details visual processing appears separate major streams initially described terms ventral pathway processing object identity information pathway processing spatial location motion information ventral stream goes lower part cortex occipital lobe temporal cortex stream goes upper part cortex occipital lobe parietal lobe reality brain areas bit complex simple story example likely pathway plays role visual information motor actions involves spatial location information certain kinds visual form information clearly pathways considered isolated processing well explore idea models described later chapter point thought distinction present retinal cells terms neurons now known types cells project processing streams likely cells play relatively unique role encoding motion information goes processed area called generally considered part processing stream now provide brief sketch processing streams ventral pathway representing visual form information object recognition thought stages lead increasingly spatially invariant representations addition complexity form information encoded representations increases stages receptive fields larger next area called appears contain number regions called specialized neurons emphasize different aspects visual information form edges surface properties color motion emphasized critical difference representations neurons exhibit feature selectivity range different positions neurons seen initial stages process ultimately produces spatially invariant object representations see kinds invariant partially receptive fields develop weights need configured achieve next major area receives inputs visual area appears primarily focused visual form processing object recognition here neurons continue process spatial invariance coding exhibit complex feature detection temporal cortex neurons finally achieve high level size location invariance measure invariance further neurons encode complex difficult properties shapes seems clear neurons provide distributed basis invariant object recognition face distributed object type details specific properties neural code see pathway represents spatial information information relevant action neurons pathway large receptive fields preferred motion fire neurons pathway information position eyes properties support processing information location objects pathway model explore level representations visual cortex area provide basis upon subsequent visual cortical processing builds important contribution model understanding properties representations described previous section explain particular types representations computationally useful nature visual world main benefits computational models cognitive neuroscience explain brain cognition certain properties provides level understanding properties computationally oriented way thinking edges represented terms correlational structure visual environment discussed reliable correlations pixels edge object objects reliably tend edges sense represent internal model environment basic pixel correlations represented edges subsequent levels processing represent higher level correlations arise regularities arrangement edges different kinds edge basic shapes etc higher higher levels visual structure see next model objective model described section show network learn represent correlational structure edges present visual inputs received via simulated thalamus recently showed network presented natural visual scenes manner generally consistent enhancement contrast properties retina develop looking realistic set oriented detector edge representations network based known biological principles intended demonstration idea sparse representations provide useful basis encoding world real visual computational models early visual representations developed aspects detailed properties representations recent models useful potential relationships biological computational properties resulting representations model present incorporates properties identified models important principled based biologically mechanisms developed part text relatively simple model standard leabra model hidden layer produces fairly looking realistic representations based natural visual inputs model cpca hebbian learning algorithm same processed pre visual scenes showed sequential pca algorithm spca produce appropriate edge detector representations blob representation shown reason believe based results conditional pca cpca hebbian algorithm organize self appropriate conditional correlational structure present image edges hebbian model learning case effectively assuming levels perceptual processing sufficiently removed particular task driven error learning play major role put way assume statistical structure input visual images sufficiently strong constrain nature representations developed purely model learning system need extra based task constraints model producing looking realistic receptive fields assumption extent see next model next layer network benefits model task learning model important properties representations orientation position size emphasized varying existing models properties dimensions model develop coarse coding representations cover space possible values dimension case actual neurons case orientation units preferred orientation respond weaker responses increasingly different orientations further expect individual units particular tuning value dimensions coding large edge light degrees location finally explore topographic arrangement dimensions neighboring units representing similar values enable subsequent processing layers representations explore role excitatory lateral connectivity producing topographic inputs model based center layers lgn project itself modeled single layer actually corresponds hidden layer area cortical layers input layer neurons cortical layer appear basically same center receptive fields lgn retina inputs cortical input layer lgn model lgn model here nice separation easier receptive field properties network presented images natural scenes etc processed pre effects contrast enhancement retina done spatial effects surround center processing units network positive negative weights activations biologically implausible separate center components separate components activations weights biological constraints presenting valued positive processed image pixels center input layer absolute values valued negative ones center input layer note absolute value reflects fact center neurons excited input negative images center surround center directly method advantage exact same input statistics comparison results critical aspect model relative scale model intended simulate roughly cortical hypercolumn relatively small structural unit cortex hypercolumn generally thought contain full set feature dimensions orientations sizes etc area retina hypercolumn neurons settle considerably reduced model hypercolumn main factor determining scale model raw number units patterns connectivity inhibition determines ways units interact extent process same different input patterns model seen representing hypercolumn basis following connectivity patterns units connected same lgn inputs actual neurons hypercolumn different individual connectivity patterns roughly same part lgn lateral connectivity relatively large portion neighboring units units common inhibitory system kwta layer kwta based average inhibitory function ensures units active time critical specialization units discussed consequence scaling small overall image presented inputs network hypercolumn similarly processes small overall retinal input units fully connected input layers contrast models require initial spatially topographic connectivity patterns obtain specialization units function subsequent models simulating larger visual cortex model want include spatially restricted topographic connectivity patterns mechanism topographic representations excitatory interactions neighboring units implemented circle connections surrounding unit strength connections gaussian function distance unit active tend activate closest via lateral excitation learning cause neighboring units active develop similar representations similar responses subsets input images discussed key idea networks explicitly specific activation onto units surrounding single active unit layer main advantage current approach kwta inhibitory function allows multiple activation multiple different features present input time further closer actual biology balance lateral excitation inhibition achieve topographic representations lateral excitatory connectivity edges unit right side hidden layer actually unit left side same top bottom shaped functional onto hidden layer important otherwise units edges middle activated problem cortex network huge edge units constitute relatively small percentage finally receptive fields need represent graded coded coarse tuning functions appropriate default weight contrast useful binary kinds representations results shown weight gain parameter weight contrast default weight gain weight offset interact higher offset needed gain offset default order encourage units represent strongest correlations present input see details open project directory notice network input layers size representing small center lgn neurons representing similar center lgn neurons specific input patterns produced randomly set larger pixels images natural scenes single hidden layer size lets examine weights network clicking hidden unit observe unit fully randomly connected input layers neighborhood lateral excitatory connectivity needed topographic representations select again network window locate order load single processed pre image training network images time memory required exploration now press observe activations network settles response input pattern observe center input patterns complementary activity patterns activity activity versa vice reflects fact center cell excited image middle edges receptive field center cell excited image edges middle true active image location important keep mind center units active positive activations extent image contains relatively region location coded unit actually negative activations encode hidden units initially random sparse pattern activity response input images note noise added processing units important satisfaction constraint settling process balance effects lateral connections feedforward connections input patterns useful here same reasons useful cube example studied noise rapid settling relatively equally good states network case lateral connectivity unit same lateral weights point hidden unit space trying activity noise needed break enable best activity level noise determined parameter overall control panel continue input patterns effects lateral weights particularly evident hidden unit activity patterns expect see weights playing dominant role determining activities hidden units now increase control panel parameter default continue increase effective strength lateral recurrent weights hidden layer change hidden unit activation patterns set back relatively subtle strength level lateral weights want network able multiple activity dominant important multiple different edges different orientations present image case let network run image develop set representations reflect correlational structure edges present input take load trained pre network point network window network trained epochs image total image select click upper left hidden unit see weights onto input layers indication vertical orientation coding vertical bar stronger weight values center bar middle input center bar adjacent left note network center bars adjacent locations same location complementary active same place arrangement center bars direction orientation coding line fact encoding edges change orientation edge order see clearly weights small magnitude click top color scale bar right network window good visual contrast weight display overall range values represented providing better contrast smaller weight values click next unit right next click back forth units observe center bar remains roughly same position center weights unit switch left surrounding central center bar reflects receptive field left unit organization center center region next unit organization center region center ones unit left goes back field center bar right further vertical orientation going right represents different related orientation coding compared vertical previous units looks smaller size individual units weight values aspects dimensions coded units topographic organization difficult overall sense units representations looking time single display receptive fields time view display press control panel select response file appears able press now see grid log presents pattern receiving weights hidden unit center values center ones single plot receptive field hidden unit positive values red going maximum yellow indicating center center excitation vice versa negative values blue going maximum negative magnitude receptive fields hidden unit correspond hidden units network verify look same units examined upper left grid log see same features described keeping mind grid log represents difference center center values clearly see topographic nature receptive fields full range different receptive field properties dimensions receptive fields different hidden units vary observe topographic organization different features causes neighboring units share value least dimension similar least dimension keep mind units far right similar far left etc observe range different values represented dimension space possible values combinations values reasonably well covered interesting aspects representations network well neurons orientation receptive fields systematically degrees topographic space units seen starting hidden unit located top column number noting orientation units circle units wide surrounding units circle orientations seen consequence varying neighborhood relationships unit similar distinct orientation coding expect proceed circle goes further away unit middle represents relate neighboring unit values occur relatively changes orientation mapping short distance hidden unit topographic space phenomena found topographic representations real neurons provide important source data models account order fully replicate real system directly examine weights simulated neurons model possible biological system indirect measures taken order map receptive field properties neurons commonly measure activation neurons response simple visual stimuli vary critical dimensions oriented bars light replicate kind experiment model pressing environment containing events represents edge different orientation position explain sense probe stimuli represent edges reference relationship patterns control panel observe resulting hidden unit activations compare based weight receptive fields shown grid log described displayed here select network window press panel continue next events note relationship case based weight receptive fields shown grid log groups hidden units activated probe events correspond well based weight receptive fields explain interested new patterns probe events present same procedure described particular interesting see network responds multiple edges present single input event finally order see lateral connectivity responsible developing topographic representations load set receptive fields generated network trained topographic organization resulting receptive field grid log indicating strength lateral connectivity provided neighborhood constraints patterns look similar networks trained lateral connectivity receptive field plot suggests interaction topographic aspect representations nature individual receptive fields themselves look different case lateral interactions kinds interactions documented brain sense computationally lateral connectivity important effect response properties neurons responsible tuning receptive fields place via cpca hebbian learning rule model illustrates hebbian learning develop representations capture important statistical correlations exist individual elements edges exist perceptual environment resulting representations capture important properties actual receptive fields meaning model provide computational explanation properties arise brain representations provide building blocks elaborate representations illustrated next simulation likely similar principles apply learning sensory pathways next model visual form object recognition pathway range important issues model edge detector representations level builds way spatially invariant representations enable recognition objects regardless appear visual input space range different sizes characteristic processing pathway brain transformations performed collapse different spatial locations sizes distinctions different objects forms invariance location size simulated here likely same mechanisms lead least degree invariance well note clear extent brain exhibits invariance representations important challenges spatially invariant object recognition problem binding problem discussed problem arises recognizing object encode spatial relationship different features object particular edge right left hand side object same time collapsing overall spatial location object appears retina simply encoded feature completely separately spatially invariant fashion tried recognize objects basis resulting collection features spatial arrangement features relative objects same features different example clearly problem solved encoding limited combinations features way reflects spatial arrangement same time recognizing feature combinations range different spatial locations repeatedly performing type transformation levels processing ends spatially invariant representations encode spatial features model depends critically learning hierarchical series transformations produce increasingly complex terms object features spatially invariant representations general principle hierarchical representations likely important aspects cortical processing ability learn representations task model learning provides key demonstration general principle researchers suggested object recognition operates roughly hierarchical fashion existing models implement specific versions idea important difference present model model versions separate process increasingly complex representations increasingly invariant representations different stages processing training easier well properties visual system appears achieve increased complexity spatial invariance same stages processing contrast present model constrained way develops aspects representation simultaneously model demonstrates hierarchical sequence transformations work effectively novel inputs generalization see new object learned relatively rapidly small set retinal positions sizes recognized large percentage time further learning positions sizes accomplished performing spatial invariance transformation common structural features shared objects higher levels network contain spatially invariant representations object features need associated unique combinations define particular objects assumes possibly large set underlying structural regularities shared objects ensure case model likely true objects real world seen composed different component shapes etc particular exists component shapes exactly structural regularities learning automatically find model general approach taken towards object recognition problem here consistent features actual visual system results effective solution size invariance ways people thought solving problem example suggested dynamically object recognition system spatial transformation system effectively object back position size orientation representation easily recognized simple pattern recognition system similar ideas suggested different general approach well supported known properties visual system effective implementation kinds model ideas proposed computer solve object recognition problem kind gradual hierarchical parallel transformations brain well suited performing finally important limitation current model processes single object time see later section spatial based object representations interact multiple levels perception complex potentially visual displays containing multiple objects objects object recognition model environment experienced model contains objects composed combinations horizontal vertical lines regularity objects composed same basic features critical enabling network generalize novel objects described network trained objects range positions sizes last testing generalization training positions sizes testing recognition performance whereas previous model represented roughly single cortical hypercolumn model relatively wide cortical area hypercolumns previous case means model reduced terms number neurons again connectivity patterns determine effective scale model need represent large cortical scale requires additional structure far particularly respect way inhibition computed idea structure inhibition acts multiple scales stronger inhibition neurons relatively close single cortical hypercolumn significant inhibition communicated larger via longer range excitatory cortical connectivity projects onto local inhibitory interneurons represented model levels kwta inhibition units single hypercolumn amongst themselves kwta activity level same time units layer larger kwta activity level reflecting range longer inhibition level unit level inhibition maximum computed kwta computations aspect relatively large scale model connectivity patterns previous model assume units hypercolumn connected same inputs further assumption neighboring hypercolumns partially overlapping offset inputs due overlap coding input area multiple hypercolumns basically different hypercolumns process different information hypercolumn units processing different parts input basically same thing extracting features multiple locations sizes part input different objects appear speed learning significantly amount memory required implement network hypercolumn units share same set weights weight weight sharing hypercolumn layers model weight sharing allows hypercolumn benefit experiences hypercolumns learning time verify playing substantial role resulting performance network control network run weight sharing network took memory longer train expected resulting representations overall performance similar network weight sharing explore here sense hypercolumn experience same input patterns time develop roughly same kinds weight patterns hebbian learning undoubtedly important similarity weight shared separate weight networks tends reliably produce same weight patterns purely error driven backpropagation network run comparison perform similarly cases type network tends high level variance learned solutions hypercolumns different things led worse performance terms spatial invariance properties resulting network network seen extension previous input lgn separate center layers represented objects bars light pixel wide showed bars activity center lgn input center input activity ends bar representing light end information important widely thought neurons early visual cortex kind information represent lines particular length example included center activity length bar redundant center representation required kinds bars center right center left lgn units previous simulation right unit left same bottom top bar features pixels long depending size object sizes corresponding pixel length bars lower left hand side objects located grid total different unique locations combined different unique images object previous model area processes lgn input simple oriented detector edge representations already demonstrated kinds representations develop response natural images want number units extra units required organizing self learning develop properly discussed simply fixed representations encode horizontal vertical lines possible locations receptive field trying combinations different center inputs set units encode bars center field encode bars center field receptive field size lgn input horizontal vertical bars center center total units hypercolumn next layer cortical object recognition pathway contains neurons larger complex receptive field properties assume units represent combinations edges larger receptive fields enable slightly spatially invariant representations encode same feature multiple far input locations next layer continues towards increasingly complex spatially invariant representations due limited size model representations visual input space produce fully invariant representations entire space larger realistic model cortex next layer processing temporal cortex fully invariant representations possible units able cover input space relative simplicity objects simulated environment enables representations sufficient complexity terms feature combinations distinguish amongst different objects effectively layers layer last layer network output layer enables based task driven error learning addition based model hebbian learning train network assumed correspond number possible task outputs example different objects different sounds network corresponding representation modalities feedback improve ability visual system identify different objects accurately predict similarly objects different physical consequences themselves etc serve digits letters case simply distinct output unit object network trained produce correct output unit image object presented input based task learning important successful learning network parameters network standard amount hebbian learning set connections lower amount hebbian learning here necessary due weight sharing same weights updated times input causing hebbian learning dominant learning rate epochs minimize interference effects learning likely brain exhibits similar kind slowing learning note simulation requires minimum run open project directory see network looks skeleton big network units connections project file build network moment skeleton important aspects network structure see lgn input layers see see layer grid structure grid elements represents hypercolumn units hypercolumn contain group units network built units connected same small region lgn inputs discussed neighboring groups connected overlapping regions lgn see clearly network connected addition connectivity groups organize inhibition layer described kwta level set units hypercolumn entire layer hypercolumns activity activity need distributed manner layer organized grid hypercolumns time size hypercolumn units again inhibition operates hypercolumn entire layer scales here units active hypercolumn entire layer hypercolumn units receives layer neighboring columns again overlapping receptive fields next layer represents single hypercolumn units units single inhibitory group receives entire layer finally output layer units different objects seen organization layers skeleton view lets build network press overall control panel see network units connected now switch click units layer see hypercolumn units receives input layers neighboring hypercolumns receive overlapping clicking units see connectivity patterns similar form larger size now lets see network trained back viewing input image shapes shown random location size pressing different input patterns look takes network trained load weights trained network network trained epochs object inputs epoch object took roughly epochs object performance approach corresponds network seen object location size locations sizes object objects assuming perfect distribution case clear considerable amount generalization training due weight sharing weight sharing help size invariance example network relatively large actual weight values themselves file weights network window menu left hand side selecting now back control panel press trained network perform task see plus minus phase output states same meaning correctly recognizing objects presented record performance text log window lower left screen column shows error pattern see error column errors associated sizes low end resolution network feature detectors correspond objects real world particularly informative patterns activity layers network response different inputs further units directly connected input view weights easily see representing important technique activation based receptive field hidden units network activation based receptive field shows units activity correlated layers activity patterns measured large sample patterns example want know units activity lgn patterns activity unit reliably responds input pattern resulting activation based receptive field input pattern unit responds equally set input patterns result average patterns corresponding mathematical expression receptive field corresponding lgn unit example activation unit receptive field computing unit example activation unit layer computing receptive field lgn example index input patterns usual note similar cpca hebbian learning rule computes saw previous simulation lgn example activation based receptive field procedure compute weights receptive field values layers unit directly connected example useful look units activation based receptive field averaging output images see object unit representing now lets take look activation based receptive fields different layers network press overall control panel select press file selection window input receptive field units center lgn layer file selection window appear move examine current display weight sharing need look hypercolumns worth units displayed scale large grid window grid elements grid representing weighted average input patterns activation based receptive field unit note units lower left hand hypercolumn layer receive corresponding lower left hand region input receptive fields emphasize region notice units looking receptive fields appear represent orientations positions tuned explain level corresponds terms activation based receptive field computed taking advantage answer previous question describe characteristics receptive fields observe here terms selectivity particular input features particular kind evidence see conjunctive representations bind different features relevant different objects see evidence level spatial invariance single units respond features range different positions explain characteristics receptive fields overall computation performed network press file selection window bring next receptive field display again put next moment shows input receptive fields center lgn input layer observe clear receptive field patterns center inputs indicating neurons encoded linear represented center fields lines end represented center fields now press next file selection window bring next receptive field display shows output layer receptive fields same units enables see objects units participate representing notice appear correlation input output selectivity units highly selective input coding representation objects versa vice expected highly selective input tuning units highly selective objects represent fact representing shared features objects participate representation multiple objects images objects shown same configuration output units explain units particular output representation based features shown input receptive fields hint pick unit particularly selective specific input patterns specific output units things easier see press bring next window showing input receptive fields units sure notice scale shown bottom window tells large maximum values window describe receptive fields indicate selectivity units specific input patterns terms spatial invariance property units finally press view output receptive fields units again sure notice scale shown bottom window want manipulate scale new numbers pressing control buttons bottom right match scale grid log receptive fields entire objects parts different objects explain answer question now compare relative selectivity units particular output units objects compared units units selective explain hint determine number objects unit participate tell relative complexity input features units selective different layers note complex representation objects complex combination features probe stimuli test response properties units model perspective units encoding visual inputs probe stimulus technique observe units responses previous simulation present case record sum units responses display grid log based activation receptive fields examined press overall control panel select press grid log window appear file selection window appear again move examine current display log displays responses hypercolumn units different probe stimuli consists adjacent lines lines object space view responses display presented possible locations lower left set lgn locations plotted units responses probe grid shown display position probe corresponding shown responses locations lower left hand probe figure shown lower left cells overall grid upper right probe figure upper right cells overall grid unit probe locations highly spatially invariant representation show display solid yellow active color corresponding grid subset features compared significantly respond least activity conclude selectivity units number different locations unit active area pixels display say units least degree spatial invariance results correspond based activation based receptive fields examined previously press next file selection window bring probe receptive fields units range different feature levels spatial invariance indicated figure specifically units respond feature locations single yellow unit respond probe stimulus tend relatively spatially invariant probe stimuli seem units explain units sensitive terms features present probe stimuli see explain response properties unit appears time similar way testing representations network actual object stimuli systematically positions input different sizes record statistics resulting activity patterns important statistic number different unique patterns occur layers different positions object recall position lower hand left object different sizes average different patterns compared different object images conclude time representation fully invariant object interestingly true novel objects network seen critical generalization test described detailed report analysis found file simulation directory results object item size object size spatial errors correlations subsequent representations number unique activation patterns techniques obtained insight way network performs spatially invariant object recognition particular appears build invariance gradually multiple levels processing similarly complexity representations increases increasing levels hierarchy simultaneously stages multiple levels network able recognize objects environment depends critically detailed spatial arrangement features apparently binding problem described previously addition receptive field measures networks performance perform behavioral test ability generalize spatially invariant manner objects numbers presented network training now train objects restricted set spatial locations sizes assess networks ability respond items novel locations sizes indication work test showed network produced spatially invariant responses layer novel stimuli presumably network needs learn association representations appropriate output units good generalization result spatial locations addition presenting novel objects training present familiar objects otherwise network catastrophic interference see discussion issue following procedure trial chance novel object presented chance familiar presented novel object presented location chosen random grid center visual field possible locations total roughly locations size chosen random pixels possible sizes sizes familiar object presented size position chosen completely random procedure repeated epochs objects epoch learning rate network getting novel objects correct epochs longer training new knowledge well testing testing performed analysis training new objects detailed results contained file attention column shows errors function total item object overall results object roughly errors testing possible locations sizes object considering trained possible input images good generalization result note evidence interference training objects observed detail comparing file looking specifically size generalization performance object generalized novel sizes show sizes results file level errors different trained sizes similar results object errors novel sizes compared trained ones good evidence network able generalize learning set sizes recognizing objects different sizes seen determine learning primarily occurred examine difference pre generalization training weights load weight differences network viewed clicking click objects output layer see magnitude weight changes units now compare output units layers lower network clear primary learning occurred evidence weight change units accounts observed interference kind interference learning constantly weights important general result test demonstrates hierarchical series representations operate effectively novel stimuli long structural features common familiar objects present case novel objects built same line features objects network learned represent terms increasingly complex conjunctions increasingly spatially invariant role driven error hebbian learning missing properties size missing dimensions color depth etc need generalization training prevent interference items described attention plays important role solving binding problem restricting focus processing related set features object enables resulting pattern distributed representations sense apply features actually related opposed random combinations features different objects aspects environment problem previous model presenting simulation object time object multiple binding problem different feature multiple binding problem previous model multiple objects necessarily want form conjunctive representations encode multiple objects simultaneously objects completely possible objects commonly seen eyes face network same kind hierarchical conjunctions feature spatial invariance solution multiple object binding problem feature multiple section next develop closely related models attention end result model extension object recognition model network able multiple objects environment sequentially focusing attention turn focus case attentional effects object recognition specifically role spatial representations controlling attention principles apply generally processing pathways begin general issues regarding nature role attention proceed explore models need emphasize point attention separate mechanism framework emergent property activation dynamics representational structure network particular inhibition responsible limitation total amount activity set representations layer constraint satisfaction operating network representations active context emergent property convenient refer attention mechanism addition thought distinct mechanism attention associated types mediated spatially effects modeling here mechanisms underlie attention cortex same kinds issues discussed here context object recognition apply generally well see following models particularly special mechanisms representations model attentional spatial object processing pathway attentional contribution well attention obviously represents limitation processing providing important functional benefit enabling system take advantage powerful distributed representations confused limitations course possible brain large number redundant specialized processing pathways own system distributed representations things processed parallel limitations imposed attention issue actual benefit processing objects time ultimately resulting representations fashion ongoing processing perform specific task sense focus sequentially individual objects particular ongoing processing activation common object processing pathway directly interact representations aspects processing object pathways relevant particular case learn appropriate set weights processing pathway capable appropriate ways processing pathways words assuming objects likely relevant time well processing object time place possible imagine tasks objects simultaneously relevant generally seem case objects relevant solving task example large number objects pieces potentially relevant considered properly relationships effects pieces ones overall etc cases specific properties individual objects relevant useful consider objects same time otherwise important exception general point case objects discussed functional benefit attention fact need act coherent focused manner order effectively solve tasks important survival attention forces processing lead inevitably limited resources time energy etc end basic tradeoff parallel processing sequential processing different advantages case brain appropriate balance forms processing way benefits models follow see structural principles developed specialized processing pathways lateral interactions critical determining ways functional resolved specific case interactions spatial object representations specifically parallel processing takes place lower levels system higher levels greater focus single objects locations space representations space object interact via level low spatially feature simpler model begin explore ways spatial representations interact kinds based object processing developed previous model parietal lobe cortex contains different types spatial representations lesions lobe cause deficits spatial processing discussed assumptions nature functions parietal representations basic spatial properties fundamental assumption underlying beneficial role spatial representations object processing objects tend spatially useful group focus processing upon subsets features spatial region simple idea spatial object representations interact via top effects spatial feature map provides inputs subsequent levels processing important problem idea positive feedback characteristic spatial representations focus top attention particular spatial locations ability activity coming spatial locations attention new location system able switch attention new location top projections spatial system relatively weak weak resulting level spatial attention incapable sufficiently focusing object processing pathway object spatial representations attentional effect via direct connections object processing system spatially organized lower levels alternative model explore here shown includes lateral interconnectivity spatial pathway lower spatially organized levels object processing pathway areas object model described previous section emphasized figure influences spatial processing object processing stronger opposite direction bottom pathway stronger top efficacy spatial modulation object processing limited need keep spatial system sensitive bottom input model include multiple spatial scales processing object pathway case previous model spatial pathway important note model includes top projections object pathway lateral projections spatial system object pathway allow possibility based object attentional effects considerable debate field exact nature based object attention essentially attention spatial region object located specifically attention particular features object model allows potentially complex interactions pathways processing rise interesting effects explore issue model well developed object pathway posner spatial attention task cue attention region space reaction times detect target faster cue valid target appears same region invalid simplest influential studying spatial attention posner task attention region space box side display subsequently affects speed target detection object recognition attention same region target subsequently appears subjects faster detect target attention opposite region interaction spatial object processing captured general framework discussed shown specifically activation spatial processing pathway processing objects appear part space processing objects parts space time competing virtue inhibition activating corresponding feature representations object pathway observe dynamics model todo positive feedback model capture model effects lesions parietal cortex performance posner spatial task spatial processing effects object recognition generally argued lesions parietal cortex lead specific deficit ability disengage attention particular spatial location showed apparent disengage deficits accounted terms model spatial attention common explore following explore effects parietal lesions associated posner task note model form shown account specific parameter needed work model provide robust richer basic points virtue lateral interactions spatial object processing pathways discussed shown modeling exercise provides example based biologically principles understanding cognition rise different functional cognitive phenomena specific disengage mechanism model argued basis traditional box model thing model spatial attention effects parietal lesions generally consistent associated parietal damage typically right patients lesions show side visual space reference seen address model account terms damage side representations visual space difficult patients focus attention damaged side start opening project lets step network structure see network basically mutually interconnected pathways layer contains spatially feature simple case assuming object represented single distinct feature organized single spatial dimension row units represents objects feature serves cue stimulus different locations row represents objects feature serves target same locations object processing pathway sequence increasingly spatially invariant layers representations unit collapsing adjacent spatial locations layer spatial processing pathway similarly represents adjacent spatial locations object pathway sensitive particular features units location spatial pathway todo represent distributed representations likely present brain useful effects partial damage pathway select network window click object spatial units see function via connectivity patterns note output layer taken level object pathway reaction time detect objects settling whenever target output object gets activity happen settling stops cycles locate overall control panel contains number important parameters values determine relative strength pathways network pathways default strength relatively weakly activated visual inputs compared top projections object spatial pathways allowing dominated spatial object attention spatial system bottom inputs influences object pathway relatively strongly compared relatively zero non impact spatial pathway specialized pathways influence lower layers note apply projections spatial system back apply top projections object system taken spatial system dominant activation object pathway inputs parameters show relatively slow settling adding noise processing simulate subject performance now lets see network responds perception multiple objects presented simultaneously function spatial location objects provide introduction kinds interactions spatial object processing happen relatively simple model lets begin viewing events present network press overall control panel select event different objects features present different spatial locations note target object slightly higher activation result reliable selection object next event same objects presented different locations finally last event different objects same spatial location imagine events look actual visual display complex objects different letters events terms relative speed recognizing target object answer now lets test model locate followed present event network settling updating networks activations cycle time target units activation exceeds threshold remaining events graph log lower right shows settling times event todo settling time represents settling times match terms interactions excitation inhibition flows network run batch runs pressing panel report resulting average settling times text log immediately left graph log original results observed spatial representations facilitate processing objects attention object key contrast objects same location spatial attention longer separate leaving object pathway try process objects simultaneously see complex realistic example effect next model now lets see model posner spatial task overall control panel select task represented groups events shown here correspond cue cue case presentation target object left location valid case event cue presented left followed target event target left activations primary groups events opposed list separate events todo specify finally invalid case same cue event target event target showing right opposite side space again process control panel note network responds conditions posner task continue cases press control panel run batch runs subsequent processing target valid invalid cases typical reaction times human subjects task roughly valid invalid appears difference side effects attentional focus general pattern results obtained fit data precisely add constant offset roughly correspond factors task included simulation course relationship cycle processing human reaction time automatic simply particular experiment now lets explore effects parameters networks performance try reducing based comparison results default network effects spatial processing object processing set back reduce value turn network display back invalid target case explain results set back continuing manipulate parameter controls strength input determine role top modulation activations via spatial system fixed strength increase value effectively reducing influence top projections set find longer associated invalid trials reduced indicating attentional modulation important contribution overall behavior model todo important set back continuing additional manipulation visual distance cue target presented closer expect attentional effect tested control panel selecting case see overlapping set spatial representations activated run effect invalid case relative case strong valid trials switch environment back continuing mentioned earlier showed patients lesions parietal cortex exhibit performance invalid trials posner spatial task specifically cue presented side space processed intact lesion target processed lesioned patients showed difference valid invalid cases roughly compared control subjects showed roughly invalid valid difference valid model data need change mapping cycles settling model reaction time patients generally exhibit overall slowing reaction times due patients generalized effects damage addition specific effects recall cycles settling model normal subjects reaction times add normal scale cycles settling model control subjects valid divided normal valid models results example found cycles difference valid invalid trials scaled appropriately controls provides good match observed data apply scaling procedure patients data results fit well take valid patients divide same invalid valid difference factor significantly smaller invalid valid difference patients patients additional slowing ability disengage attention region space accounted overall reaction times attempt replicate additional slowing similarly lesioning spatial processing pathway network resulting reaction times posner task press control panel select lesion levels spatial representations location right space select units back units right weights resulting scale invalid valid difference overall slowing patients responses compare patients data intact models performance turn network display back control panel explain lesioned model invalid trials terms activation dynamics network found replicate apparent disengage deficit specific mechanism determine importance presenting target lesioned side space run posner task locations cue target clicking compare previous lesioned results intact network explain interesting additional lesion data comes patients parietal lesions called interestingly exhibit level attentional effects posner task smaller invalid valid difference emphasized data provides important argument disengage explanation parietal function posner predict problems invalid trials naturally type model weve exploring simulate lesioned network explain results differ lesioned network finally explore effects lesion parietal spatial representations provide better model known typically referred described results lesions parietal cortex right cause patients generally lesioned side space left side space due visual information cortex simulate similar lesion lesion units location specifically now back case object lesioned side space network attention intact side specifically lesioned side case causes network activate cue object representation network settling resulting full cycles settling now choose again see similar phenomenon network completely incapable switching attention damaged side space order detect target again resulting full cycles settling invalid case interestingly case attentional effects completely settling times same conditions roughly cycles case network incapable processing stimuli damaged side space competing stimulus good side space easily explain general tendency rare actually competing stimuli relatively weak competition coming intact side space attention focused intact side smaller level damage produces disengage deficit posner task closely associated phenomenon parietal lesions shows relatively strong competing visual stimulus presented good side space cue invalid trials posner task model able account wide range different spatial processing deficits associated parietal damage depending location damage object based spatial modulation side object etc emphasize role inhibition simple model explored provides useful means understanding basic principles spatial attention interacts object processing address main spatial attention place avoid binding problem multiple objects next simulation basically previous models full object recognition model simple spatial attention model result model spatial attention restrict object processing pathway object time enabling successfully perform environment containing multiple objects essential ideas covered discussion previous models proceed directly exploration model note model currently inconsistent previous object recognition model updated model principles discussed apply new model open project notice network comes essentially same object recognition model addition spatial processing layers layers interconnected respectively amongst themselves providing interactions rise spatially mediated attentional effects network window menu left hand side select weights network trained objects presented main reasons done due involved providing feedback training signals based attentional selection imposed network itself sure feedback name object appropriate object network actually selected done spatial representations training signal equivalent plausible ways implementing kind training results likely similar produced network trained single objects weights select network window examine connectivity patterns spatial layers similar spatial extent corresponding object pathway representations unit unique location spatial system information location needs represented todo specify contrast now simply test network patterns composed objects random locations note objects smaller size previous object model necessary fit multiple objects display locate process control panel followed see input pattern presented network activations updated cycles settling notice updates representations objects activated features objects remain active result activation coming back spatial pathways supporting object specific object gets selected depends randomly level activation corresponding features object pathway particular case network object lower left hand object number note spatial locations represented upper left hand regions represent inputs lower left hand areas active case activation features object recognized object processing pathway robust amount interference appears network performs better allowing bit activation spatial constraint stronger weights spatial system object system allowing spatial units active note evidence based object attention activity pattern well case entire set features location activated result focusing spatial attention location todo case difficult see network capable determining object present location seems clear successful network performance requires spatial attention essentially role favors activations directly activate features itself grid log right network shows output pattern produced end settling left grid allows compare target output pattern right grid actual objects present input pattern network said performed correctly unit activated present target standard sum squared error grid log appropriate error measure case column computes error unit target pattern otherwise provides better measure actual network performance continue patterns environment notice network typically hard time patterns directly overlapping occurs frequency gets right hard correctly sense due limitations ability network obtain perfect recognition training network errors single patterns due fact spatial system present training continue patterns text log right report total count statistics roughly events todo greater total relatively large percentage errors due overlapping stimuli network actually performing reasonably well order determine spatial processing pathway contributing performance set parameter reduce impact spatial input meaning network incapable focusing individual objects again see object pathway gets activation objects produces output correspond objects presented presumably network confused features different objects unable bind single object set back continuing spatial attention useful enabling single object processing pathway focus single object object environment useful attention sequentially different objects present scene implement simple version kind attention switching taking advantage accommodation properties neurons described object processed network currently active neurons due accommodation new set neurons activated presumably new neurons represent object present environment press button control panel click button turn accommodation channels let potassium ions unit active period time switch new environment events overlapping non objects extend settling time activation dynamics time play now press process control panel see same event saw previously notice correctly recognized object lower left focus activation object lower right correctly recognized point network goes correctly recognize object performance subsequent events far perfect clear network least partially successful switching attention different objects likely better performance kind task require explicit training control system provides top activation spatial representations order direct attention towards different objects kind accommodation mechanism allowing attention further provides useful demonstration kinds looking dynamics emerge fairly simple mechanisms see far model taken order account phenomena visual search see simple model phenomena key properties common weve exploring present treatment different attentional object recognition phenomena same computational principles models described chapter memory defined effect experience general forms take neural network activity based activation memory changes weights based weight memory relationship learning memory weight changes clearly equivalent learning network ability maintain activation states learning explore ideas relatively generic leabra network provides reasonable model general memory cortex slow learning information integrating different experiences learning thought taking place cortex activation similarly cortex well memory general highly distributed take different forms different representations cortex find useful general distinguish cortical memory systems see basic model suffers important limitations level interference subsequent learning ability maintain activity absence input limitations result basic tradeoffs discussed generic cortical model specialized systems specialized system rapid learning arbitrary information hippocampus related structures hcmp specialized system robust rapidly maintenance activations prefrontal cortex pfc further useful distinguish set effects experience generic cortical system relatively short term activation results form short term priming items processed rapidly immediately similar items addition relatively small weight changes resulting processing item rise long term priming effects last long time period due based weight nature term longer effects learning experiences leads memories characterized semantic memory memory mediated types memory mediated hcmp memories characterized nature define nature hcmp memory according underlying mechanisms enable hcmp system learn rapidly interference specifically sparse conjunctive representations finally mediated pfc active memories constitute memory characterized representations internal context models previous chapter provide examples slow learning gradually new knowledge cortical processing pathways resulting semantic memory object recognition model learning features different objects world addition gradual process effects take time see following exploration slow learning immediately effects example small weight changes properties existing representations ways lead weight based long term priming try same system rapidly learn novel information find rapid learning distributed representations causes interference see subsequent exploration need separate specialized system achieving rapid learning overlap different representations model system hippocampus presented main behavioral studying priming completion subjects study list words asked simply come words complete initial word words constructed possible increased probability coming studied word relative control subjects words initially taken indication residual effects processing word words initial studying word subsequent processing word importantly turns form priming intact normal levels patients hcmp lesions indicating effects single presentation word mediated intact cortical system results surface appear challenge idea cortex learns slowly experiences see easily accounted framework way completion task modeling purposes mapping input different possible words related priming explicit different spelling task subjects name asked critical word word read note words pronounced same thing actually said input ambiguous pronunciation output possible model trained associate different output patterns input pattern simplicity random distributed patterns input output patterns initial period slow training allows network appropriate associations simulate subjects prior experience results relevant knowledge trial network produces outputs response input pattern training followed testing phase network presented particular association input tested see word produce input see single trial learning same learning rate network information initially results strong bias towards producing output now lets explore model open directory notice network standard layer structure input presented bottom output produced top press events shown event bottom pattern represents input top represents output able tell set events set events same set input patterns different output patterns events reflect event random input pattern number corresponding output pattern labeled event same input pattern corresponding output pattern labeled total different input patterns total output input combinations events iconify environment window train network standard combination hebbian driven error learning turn network turn complete training graph log shows statistics training reber grammar network possible outputs input standard error measure target closest event statistic find event training environment closest similar target output pattern output pattern network actually produced minus phase statistic results distance closest event thresholded usual output exactly matches events environment name closest event appear graph log value appear testing log later event currently presented network otherwise think binary distance error measure computed sum closest event epoch training plotted yellow graph log network starts producing outputs exactly match valid outputs environment necessarily appropriate outputs input pattern approach zero statistic graph log shows plotted blue input name looks part event name input pattern portion name possible outputs approach zero network noted ability learn mapping task depends critically presence kwta inhibition network standard backpropagation networks learn produce blend output patterns learning produce output inhibition helps network choose output active same time inhibitory constraints mapping adding small amount noise membrane potentials units processing provides selection output produce finally seems hebbian learning important here network learns task better hebbian learning purely error driven manner hebbian learning help produce distinctive representations output cases virtue different correlations exist cases trained network appropriate semantic knowledge now assess performance priming task press control panel bring new process control panel running testing text display results perform different kinds tests assess biases exist respond input patterns outputs baseline responses done learning turned turn learning train network produce adjust weights done training see trial learning produce particular output substantial effect probability producing output lets baseline values learning determines learning point press control panel epoch events see column larger text log events presented sequential order outputs presented outputs testing now outputs actually presented plus phase run minus phase means list basically same same input patterns order event presentation sense turn training differences point due noise added unit membrane potentials columns pay attention larger text log event name actual output produced network described generally producing correct outputs critical issue response see fairly random set output responses determine times network produced output pattern corresponding particular target output current event look column refer simply time same name event produced same output pattern otherwise observe roughly time produces same name current event smaller log shows result column column monitor actual network producing correct outputs note statistic perfect errors magnitude explain target output actual output roughly now lets turn learning see obvious effects trial learning input pattern subsequent performance same input set event press again process control panel now learning event expect result seen output associated particular input pattern list events network likely produce output comes input pattern again baseline case observe systematic difference responses network patterns recall network trained produce output output log actually produced minus phase reflect effects prior training training produce producing output events time same inputs patterns affect statistic summary value now again report happens time events presented explain behavior relate priming results humans described seen simple model cortex slow learning develop useful distributed representations show effects single trials stimuli now see cortical model task rapidly learning arbitrary associations further explore particularly challenging case high level overlap associations learned case valid required remember information similar things ones opposed previous task studied human subjects commonly known list learning task represents set words associated different sets words example word window associated word reason list window associated list studying list subjects tested appropriate associate words subjects study list multiple subsequently tested lists recall learning list subjects exhibit level interference initially learned associations result learning list remember reasonable percentage see data comparison human backpropagation network performance list learning task tried standard backpropagation network perform list learning task found network described catastrophic interference comparison typical human data networks performance shown human performance goes correct recall list immediately studying roughly learning list network immediately recall well list learned model explore here start catastrophic interference effect standard leabra network neural networks good models human cognition see significantly reduce interference important changes networks parameters original catastrophic interference finding amount subsequent research consistent basic idea interference results same units weights learn different associations weight changes learn association necessary learn obvious way avoid interference different units represent different associations previously solution distributed representations same units participate representation different items see set parameters result distributed representations case catastrophic interference set parameters result units active separated representations avoid interference benefits distributed representations explore model hippocampus related structures appears providing specialized memory system based sparse separated representations rapid learning interference basic framework implementing task input patterns represents stimulus represents list context assume subject develops internal representation different lists serves means produced input patterns hidden layer produces output pattern corresponding associate explore different representations word items distributed representation random bit patterns localist representation overlap different items begin open project directory lets look training environment press distributed patterns windows come show list lower left list upper right bottom patterns item list context upper pattern associate note item list same pattern item list true corresponding items lists items list share same context pattern items ahead iconify windows now lets see well network performs task locate see graph log updated network trained initially list list red line shows error number items produced units right side training set yellow line shows error testing list note red yellow lines start roughly correlated identical due fact testing occurs training weights different item presented epoch red line training error gets zero epochs pass getting zero network automatically training list see red line immediately new set training events results see yellow line immediately well indicating learning list prior learning list run batch subjects summary average statistics taken end list training subject appear text log lower left shows training error shows testing error list batch run simulated subjects results compare human data presented error list introduction list turn network hidden unit representations unit seems point active input pattern seems obviously problematic interference perspective list going activating same units reducing extent hidden unit representations distributed able encourage network separate representations learning lists items lets test idea reducing parameter units active time hidden layer result distributed representations activity report resulting average testing statistic describe effects manipulation number epochs takes network reach maximum error list introduction list results compare human data presented reason network performed well expected done encourage different sets units represent different way encourage increase variance initial random weights unit pattern responses encourage different units encode different change thing improve performance enhance contribution list context inputs relative stimulus list context different changing increases weight scaling context inputs imagine processing subject finally increased amounts hebbian learning contribute better performance perfect correlation items list associated list context representation emphasized hebbian learning lead different subsets hidden units representing items lists different context representations setting batch run new parameters testing shown basically best performance obtained network now good model human performance important dimension weve emphasized speed network learns clearly learning fast human subjects further appears manipulations weve improve interference performance longer training times observed cases maximum epochs reached learning items need play parameter see speed learning network keeping same optimal parameters set training run note training list stops epochs red line zero observed learning significantly faster list learned actually learning faster network learning item previous learning items evidence list interference greatly reduced order substantially increase rate learning words network appears want learning same kind slow integrative way try representations sparse separated slow learning rate necessary kind integrative learning meaning addition interference network fail learn fast human subjects weve seen improve performance moving away distributed overlapping representations towards separated representations seems simple architecture somehow incapable matching human performance task well see later hippocampal system specialized neural architecture properties appear particularly well suited role rapid arbitrary learning system particularly biologically plausible thing try network localist representations items context distributed ones pressing button control panel select active item context note switch localist representations switch simple kwta inhibition function output layer done localist layers due average unit based average kwta default distributed representations hidden layer continues based average kwta localist representations batch run training runs report average testing error better model human performance human explain localist patterns task easy network well press button control panel batch run report results understand function hippocampus related structures hcmp terms tradeoff standard cortical model learning new things rapidly requires different overlapping non sets units minimize interference need distributed overlapping representations represent underlying structure environment expect different specialized systems support different types learning memory considerable amount data indicates hcmp rapid learning system cortex slow learning form weve explored previous models section see model incorporates important biological properties hcmp perform rapid learning significantly interference distributed cortical network model based bidirectional connectivity hcmp specifically cortex wide range cortical areas adapted described hcmp top cortical hierarchy receiving wide range different types information cortical areas see hcmp receives inputs represent essentially entire cortical state time cortical representation current state environment order serve role memory system encode input pattern fraction original pattern original whole critically retrieval occurs initially hcmp spread back cortex resulting original cortical representation words hcmp cortical representations somehow particular memory detailed content memory cortex hcmp simply subsets cortical representations hcmp semantic information represented interconnectivity overlapping distributed representations cortex bind pieces semantic information isolated facts chance hcmp memories characterized memory associated particular example encode report details hcmp representation typically represents particular whereas cortex represents eventually associations cortical representations initially encoded hcmp learned cortex itself standard slow learning process detailed treatment set ideas issues see critical functional properties hcmp system summarized competing mechanisms pattern separation pattern completion pattern separation leads different relatively overlapping non representations hcmp different subsets units encode different memories interference different sets weights involved pattern completion enables partial trigger activation complete previously encoded memory pattern separation operates encoding new memories pattern completion operates retrieval existing memories pattern separation hcmp indicate representations composed patterns active units cortex overlapping relatively large proportion active units here hcmp sparse representations smaller lead overlap separation units hcmp conjunctive activated specific combinations activity cortex shows representations units active likely different subsets units represent different patterns pattern separation principle tried cortical model list learning relatively small effects right direction see hcmp sparseness scale larger effects way thinking sparseness produces pattern separation terms higher threshold units active come greater levels inhibition producing large inhibitory excitation overcome threshold input patterns overlap same unit receive activation patterns high threshold smaller words high threshold increased levels specialization representations particular input patterns think highly specialized due high threshold competition same basic principle evolution specialization competition effect described context hippocampus detailed mathematical treatment found effect sparseness individual units conjunctive meaning activation depends conjunction multiple input features units understood terms higher threshold activation unit relatively large number active inputs conjunction inputs threshold implicit notion individual units receive subset input units units active good subset receive input subset active subset connectivity increasing variance random initial weights list learning model saw bit free interference learning ideas sparse conjunctive representations pattern separation critical understanding functional role biological properties hcmp important understand need pattern completion pattern separation work previously stored information exactly same original input activation pattern hcmp new pattern separated version input recognizing retrieval cue existing memory order actually memories stored hippocampus mechanism pattern completion needed pattern completion mechanism takes partial input pattern subset stored memory missing parts asked chance input cue sufficient trigger completion full encoded memory enabling respond pattern completion particular properties hcmp system strong set lateral connections particular layer hcmp perform same kind pattern completion explored fundamental pattern separation pattern completion consider following event good starts story happened story hippocampus know information new memory keep separate pattern separation memories complete information existing memory story case hippocampus produce completely new activity pattern produce completely perfect memory presented exactly same way time problem obvious solution memories noisy inputs require call basically tradeoff operating hcmp itself pattern separation completion tradeoff actually understand features hippocampal biology hcmp model areas connectivity corresponding columns input example activity pattern note sparse activity intermediate sparseness shows diagram hcmp model contains basic anatomical regions hippocampal including fields including well cortex serves primary cortical output input pathway hippocampus output input area represented here likely play similar role greater emphasis subcortical motor representations describe layers pathways model summarize known biological properties contribute pattern separation completion functions model based model model provides framework functional properties memory mechanisms pattern separation learning synaptic modification pattern completion further mechanisms underlying anatomical properties hippocampal model basic computational structures hippocampus feedforward pathway area via important pattern separation pattern completion recurrent connectivity primarily important pattern completion model sparse random projections feedforward pathway strong inhibitory interactions form sparse random conjunctive representations emphasize importance region providing means separated representation back language necessary recall information happen forms representation pattern pattern rise place rough estimates size hippocampal areas expected activity levels corresponding values model data layer hcmp same basic excitatory inhibitory structure described cortex importantly layer inhibitory neurons form local feedback serve activity levels system simulate inhibition kwta inhibition function same kinds learning mechanisms work cortex hippocampus hippocampus research synaptic modification done cpca hebbian learning model task encoding information model driven error task learning operating hcmp patterns firing neurons cortical neurons highly distributed activity patterns contrast neurons highly sparse conjunctive firing specific locations rough sizes activity levels hippocampal layers shown note seems sparse level activity roughly times larger layers active output input layer model roughly scaled numbers units activations generally higher order obtain sufficient absolute numbers active units reasonable distributed representations additional evidence regarding sparse importantly conjunctive nature extent neurons example shows patterns neural firing neurons fire particular location particular direction explained neurons activated particular conjunctions sensory features present specific locations amount known detailed patterns connectivity hippocampal areas starting input structure topographic projections different cortical areas projections broad projection known pathway sparse focused topographic neuron receives synapses projection synapse widely significantly stronger inputs contrast cortex projections feedforward direct feedback regions known exist lateral recurrent projections project widely neuron receive large number inputs entire similarly connecting wide range finally interconnectivity relatively point point projections now put biological properties work model explaining encoding retrieval memories works terms areas projections general scheme encoding activation comes cortex flows pattern separated representation sparse distributed set units bound rapid hebbian learning recurrent plus learning feedforward pathway helps encode representation simultaneously activation flows pattern separated representation association representations encoded learning connections encoded information way retrieval partial input cue occur follows again representation partial cue based inputs cortex goes now prior learning feedforward pathway recurrent connections leads ability complete partial input cue original representation representation activates corresponding representation capable complete original representation input pattern novel weights particular activity pattern strongly driven activity pattern corresponds components previously studied conjunctive nature representations prevent successful recall addition performing pattern completion complete memory partial cue output hcmp presumably recognize items events etc new ones known recognition memory divided mechanism specific information item mechanism provides new items hcmp appears play central role component recognition model successful recall input probe pattern back hippocampus section further details model generally based biology shaped reasonably simple working model details details input representation mapping problem arises recall input representation incorporates topographic characteristics different cortical areas areas sub represented different slots thought representing different feature dimensions input color semantic features etc slots units slot unit slot active unit representing particular feature value input patterns constructed randomly selecting different feature values random subset slots distinct layers receives input cortical areas projects hippocampus receives projections projects back cortex deep representations layers different details assume equivalent same representations serve separated pattern representation back activation patterns pattern completion representations mentioned same time achieve amount pattern separation minimize interference learning mappings pattern separation explain hippocampus actually directly back input challenge implementing requires systematic mapping pattern separation requires highly linear non mapping done model training mapping pieces referred columns separated pattern representations entire representation composed systematically additional learning different combinations representations different columns column conjunctive pattern separated columns units entire composed columns column receives input adjacent slots units consistent relatively point point connectivity areas weights column trained taking column activity level training combination patterns slots conjunctive pattern separated representation patterns slots scheme units required column nonetheless consistent relatively greater humans relative hippocampal areas function cortical size further benefit certain combinations active units column correspond valid patterns allowing invalid combinations due interference imagine real system slow learning develops mappings columns separately time finally basic problem recall system needs able distinguish activation due item input directly via due activation coming recall solution problem suggested lesions phase respect drives point units otherwise providing means driven activation approximate mechanism simply turning inputs testing assess quality hippocampal recall comparing resulting note model requires least run exploration hcmp model same basic list learning standard cortical network hcmp able learn new causing levels interference original associations able rapidly possible cortical model least distributed representations here start open project hippocampal due ability model simulate wide range hippocampal phenomena directory shown large models stored skeleton needs built button overall control panel observe process activation network training locate control panel see particular input pattern training set presented network input pattern composed parts representing associate items representing list context source information list context slightly different item considerable overlap amongst context representations same list lists training parts presented testing associate pattern completion hcmp based partial cue stimulus context see activation flows layer pathway simultaneously sparse representation associated representation back end settling notice active units network selected remain selected settling next pattern allowing easily compare representations subsequent events observe relative amount pattern overlap subsequent events layers report general terms amount overlap layers explanation pattern separation mechanism relative levels activity different layers explain results epoch training consists training events followed testing sets testing events testing set contains list items contains list items contains set novel items sure network treating novel items network automatically testing pass training events smaller text log lower left display training events processed lets speed things bit turning display training events network select overall control panel train process monitor text log event presented number point network display again start updating test event network trained epochs detailed testing observe testing process point notice input pattern presented network missing middle portion corresponds associate activation network capable missing part pattern completion result activation back via processing stops epochs training testing now testing process detail control panel immediately right weve propagation network cycle cycle basis lets clear larger text log right side display contains record testing events multiple testing see testing event layer corresponds studied stimulus stimulus list context representation list studied likely network able complete pattern able see activation pattern gets network settling pattern compare activity pattern produced original stored activity pattern pressing button network window see target pattern comparison purposes layer active units selected relatively easily compare compare switching back forth act now look line testing text log critical columns here shows proportion units activated pattern shows proportion units activated active present pattern measures zero network correctly original pattern large indicates network otherwise different pattern relatively rare model large indicates network recall probe pattern common novel items last testing set see now look graph log right testing text log shows plotted axis axis event showing particular location extent lower left hand network accurately log automatically new testing set sure particular testing environment order code discrete responses network need set statistics following chosen provide reasonable performance say hcmp item shown column text log successful ignore columns point relevant simulation continue patterns sure understand relationship networks performance statistics turn back recall event continue step studied items tested again see field increments now presenting same stimulus list context items trained expect network successfully recall items learned next items network notice towards smaller values caused similarity studied items overlap next set testing events completely novel set items network high aspect items report total number responses testing text log testing respectively now trained tested resulting performance train network epochs examine testing log network automatically tested training epoch looking interference back control panel value field select process sure press effects prior training monitor training testing text network trained tested testing text log testing respectively epoch training find evidence interference learning testing results compare contrast performance hcmp model cortical model human data same basic task particular attention interference number epochs necessary learn found explorations biological properties hcmp result system particularly learning new information rapidly interference previously information relatively similar issues scaling capacity switching train test phenomena space conditioning terms based activation memory simple attentional model illustrates activation left processing cue stimulus affects subsequent processing target explore residual activation result activation based short term priming effect similar based weight lasting long residual activations useful form active memory longer periods time requires active maintenance mechanism keep activation going time see subsequent exploration recurrent excitatory connectivity provide mechanism turns distributed representations cause similar problem active memory system based weight time activation overlapping representations network required hold activation state support environment need separate specialized system achieving robust maintenance active memories model system pfc presented further elaborated next explore based activation properties standard cortical network saw model posner spatial task activation states affect subsequent processing explore similar effects priming simulation described behavioral data effect essentially same based weight priming case weight activation based changes result similar types primary difference based weight priming relatively lasting long based activation priming relatively model simply turning learning simulation priming effects due recent activation easily perform manipulation humans resort techniques way types priming prior immediately preceding experience builds weights favor response experience favors explore idea greater detail context phenomenon begin open project simulation weight priming assume already familiar network environment begin trained pre network network menu select configuration default lets test network different last time obtain baseline measure performance press test responses output presented allows determine impact seeing case response case presented output response pattern value case presented input pattern presented output response serves comparison pattern actual output produced network input observe network produces response trials time indication priming completely activations event presented roughly response reflects random biases trained network parameter controls extent activations reset event processed lets change activations completely intact trial next now again notice increased tendency network respond trials test trials explain effect performance now explore ability standard cortical network maintain active memories longer periods time absence supporting input presence inputs kind active maintenance useful keeping information immediately active form longer periods time residual activation explored previous section obvious neural network mechanism achieving active maintenance recurrent bidirectional excitatory connectivity activation constantly active units maintaining activation implement bidirectional excitatory connectivity context distributed representations find original activating input pattern removed maintenance pattern activation distributed representations resulting original information distributed representations again appear important form memory standard model cortex necessary distributed representations perform active maintenance same time tradeoff resolved prefrontal cortex pfc specialized robust self active memory discussed further related items terminal represented overlapping distributed fashion common features explore ideas simple overlapping distributed representations terminal terms features monitor shown open project see network hidden units representing features input units provide individual input corresponding hidden unit view connections hidden units enable maintain representations input pattern turned nature distributed representations unit needs able support unit different items involve different combinations features weights feature units press button grid log shows activity input hidden units event input present event input removed see features active input activates appropriate hidden units corresponding distributed representation terminal input subsequently removed activation remain features feature impossible determine item originally present spread occurs simply units interconnected problem weights exactly same connections likely true brain manipulate parameters control panel introduce different amounts variability recurrent weights see spread problem explain results obvious solution spread problem eliminate interconnectivity units isolated allow maintaining self add excitatory connection self unit press button control panel switch network type connectivity verify connectivity set back respectively press button network observe network now able maintain information difficulty now ability perform useful computations require knowledge features unit separated now lets active maintenance task challenging explore role recurrent connections noise change leak current reduce controls relative strength recurrent weights now network weaker recurrent weights case see longer capable information long todo elaborate enables network hold onto information time add noise membrane potentials units setting again multiple times network ever maintain information whole time find value increments time allows network maintain information repeated runs explain noise requires stronger recurrent weights section explore development interaction activation based weight memory previous sections focused forms memory separately here consider cases changes weights activity interact same system support competing responses competition arise weights response pathway network result repeated practice activity recent processing favors competing pathway addition exploring interactions model provides introduction important issues development neural network models provide important understanding experience genetic factors interact producing patterns changes take place develops studied extensively human infants provides good example competition activation based weight memory task subjects toy location typically allowed search object short delay procedure repeated subjects object new location following short delay human infants search error based weight memory support reaching activity dependent development prefrontal cortex pfc direct correct reaching consistent lesions prefrontal cortex performance task important note researchers tests infants knowledge require performance factors reaching demonstrate underlying object infants underlying fail tasks due performance limitations assumption researchers experiments demonstrate earlier example reduced production errors observed expectation looking reaching behaviors differ task looking response accurate infants errors task observe finally expectation task infants look longer toy hidden nonetheless search sensitivity objects new location commonly evidence infants know hidden task search due deficits external knowledge representations see understand findings terms competition activation based weight memory neural network framework knowledge deficits model explore ways activation based weight memory support competing responses task based weight memory implemented standard cpca hebbian learning occurs function activity units network activation based memory implemented recurrent connections representations network discussed style pfc active memory system previous section note original model exercises presented here simplification recurrent weights increased hand simulate effects development potential role experience shaping weights demonstrated input representations model based idea spatial location object identity processed separate pathways discussed pathway represents location hidden represents cover hidden hidden layer represents locations receives input object representations hidden layer recurrent self connections representation location represents pfc active memory system model output layers reaching expectation single difference frequency responses updating unit activity task expectation layer responds input reaching layer responds inputs corresponding stimulus reaching distance updating constraint output layers capture different frequencies infants reaching trial infants reach contrast prevents infants underlie longer looking impossible events trial similarly infants restricted experiments infants nonetheless reach show simulations responses change dynamic active resulting looking reaching networks initial connectivity includes bias respond appropriately location information look location presented infants appear enter experiments biases exploring model open project notice location units corresponding locations network cover input units corresponding default cover type different cover type toy units corresponding default toy different toy type now lets click observe connectivity input layers fully connected hidden layer hidden layer fully connected output layers see initial bias same locations strongly activating weights locations initial connection weight connections toy cover units relatively weak hidden output layers recurrent self excitatory connections back unit initially magnitude change improve networks ability maintain active representations starting relatively weak ones simulate active maintenance now examine events presented network click overall control panel types trials represented events see window trials pre corresponding practice trials provided start experiment infants reach trials trials trial types repeated multiple times events trial version running task consists trials pre trials trial trial consists corresponding trial follows place infants attention particular location infants attention location place presents place infants response reaching patterns activity presented input units corresponding visible aspects stimulus event input units activity levels input activity represent aspects stimulus aspects toy producing activity now lets run network easier tell going network looking grid log display viewing trial activation separately press button control panel training process network run entire experiment record activations weights grid log updated dynamically processing button turned simply specific points experiment buttons control panel press see trials pre main columns now column tells event presented next column shows activations network event columns show weights difficult interpret point well notice toy presented activated network looks location expectation layer hebbian learning taking place trial units active experience weight increases case increases network activate location representations press see testing trials networks tendency reach location note result hebbian learning hidden output units active here trials pre networks internal representations output reach responses delay choice trials observe network error explain network performing terms interactions weights learned prior experience trials recurrent activity representation now increase parameter default again describe network responds time based experience weights basically same cases explain network performs now decrease parameter intermediate value run happens trial time explain network exhibits different responses relate behavioral data showing expectation reaching measures task infants typically perform better errors delay reach simulate pressing button default delay press try explain effects delay networks behavior set back reduce think effect manipulation press case representation recurrent weights explain going here causes increase nature knowledge task interactions different forces importance mechanistic models wide range memory phenomena potentially explained terms interactions specialized brain areas described posterior motor cortex pmc prefrontal cortex pfc hippocampus related structures hcmp sketch areas contribute phenomena section explore explicit models due magnitude complexity models required hcmp model account contribution hippocampus range recognition memory phenomena hcmp enable subject experience seen item recognition addition hcmp recognition pmc form priming signal idea here sensitive effects small weight changes resulting item enabling people recognize item familiar computationally biologically account consistent process models recognition memory subjects signal fails list length list strength effects etc effects pfc hcmp interactions subjective organization sequential recall etc frequency level encoding etc pfc hcmp contributions capacity experience etc language particularly interesting domain study cognition plays important role cognitive phenomena depends wide range specialized processing pathways question regarding relationship language thought require neural network perspective provides useful intermediate based language input output representations associated wide range representations different parts brain enabling mutual interaction shaping representations verbal input activate perceptual verbal non motor representations via associations versa vice similarly time learning shaped influences language perception perception language short language viewed set specialized processing pathways perceptual motor association areas participate overall pathways produce human cognition issues dominated study language phenomena neural network perspective nature relationship rule processing processing exceptions regularities language contains aspects rules rules absolute seem exceptions metaphor computer perspective implement system set rules table exceptions dependent knowledge embedded nature neural network processing require separation types processing neural network models allow parsimonious accounts complex regularities exceptions language modeling unified system exhibits sensitivity regularities exceptions function frequency perceptual level language involves form object recognition takes visual sensory information representations primary building blocks language letters phonemes words perceptual pathways appear located appropriate general pathways different modalities ventral object recognition pathway recognition word processing viewed specialized versions apply ideas developed modeling visual word recognition provides representations words orthography explore basic features speech develop phonological representations capture features output end language able internal representations correct primary output pathways speech writing explore mapping perceptual motor output context simple models reading visual word input speech output domain regularities exceptions sound spelling orthography phonology mapping extensively studied modeled important challenge models ability account ability pronounce novel nonwords systematic ways generalization tests complex interactions different arise regularities mapping applicable apply locally different mapping system internal semantic representations word meaning speech output study words explicitly language feature specific types information specific case tense past inflectional system played large role application neural networks language phenomena period regular tense past inflection rule add producing things went taken indication based rule system applies rule detailed pattern data modeled neural network framework see correlational sensitivity hebbian learning combined driven error learning important capturing behavioral phenomena purpose language convey meaning naturally issue semantics assume semantics involves associations language representations rest cortex complex issue way language input shape semantic representations occurrence relationships different words idea words occur likely semantically related way shown hebbian based pca mechanism develop useful semantic representations word occurrence large text representations appear capture sense common relationships words explore model idea cpca hebbian learning developed taken different orthographic phonological semantic representational systems provide distributed representation words traditional models language discrete possible isnt pure area brain representations words simply distributed number different pathways specialized different aspects words idea distributed model language neural network perspective begin chapter model idea orthographic phonological semantic representations words interact activation areas produce appropriate corresponding activation areas relatively scale small model provides nice overall framework understanding relationships different component aspects distributed examined greater detail described finally language individual words sensory input reading stream speech somehow integrated time produce representations meaning scale larger structures etc similarly complex internal representations sequence simpler speech production writing temporally extended sequential processing developed critical understanding aspects language specialized memory systems hcmp pfc discussed likely important temporally extended structures language characterized see networks learn simpler real natural language capture important features behavioral data suggests regularities natural language highly specific case interpretation sentence depend specific words involved again easy account embedded knowledge dependent representations neural network explore interaction semantics grammar sentence model todo language involves range different brain areas models follow identify main brain areas potential interactions discuss relevant aspects output input modalities language assume familiar visual properties words focus details phonology people explicitly familiar details obviously implicit ability produce speech location areas showing relationship relevant specific areas motor output visual word processing psychology familiar brain areas commonly described important language area located posterior prefrontal cortex areas left side brain apparently responsible production speech output including higher order aspects patients damage area said deficit aspect speech unable produce speech function words structure sense think area level higher control area motor outputs speech production area located right temporal lobe parietal occipital area see apparently responsible aspects semantic content language patients damage area said exhibit speech produce semantically speech sense area important language processing representations temporal lobe semantic information encoded occipital lobe visual semantics parietal lobe spatial functional semantics thought complementary deficits surface properties speech production deep semantic properties language production respectively model examine similar set complementary deficits damaged different places domain reading speech observed model terms refer deficits reading ability dyslexia number brain areas damaged produce language cover detail useful generalization language seems areas surrounding part temporal cortex including adjacent parietal frontal occipital cortex further language function appears typically left left individuals effects right damage language catastrophic left damage detailed neural basis language see phonology sounds speech complementary terms speech production characteristics human sound producing terms resulting sound system clearly relationship differences way phoneme produced typically corresponding differences characteristics focus aspect phonology models focus speech production major features responsible producing speech sounds human speech production system based air aka nose pathway called illustrated open air speech sounds open phoneme said whereas closed changing positions things tongue affects properties air come different phonemes defined largely positions parts system discussion details distinctions phonemes critical representations models distributed representation roughly captures similarity structure different phonemes general categories phonemes vowels consonants own sets phonological characteristics discuss turn starting vowels standard system different phonemes labels here require non take character represent models operate things simpler standard english letters represent phonemes adopt known pmsp presented show examples sounds represent dimensional organization vowels according position tongue back representations vowels pmsp phoneme labels features based location tongue position short long note long vowels represented phoneme letters represent vowels form central word provide means air word said vowel sound same true consonants different dimensions vowels vary captured representations shows dimensions based position tongue back dimensions dimensions positions todo here length vowel sound short long long vowel typically similar short vowel shows vowel example try values dimensions activity pattern corresponding vowel network contains unit active group representing back representing possible combinations long short done avoid features active representations network itself develops representations consonants pmsp phoneme labels features based location soft manner vowel liquid vowels consonants typically produced restricting causes distinctive sound depending occurs way done called manner example sound air small opening produced tongue aka sound same closed critical features consonant representations location restricted way done different locations called soft follows phoneme air restricted constant sound phoneme produced lot vowel phoneme liquid phoneme finally involves complete air nose primary phoneme see full consonants features phonological representation word case centered vowel repeating consonants onset coda slots finally need scheme combining individual phonemes representation whole word idea adopt here commonly vowel centered representation slots side onset coda consonants surround word sufficient words models chapter onset coda slots consonant cases different consonants slots alternative scheme blank phoneme extra slots high frequency learning enables systematic orthographic phonological mapping developed words share onset coda position example overlap onset coda overlap purposes reading words represented distributed fashion orthographic visual word recognition phonological speech output semantic areas number separable representational systems different brain areas potentially involved representing aspect word large restrict focus areas essential reading number consisting principally orthographic visual word recognition representations phonological speech output representations semantic representations section explore model based developed demonstrates interaction different areas complete bidirectional connectivity see model provides distributed idea number relevant behavioral phenomena reading model address primarily understand effects brain damage reading ability dyslexia generic term reading problem different types categories reading problems identified type best known dyslexia focus main categories dyslexia phonological deep surface dyslexia phonological dyslexics selective deficit reading nonwords compared reading words terms model shown phonological dyslexia understood lesion direct pathway connecting orthography phonology preserved reading goes via semantics representations nonwords deep dyslexics phonological dyslexics read nonwords exhibit significant levels semantic errors read words semantically related words reading deep dyslexics visual errors reading combined visual semantic errors read presumably via tend errors abstract words compared concrete words reflect semantic representations items summarized implemented model way think deep dyslexics phonological dyslexics damage semantic pathway terms model shown deep dyslexia produced lesion direct pathway connecting orthography phonology preserved reading goes via semantics additional damage semantic pathway semantic errors produced see necessary additional damage semantic pathway semantic pathway come rely presence direct pathway direct pathway lesioned errors reflect similarities semantic representations semantic errors account deep dyslexia reflects damage direct pathway phonological dyslexia surface dyslexia contrast phonological deep dyslexia characterized preserved ability read nonwords problems semantic information written words todo provide example break bit critical difficulty reading exception words exception words words pronunciation follow regularities present words presumably easier mapping learned via semantics directly orthographic phonological representations mappings direct pathway surface dyslexics visual errors semantic errors interpret surface dyslexia resulting damage semantic pathway model shown preserved ability direct orthography phonology pathway reading simulation exercises see forms dyslexia phonological deep surface emerge kinds damage important feature model explore arises interactive bidirectionally connected nature regardless activation orthography semantics phonology flow simultaneously direct indirect pathways allowing contribute activation representations pathways allow division labor processing different types words example reading word based orthographic input direct pathway phonology indirect pathway via semantics participate pathway relatively important word division labor established network learns fact pathway find type mapping easier rely greater extent unless pathways damaged relatively difficult see effects division labor addition obvious effects pathway damage damage direct pathway ability read nonwords model exhibit subtle effects pre division labor certain words different pathways pmsp actually simulate full set pathways showed kind division labor interactive neural network explain characteristics surface dyslexia showed partial lesioning model direct orthography phonology pathway produce right patterns deficit cases surface dyslexia simulated effect semantic pathway actually implementing providing partial correct input appropriate phonological representations training direct pathway model reduced amount input network learned simulating semantic pathway lesion showed network general properties surface dyslexia virtue learned presence simulated semantic pathway direct pathway dependent imagine words direct learn regular words relatively semantic pathway removed exception words particularly low frequency ones dependent upon lesioned semantic pathway phenomenon division labor different pathways function learning network apparent pathways damaged novel mechanism patterns behavior dyslexia neural network models naturally provide explore aspects phenomenon model network completely interconnected kind form particularly interesting result effects possible tell simple story complementary direct semantic pathway lesions producing deep surface dyslexia respectively semantic pathway comes depend partially orthographic input phonological system errors input missing errors semantic nature deep dyslexia due overlap semantically related items difference phonological deep dyslexia account lesion phonological pathway see todo traditional based rule accounts reading reading direct based explicit rules mapping orthography phonology containing based word table exceptions rules related type mechanism traditional accounts seem interactive distributed model approaches debate actually general existence basic pathways discussed direct orthography phonology pathway indirect pathway via semantics central issue debate versus single nature processing mechanisms taking place real contrast explicit rules direct pathway traditional model compared neural network sensitive mappings contrast understanding way divide labor discussed neural network direct pathway sensitive regularity frequency whereas based rule system concerned regularity direct pathway neural network learn regular frequency high irregulars frequency low irregulars rely semantic pathway provide specific word input contrast based rule system expected frequency high irregulars further discrete switch network approach involves interactive processing division labor present model small number words focused capturing level relationships different types capable frequency regularity effects play important role debate nature processing direct pathway well revisit issues next section explore elaborated realistic model direct pathway based pmsp model results model show neural network implementation direct pathway learn high frequency addition irregulars general empirical data appear support neural network approach pmsp model full bidirectional connectivity orthography phonology semantics model based directly developed model deep dyslexia same set words roughly same representations words described original model pathways orthography semantics phonology direct pathway orthography phonology pathway original model assumed completely lesioned deep dyslexia further damage parts original model allowed replicate effects seen deep dyslexics model includes direct pathway able explore range phenomena including surface phonological dyslexia possibly simpler account deep dyslexia model looks essentially identical shown hidden units direct pathway hidden layer semantic pathway hidden layers distinction concrete abstract words described detail semantic hidden layers units semantic hidden layer divided groups corresponding concrete semantic features units abstract semantic features units interconnected layer units same group implement idea emphasized main representational layers orthography semantics phonology connection self recurrent weights implement pattern completion via dynamics encourage layer settle trained patterns initially pattern known reduces number blend responses contain components different words words simulation concrete abstract words roughly orthographic similarity shown phonological representation word train network representations developed orthography phonology semantics set words words concrete physical objects abstract choice word types mentioned deep dyslexics treat types words specifically deep dyslexics semantic errors abstract words concrete ones presumably semantic representations concrete words richer robust damage sets words closely orthographic features orthographic representations distributed representation letter features letter representation appearing right left order letter slots word phonological representations slots individual units slot representing phonemes appear slots word phoneme blank unit represented fact semantics developed set features characterized properties word features concrete words shape main found abstract ones location difficulty quality obviously semantics important distinction concrete words features richer semantic representations abstract words distinction models ability simulate deep dyslexics performance categories words cluster plot semantic similarities shown shows abstract items cluster concrete items main clusters corresponding roughly versus things cluster plot similarity semantic representations different words abstract words central cluster main concrete words correspond roughly things slightly modified representations model consistent simple representations orthography phonology orthography individual units represent letters slots simpler consistent phonological representations eliminate representation similarity similar letters level similarity otherwise networks behavior visual confusions based words sharing letters common captured orthographic representations phonology representations consistent subsequent simulations specifically pmsp phoneme labels described distributed features phoneme individual units original model slots original model repeating consonants scheme reasons described earlier network trained randomly selecting trial main representational layers act input layer setting network learns take aspect words representation map onto corresponding representations standard parameters hidden unit activity open project directory see network constructed stored skeleton keep project file relatively small network takes time train epochs load trained pre network begin overall control panel initial exploration observe behavior network words presented orthographic input layer press overall control panel select process control panel press see activation flow network settle correct pronunciation semantics word large square see record networks performance word including name input word column name word closest output produced network column case same input word pronounced correctly distance actual network output closest event shown column column fact same name input reflected zero column error here indicated meaning same step patterns monitor values sure reading correctly step words pay attention network display point phonological representations activated direct input via orthography indirect input via semantics cases initial phonological pattern subsequently later input describe find discuss behavior network damage pathways now explore ability network read pathways phonology removed action relatively simple manipulation provides insight networks behavior expect semantic pathway leaving intact direct pathway characteristics surface dyslexics read words semantic representations visual errors direct pathway reading semantic pathway potential semantic confusions arises expect replicate effects deep dyslexia finding semantic visual errors note phonological dyslexia form deep dyslexia explore perform amounts partial damage lesioning entire pathways hit overall control panel select pathway lesion actually units network structure lesion entire layer note entire pathway network rely intact means errors expect associated properties intact pathway lesioned example lesioning direct pathway network rely semantics allowing possibility semantic errors extent semantic pathway things right missing direct pathway completely lesioning semantic pathway itself lead semantic errors semantic information left errors based test network process control panel patterns want network reading overall count number column look counted notice column indicating phonological output closely known word larger threshold output novel blend phonological features correspond closely words training set fact say properties network errors compare word network produced input word produced word similar input word called visual error based visual properties semantic properties word lesioned semantically networks errors visual now lets try opposite lesion selecting time again errors visual errors common sense cluster plot common sense determine semantic similarity response input word semantic errors count number cases think true now scroll bar bottom scroll log see columns right hand side log reflect automatic networks responses according visual semantic overlap word patterns set columns applying concrete words set abstract ones input orthography response orthography determined response phonology overlap letters error visual error semantic errors difficult code variable amount activity pattern semantic representations overlap measured normalized product patterns semantic error formula goes overlapping non patterns completely overlapping ones value good job including cluster plot shown column semantic visual errors blend responses column apply summarize results far seen lesion semantic pathway results purely visual errors lesion direct pathway results combination visual semantic errors order approximation observed surface deep dyslexics respectively shown pmsp model surface dyslexics actually likely errors frequency low irregular words manipulation simple corpus words examine aspect performance critical difference here surface dyslexics semantic errors deep dyslexics errors visual errors now perform realistic form lesion sequence series different lesion types corresponding different layers semantic direct pathways units increments layer effects reading performance lesion types semantic pathway hidden layers replicate effects surface dyslexia next direct pathway replicate effects phonological dyslexia high levels produce deep dyslexia next lesion types semantic pathway hidden layers again pathway corresponds model deep dyslexia explored finally last lesion type direct pathway hidden layer again full lesion semantic pathway produce extreme form surface dyslexia included particular motivation begin lesioning select overall control panel choose see graph appear process control panel graph right display performance automatically error types function level damage type graph left display results last lesion level different type lesion lesion types axis order described starting press process control panel network display update times beginning lesion turned final corresponding actual lesion network lesion items read network want keep graph display results see point note level damage intact network interesting next time network display updated notice units layer missing reflects partial damage increase next time notice right graph begins reflect presence visual errors concrete abstract words damage levels axis value damage levels axis value right graph start next layer proceed layer lesion types final damage performance cases shown left graph log axis indicating lesion type again values correspond sequence layers described correspond indirect semantics pathway hidden layers same time complete lesion direct pathway value corresponds lesions direct pathway complete lesion indirect pathway able sense results obtain random nature lesion need average different instances lesion type reliable statistics expected types errors lesion type process ran configured run type lesion run function control panel lesion produced read easier main results error types semantic pathway lesions intact direct pathway point shows results different lesions semantic pathway intact direct pathway lesion types concrete abstract plotted separately generally case complete lesion semantic pathway network visual errors generally consistent surface dyslexia expected semantic errors lesioning semantic pathway remember intact direct pathway providing orthographic input directly phonological pathway input generally phonological output related orthographic input prevents semantic errors words think damage semantic pathway causing want semantic errors direct orthographic input well see moment direct input removed semantic errors interestingly lesions semantic layer itself produce errors concrete versus abstract words understood considering division labor develops semantic system better able process concrete words compared abstract ones direct pathway take processing abstract words effects damage understood complementary ways perspective damaged semantic pathway concrete words depend damaged pathway perspective intact direct pathway abstract words depend intact pathway shows general effects semantic variables network specific confusions words similar semantics semantic errors aware data surface dyslexics existence effect humans possible factors frequency regularity effect error types semantic pathway lesions conjunction completely lesioned direct pathway point shows same semantic pathway lesions previous figure time conjunction complete lesion direct pathway lesion types corresponds type lesion studied model deep dyslexia levels semantic pathway lesion now see semantic errors visual errors relatively large number errors pattern errors generally consistent deep dyslexics kinds errors comparing previous figure see direct pathway playing important role generating correct responses particularly semantic confusions semantic pathway otherwise compare bar graph corresponding case direct pathway lesion damage semantic pathway subsequent bars additional semantic pathway damage appear necessary produce semantic error deep dyslexia explain direct pathway lesion leads semantic errors thing notice figure relative number semantic errors concrete versus abstract words characteristics deep dyslexia patients semantic errors abstract words relative concrete words concrete abstract words number semantic errors representations types words error types direct pathway lesions intact semantic pathway full complete semantic pathway lesion shows effects direct pathway lesions lesion type lesion type intact semantic pathway lets focus case intact semantic pathway full figure notice smaller levels damage relatively semantic errors produced errors visual corresponds well phonological dyslexia assuming damage direct pathway pronunciation nonwords presumably read via direct orthography phonology pathway test aspect model words scale large model direct pathway described next section produces nonword pronunciation deficits event relatively small amounts damage interestingly level damage increases model increasingly semantic errors performance high levels damage provides good fit deep dyslexia characterized presence semantic visual errors plus pronounce nonwords discussed context phonological dyslexia occurs network rely semantic pathway previously intact direct pathway now pattern overlap semantic representations causes semantic confusions reading further see aspect deep dyslexia data greater proportion semantic errors abstract words concrete ones add semantic visual semantic errors abstract words finally last case direct pathway damage completely lesioned semantic pathway produces visual errors clear relationship model illustrates words represented distributed fashion set different specialized areas layers damage layers produces behavioral results similar observed different types dyslexics general framework interactions orthographic semantic phonological representations model elaborated subsequent models model specific pathways greater detail general point results difficult tell damage based patterns error example levels direct semantic pathway lesions produce roughly visual errors larger direct pathway lesions semantic errors start appear important demonstration purely behavioral approaches cognitive neuroscience require evidence additional including computational models previous model included direct pathway orthography phonology small number words real exploration important issues regularities exceptions mapping larger number words needed regularity complexity nature mapping english described mapping written word spelling orthography sound phonology studied extensively roughly english words series influential models models central issues relationship processing regular versus exception words specifically single system process ability replicate systematic performance human novel nonwords depends properly encoding subtle regularities letter typically pronounced understand issues understand nature regularities exceptions regularity defined mapping letter phoneme present relatively large number examples language example consider pronunciation vowel words hint etc case pronounced same way short vowel regular pronunciation group words define regularity called neighborhood contrast word pronounced long vowel exception word critical note respect vowels regular pronunciation depends letters word example words mind find etc form neighborhood regular pronunciation long form vowel time further consider example familiar rule regarding effects final produces regular neighborhood words long vowel pronunciation fine etc non english aware regularities english simple form good reason believe single system necessary perform mapping factors need taken account determine regular response irregulars thought extreme examples process pronunciation dependent configuration entire word general continuum regularity neural network appropriately trained weights deal continuum naturally taking appropriate contingencies account response contrast traditional based rule account direct sound spelling pathway requires elaborate collection rules deal properties mapping existence regularities described sense english asked pronounce novel nonword regularity pronounce complex nature regularities systematic behavior appropriately sensitive conjunctions letters cases pronounced pronounced conjunctions initial pronounced same correct way independent letters system sensitive wrong letter conjunctions respond word analogy familiar word ignore pronounce existing word correct generalization performance nonwords depends critically developing representations appropriate combination conjunctive depending letter conjunctions combinatorial allowing arbitrary combinations initial combined goal neural network models reading show appropriate representations develop learning pronounce known english words significant neural network model reading developed model learned perform mapping english words due choice output input representations model generalize nonwords well meaning produce same kinds systematic human subjects produce same word inputs representations model called due conjunctive nature ideas represented conjunction letters phonemes entire word input represented activation combinations sequential letters phonemes including initial final blank represented contained word represented useful capturing pronunciation letters discussed allow combinatorial representations letters individual letters phonemes encoded entirely different units fact letter pronounced same way regardless surrounding context model developed pmsp model problems representations single units represent letters phonemes regardless surrounding context combinatorial representations specifically divided word parts onset vowel coda set phoneme letter representations parts example word onset plus unit described unit vowel units coda mapping letter pronunciation systematic same representations again regardless surrounding context think lead exact nature word order information part actually encoded input represented turns isnt problem english example typically different words differ letter order parts onset vowel coda pmsp model added conjunctive representations specific pronunciation consequences pronunciation extra unit active referred similar scheme developed output phonology combining combinatorial conjunctive elements short pmsp model tuned hand representations significantly simplified learning orthography phonology mapping model provides useful demonstration importance appropriate representations fact representations researchers learned network itself remains problem model present here avoid problem network appropriate representations own model based ideas developed invariant object recognition model perspective object recognition model see pmsp models same tradeoff position invariant representation recognize features objects multiple locations emphasized pmsp model conjunctive representations capture inputs accomplished conjunctive units pmsp satisfying resolution tradeoff object recognition model development hierarchy representations produce increasingly invariant increasingly complex representations expect same approach applicable word reading well reason believe word recognition special case object recognition same neural pathways same basic mechanisms explore here reading model based object recognition model basic structure orthography phonology reading model orthographic input representations words appearing possible locations letter position input next hidden layer receives slots forms locally invariant conjunctive representations next hidden layer representations map standard slot centered vowel repeating consonant phonological representation output basic structure model illustrated orthographic input presented string letter activities letter slots slot units representing different letter plus additional space simulation present related simulations words presented positions allowed entire word fit word presented starting slot analogous input patterns object recognition model number different positions smaller next layer input called layer representations object recognition model units window onto letter slots units develop locally invariant representations unit letter slots conjunctive representations encode local order information letters words units develop tuned hand pmsp input representations next hidden layer fully connected hidden layer phonological output performs mapping orthography phonology phonological representation standard slot centered vowel repeating consonant representation described actual network skeleton view slots input units total units hidden layer units hidden layer slots phonology output units total pmsp corpus words train network word presented according square actual frequency square frequency pmsp models necessary here enable network train frequency low words reasonable amount time pmsp demonstrated similar results actual square frequencies word appear positions input entire word fit slots words different positions longer ones word itself position word combination subject frequency manipulation words appear activity constraints set activity hidden layers proportion hebbian learning set value necessarily small prevent constant pressure hebbian component error term successful learning corpus further object recognition model large networks necessary reduce learning rate epochs prevent subsequent weight changes cause interference previous ones parameters standard standard non parameters activity smaller hebbian learning learning rate decrease time characteristic larger networks trained large note simulation requires minimum run open project sound spelling directory again large network looks shown constructed well build load trained network took couple train overall control panel network window select well see network read words presented standard list probe words developed locate process control panel press word network best presented left edge input read word input pattern individual letters activated slots slots corresponds location contains units corresponding letters unit blank slots begin unit lower left right right input patterns verify best fact pattern presented output pattern produced correct pronunciation verify need view phonology patterns compare phoneme slots output appropriate patterns press overall control panel select initial slots onset find clicking button consonant window see pattern matches slots output next lets look coda click button consonants window scroll button number select middle button left button patterns match produced network finally select again select look vowel pattern click button see matches central vowel slot pattern probe words test network settling times regular consistent means inconsistent whereas regular inconsistent versions examples ambiguous clear regularity exceptions exceptions regularities need pattern matching process time simulator iconify consonant vowel windows locate shown column components orthography basic phonology repeated consonant phonology network actually produce special code indicating shows mean return later explore settling time properties network pronunciation network actually produced shown column show correct output produced note output contain indicates phoneme position exactly match valid phoneme patterns shown number cycles took network settle error todo show epoch network times notice best input appears positions input differences input location network capable producing correct output spatial invariance coding explored requires network maintain information local letters best example treat entire pattern same regardless appears well see moment network developed same general solution problem previous combination locally spatially invariant conjunctive encoding continue patterns want switch network updating update trial cycle overall control panel select observe error items lower frequency irregular network pronounce words correctly now lets explore connectivity network click left portion orthography input click left button see skeleton network shown clicking back forth skeleton weights clear units receiving left letter slots letter slot group units click back left button verify groups units receive overlapping groups letter slots click units pay attention patterns weights thing notice cases unit strong weights same input letters slots press control panel select particularly good example unit unit selected middle region layer unit receives letter slots providing locally spatially invariant representation letter units layer object recognition model networks hidden layer learns pronunciation consequences associated letter particular position representation learning automatically generalize locations letter kind thing pmsp tuned hand input representations accomplish see network learned own link unit hidden layer clicking hidden layer unit unit projects strongly selecting again selecting time original unit plus examining moment selected addition single hidden layer unit see unit strongly driven unit viewing receives clear pattern central vowel slot phonological output layer weights network generally symmetric interpret unit projects looking receives sure lets select show sending weights switching back forth see generally exactly symmetric hebbian learning component symmetric driven error interpret phonological output pattern hidden unit produce need again select choose phoneme particular hidden unit produce possible outputs letter overall pattern connectivity good sense further see hidden unit receives units code letter weights unit right side layer strongly interconnected hidden layer unit click back viewing see unit represents letter input locations example type invariant coding found letter case output hidden unit projects features corresponding coda verify couple phonemes coda performs kind coarse coded representation multiple related mappings looking weights unit input shows receives letter letter similar shown looking patterns finally look complex input pattern receiving vowels plus letters serve activation coda output mapping vowel present accounts production consonants general level unit shows complex conjunctions input letters represented here layer units object recognition model model important exercise weights network seems learning right kinds representations allow good generalization representations similar layer object recognition model combine spatial invariance conjunctive feature encoding able obtain insight looking representations easily further networks complex activation dynamics picture difficult figure processing input know nature mapping problem itself lots subtle forces determine pronounce word finally fact easily interpret units weights place due hebbian learning causes weights reflect probabilities unit occurrence networks weights relatively clear example representations sense terms output input mapping performed specify letters units encode hidden units combine phonemes hidden units produce output relate analysis need spatial invariance conjunctive encoding next test networks ability generalize knowledge pronounce english words pronounce nonwords regularities spelling sound mapping number nonword sets exist sets pmsp test model set nonwords lists derived regular words exception words set constructed determine nonwords actual words pronounced better set lists control list list set nonwords derived regular exception probe word lists test network earlier start testing model nonwords back nonword list click back network looking see network correctly pronounced nonword producing turn network display see nonwords errors possible see produced similar real word different reasonable pronunciation word summary nonword reading performance raw values single provided output whereas allowed shows results alternative outputs consistent training corpus allowed errors regular exception nonwords phonological representation repeating consonant phonology cycles number cycles network took settle output actual output produced network explanation network output based training corpus errors control nonwords columns previous table errors nonwords columns previous table note described exact same network testing show slightly different errors fixed later errors nonword lists summarized output different locations input seems network solve task pmsp network view file cases error note error single invalid phoneme total error output pattern closer correct pattern didnt error error tried determine network produced output cases output reflected valid pronunciation present training set didnt happen pronunciation list following pmsp computing networks overall performance total model pmsp human data shown clearly present model performing roughly same level humans pmsp model conclude network capable extracting complex subtle underlying regularities regularities sub present mapping spelling sound english applying nonwords model single pathway processing regular exception words direct traditional theories hold direct mapping spelling sound occur regular words via application explicit rules case model exhibits significantly performance low frequency irregulars word categories expected discussed previous model indirect via semantics important words todo clear semantics phonology mapping showing separable inflectional component semantic representation past tense onto regular phonological inflection add section explore model mapping semantics phonological output presumably trying internal output end reading system shown explored general mapping semantic representation word phonology essentially random relatively rare cases interesting task issue inflectional know change inflection word convey different aspects meaning example want indicate event occurred past past tense inflection adding relevant verb think tense component semantic representation gets onto appropriate inflectional representation phonology idea illustrated types inflections applied english verbs include adding indicate person adding past adding case reading model explored previous section irregular general rule add past tense inflection went similar issues regarding processing versus irregulars come domain shaped curve irregular tense past inflection production axis shows minus overregularization rate axis beginning overregularizations right starts rate least years assume point significant frequency represented line end graph turns issue inflectional specifically past tense played important role development neural network models language issue phenomenon known shaped curve irregular tense past inflection due overregularization shaped curve performance initially good gets worse middle gets better exhibit shaped curve producing correct inflection irregular verbs initially correctly produce irregular inflection saying went long period overregularizations irregular verb regular went finally learn treat irregulars irregulars again shows overregularization pattern speech serves primary phenomenon figure clear overregularization phenomenon occurs relatively low frequency considerable individual emphasized subset studied show strong evidence early correct production irregulars subject individual well overregularization phenomenon originally reflecting action based rule system gets bit early language developed neural network model showed shaped overregularization curve argued networks sensitive regularities output input mapping tendency problems model shaped effect apparently due manipulation training set large number frequency lower regular words introduced network learned smaller number frequency high words irregulars regular words caused network start treating irregular words overregularization number network models past tense learning developed entirely successful capturing essential properties shaped curve plausible manner introducing manipulations work example model widely fully account limitations depends critically manipulation training environment starts original model small number high frequency irregular verbs adding new verbs add training set clear adding new regular verbs triggers overregularization network manipulation discrete change basic problem isnt network itself driving overregularization further unable replicate original results realistic corpus based english opposed artificial corpus original model problem model task mapping base phonology representation form particularly plausible compared semantics phonology version here possible reason prior explore here models based backpropagation algorithm important limitation overregularization data show features distinguish leabra algorithm pure backpropagation competitive kwta activation dynamics hebbian learning contribute producing shaped overregularization curve couple ways understand pure driven error backpropagation learning likely produce shaped curve general level performing error produce error curve shaped detailed level network learned achieve mapping semantic representations corresponding past tense inflection phonological output see part learning task influence weight updates weight updates driven purely error signals network regular inflection pressure further shaped curve likely short arguments clear hebbian learning activity competition enhance shaped curve again general level constitute additional constraints otherwise driven error learning specific level seen playing important role hebbian learning drives weights represent correlated activity provide sustained influence strong correlation tense past aspect semantics regular past tense inflection regularity continue influence learning network regular inflection hebbian learning care network producing correct output role activity competition obvious according results important hebbian learning way thinking effects terms competition hidden units favor regular mapping favor irregular limited number units active regular mapping occurs expect units representing regular mapping competition frequency irregular verbs producing overregularization network activity competition similar competition regular irregular mapping taking place weights dynamics activation settling pattern activation dynamics allow forth back mappings gradual weight changes allow explore ideas model past tense mapping illustrated compare results leabra standard backpropagation network test idea distinctive features important producing realistic shaped curve past tense network mapping semantics phonology basic structure model shown semantic input hidden layer phonological output layer general structure approach model based work provided initial corpus words model network trained produce appropriate pronunciation different english verbs irregulars different inflections total training items words presented square actual frequency general scale simulation roughly same order previous model inflectional forms english verbs different ways verb english shown inflections thought extra bit information verb semantic input model main components semantic pattern represents basic meaning verb constant inflections basic inflectional component represents senses verb produced attempt accurately represent semantics mapping phonology largely random anyway random bit patterns units semantic layer encode pattern sequential groups units represent inflections next units total semantic input units phonology representations standard centered vowel repeating consonant ones described additional slot end additional inflectional phoneme needed inflectional component purposes human output included extra phonemes represent inflections occurred last slot same features combination same features person sounds changes past same features hidden units active time level sparseness useful learning arbitrary mapping semantics phonology due need keep different patterns relatively separate saw hebbian learning level ran network pure driven error learning generic backpropagation network comparison open project directory again project stored network skeleton form overall control panel load trained pre weights start well observe network produces phonological outputs semantic patterns locate process control panel see semantic input presented network settling produces activation pattern phonological word base form know looking activations provides actual output produced column identity input pattern column event coded initial number representing index verb followed target pronunciation followed code indicating inflection type order regularity verb irregular regular here represents base inflection irregular verb see errors network took cycles produce output looking back network notice active units next last row semantic input units indicate base inflection produced now next word see units adjacent previous now activated indicating past inflection produced case continue remaining inflections verb now lets regular word press button process control panel enter todo now see network learned task lets try connectivity determine working interested tense past mapping well focus find hidden units selective past tense inflectional semantics click network sending weights click tense past inflectional semantics unit unit left last row click adjacent tense past units notice good consistency hidden units activated reflecting work hebbian learning rule notice different hidden units activated click last base inflectional semantics unit unit left last row person inflectional semantics unit unit right last row now lets back past tense inflectional unit strongly connected hidden units subsequent pressing menu window appears type enter comparison value see hidden units now selected now click left units see unit favor particular onset phonemes clear particular coda phonemes see phonemes correspond lets click select window containing phoneme patterns last slot inflections click event button inflectional phoneme left side window see pattern inflection click middle button blank event see pattern active phoneme produced clear hidden unit favors outputs now lets determine phonemes last next slot iconify inflections window press again select click middle button buttons see unit produce patterns last next phoneme slot similar weight patterns found hidden units clear subset units encoded regularity present tense past mapping presumably units activated irregular words able regular inflection see competition takes network learns production inflection same technique past tense inflection active units code appropriate inflectional phonological pattern describe steps took reach answer plot overregularizations total responses network learns clear initial period overregularizations network producing valid responses words overregularization minus number overregularizations begins eventually relatively provides fit data interesting see network learned task relevant empirical data course time overregularizations network learned output training pattern analyzed outputs code different kinds errors primarily interested overregularizations irregular past tense words counted number times network said things went results network human data note following overregularization plotted minus proportion overregularization errors characteristic shape evident graph graph networks performance plots proportion valid responses network showing network achieves substantial level responding prior onset overregularization note measure indicates extent network producing kind output overregularization errors take account inflectional errors tells extent network produce semantic input pattern network demonstrates correct early period irregular verb production critical aspect empirical data previous models capture manipulations training corpus parameters further overregularization continues low extended time period characteristic human data total number valid responses network prior overregularization provides measure extent early correct period prior onset overregularization hebbian learning leabra inhibitory competition leabra appear contribute relative generic backpropagation network assess influence hebbian learning activation competition length correct early period compared different types networks standard explored leabra hebbian learning leabra network hebbian learning generic backpropagation network ran networks type initial period training point overregularization starts counted number valid responses prior overregularization results show hebbian learning extent activation competition important producing longer correct early period understood terms role hebbian learning activation competition producing greater specialization representations early frequency high irregulars encoded overregularization total number overregularizations training inhibitory competition leabra appears important contribution overregularization relative generic backpropagation network clear difference significant appear hebbian learning playing significant role here finally counted total number overregularizations produced full training epochs single types networks computationally practical run results shown show leabra networks considerably backpropagation network interestingly appears work done activation competition here leabra network hebbian learning produces slightly overregularizations hebbian learning based single run impossible determine significant difference appears hebbian learning least adding top activation competition substantial difference leabra networks backpropagation due chance described activation competition seems overregularizations presumably allowing dynamic based activation competition units encoding regular irregular mappings way understanding terms output multiple network explored multiple outputs case irregular regular inflections irregular verb activation competition present possible network produce discrete different output patterns different trials competition network patterns pattern correct case backpropagation network learns overregularization results exploration show model comes closest capturing characteristic shaped learning curve hebbian learning activation competition error driven learning hebbian learning appears specifically important early correct period activation competition appears specifically important achieving large number overregularizations overall findings detailed evaluation models fit human data complicated factors human data itself data cases starting well onset language production difficult determine shaped curve truly seen subset individuals difficult achieve valid mapping network learning human learning network completely focused semantics phonology mapping trial learns rapidly compared example network achieved well valid responding irregular trials training difficult know map number trials onto learning experience assumes speech rate word contributes learning semantics phonology equivalent time period achieve valid responding amount entire time shaped curve network complete corpus roughly trials years observed empirically apply kind generic scaling networks performance likely major part difference due complexity larger language task performing expect relatively simple models provide detailed picture finally advantage model relative completely training environment networks gradual learning environment results gradual nice contrast external task modeled semantic production task undoubtedly case semantic representations gradually developing same time period place important constraints learning process captured well existing corpus techniques best interpretation model demonstrates network environment exhibit shaped curve principles hebbian learning activation competition well number reasons model need additional properties order better fit human data word occurrence illustration semantic information distributed number specific processing pathways represented here different sensory modalities representations orthography phonology associated corresponding distributed activation patterns semantics figure adapted general think semantic representations word representations distributed different specialized brain areas notion distributed semantics well captured adapted shows representations items etc different specialized processing pathways pathways represented sensory figure actually imagine number types sequential oriented task etc contribute overall distributed representation item essentially pattern activity brain contributes specific item information considered semantic representation section explore model semantic representations takes advantage word occurrence statistics form distributed representations word model implements larger distributed semantic network emerge information words tend occur speech reading shown method call semantic analysis word occurrence statistics semantic representations good job human semantic example system trained text psychology textbook able multiple choice based text performance roughly correct well short good nonetheless simple procedure perform well interesting aspect word occurrence approach captures word association semantics semantics example definition word word words highly appears important capturing structure human semantic memory example semantic priming word presented followed word read faster word expect hebbian model learning provides natural mechanism learning word occurrence statistics fact method involves performing essentially sequential principal components analysis spca correlation matrix words expect cpca hebbian learning develop useful representations follows explore simple cpca network trained word textbook probe resulting network see captured important aspects information presented text model shown input layer unit word projects hidden layer containing units training network follows trial units representing words present individual activated network settles hidden activation pattern function input settling cpca hebbian learning takes place encoding conditional probability input units active hidden units repeated text multiple times key insight method works comes fact whenever similar input patterns presented likely similar hidden units activated extent reliable correlations subsets words hidden representations encode further network learn words similar happen appear same words likely occur similar words likely tend activate common subset hidden units hidden units learn similar high conditional probability words words presented network produce similar hidden activation patterns indicating semantically related textbook language contain detailed information phonology interested learning chapter revisit fundamental issues computational cognitive neuroscience covered previous chapters towards integration chapters explore remaining challenges field addressed next generation models bring conclude ways computational models towards development cognitive neuroscience covered broad range text ions individual neuron level learning complex cognitive functions planning goal understand findings different levels interactive approach emphasizes connections neurobiological cognitive computational saw individual neurons act detectors constantly inputs responding matches pattern weights synaptic ion channels neuron compute balance excitatory inhibitory leak currents reflected membrane potential potential exceeds threshold neuron fires sends inputs neurons network rate firing encoded continuous variable activation showed individual detectors network exhibit useful properties provide building blocks cognition network properties build ability individual detectors compute balance excitatory inhibitory inputs properties include input patterns patterns emphasize distinctions collapse bidirectional top bottom processing pattern completion amplification bootstrapping inhibitory competition activity regulation sparse distributed representations multiple constraint satisfaction networks updating constraints possible built basic network properties showing neurons network learn weights according activation values sending receiving units learning local variables results coherent beneficial effects entire network analyzed main learning model learning task learning model learning causes network capture important aspects statistical structure environment task learning enables network learn specific tasks learning complementary achieved known properties synaptic modification mechanisms extra mechanisms learn temporally extended sequential tasks mechanisms main forms learning updating internal context representations temporal reinforcement based mechanism overall cognitive architecture builds mechanisms described different brain areas defined basis fundamental tradeoffs arise basic neural network mechanisms tradeoff captures distinction hippocampal system rest cortex terms learning rate system learn arbitrary things rapidly extract underlying regularities environment tradeoff captures difference prefrontal cortex posterior cortex terms ability perform robust active maintenance active maintenance suffers representations interconnected activation away memory useful kinds processing pattern completion satisfaction constraint generally posterior cortex understood terms relationship specialized brain areas generally posterior cortex seen set specialized organized pathways transformations building upon provided good example organized sequence transformations lead ability recognize objects spatially invariant fashion level units input combining features integrating different locations sizes model recognize individual objects well confused multiple objects present adding spatial representations interact object processing pathway enables system sequentially process objects scene containing multiple objects accounts effects lesions spatial processing pathway performance posner spatial task level cortical visual processing neurons encode correlational structure present visual images explored different ways neural networks implement memories weights activations saw priming tasks cortical memories taking form small weight changes produced gradual learning residual activation network saw basic cortical model short capturing human memory consistent discussion fundamental computational tradeoffs regarding tradeoff rapid learning arbitrary information slow integrative learning regularities find basic cortical model learning regularities perceptual domain performs required rapidly learn novel information suffers catastrophic interference hippocampus sparse pattern separated representations avoid interference learning rapidly regarding tradeoff active maintenance rich interconnectivity find basic cortical model fails hold onto information delays face interference prefrontal cortex isolated representations finally interaction activation based weight memory complex saw model task language explored requires number specialized processing pathways representations interact build perceptual pathways representations specialized pathways operate according same principles pathway saw distributed pathway model word representations account patterns dyslexia damaged focusing direct pathway orthographic input phonological output saw network captured regularities enable generalize pronunciation nonwords same way humans focusing pathway semantics phonology saw network learn regularities exceptions inflectional produced shaped overregularization curve learned explored model semantic representation learning based occurrence statistics words large text resulting semantic representations capture relevant aspects word similarity finally saw network learns next word produced sentence learn certain important aspects grammar including verb embedded chapter took challenge applying biologically realistic neural network mechanisms explored previous chapters modeling higher level cognition saw notion controlled processing implemented combination top biasing prefrontal cortex binding hippocampus accounts general kinds things higher level cognition higher level focusing role prefrontal cortex pfc viewed maintaining internal context representations explored sequential processing saw simple model account normal performance task important next step see dynamic control maintained pfc activations via gating implemented dopamine leads flexible task performance compared simple based weight task learning explored context simple task dopamine control pfc representation updated basal ganglia important determining pfc representations updated possible see mapping dynamic controlled gating mechanism firing production system neural network models higher level cognition able production system models higher level cognition well learning major challenge domain higher level cognition understand appropriate pfc representations learned experience enable flexible task performance brief summary clear computational models lot say different aspects cognitive neuroscience excited range different phenomena current framework address number important challenges remain future work address section attempt identify challenges sketch ideas addressed history neural network modeling dominated periods extreme extreme past entire approach based limitations research subsequently overcome limitations rule delta learning taking similarly approach find list future challenges long issues here cover question entire hand existing ignore important limitations encourage balance approach simply remaining challenges feel recent years balance same balance continue future years need simplification neural network modeling strengths strength allows extract example essential insight biological properties important particular behavior simplification largely details simplification facts relationship different levels modeling varying level detail range cognitive phenomena covered modeling extent different levels detail phenomena beneficial mutual constraints models true individual researchers face tradeoffs important study point time larger field allows multiple parallel approaches see key ultimately solving simplification problem different models different levels analysis detail overlap model simplification further detail extent models compared area overlap simplified model taking account behavior detailed model least limitations simplified model known examples effective multiple overlapping models found places book compared performance unit fires discrete spikes computes rate code compared detailed implementation inhibition inhibitory interneurons kwta simplification examples showed reasonable detailed case clear differences possible explore models otherwise language chapter explored number phenomena simplified model main representations involved reading larger realistic models explore detailed aspects performance specific pathways simplified model possible explore certain aspects behavior effects damage manner provided general framework detailed models examples sample kinds benefits multiple levels analysis sections follow discuss areas detailed models reading neuroscience brain research etc biological details biological properties largely ignored models general answer question powerful require lots biological actually implement undoubtedly fail capture mechanisms capture main effect generally speaking useful relate functional properties detailed mechanisms simpler find exactly differences matter following particularly relevant powerful simplification activity regulation kwta function likely lot biological necessary keep neurons firing right activation addition basic feedforward feedback inhibition includes lots channels lines discussed section regulation self chapter factors expression major simplification neural network models initial networks simply starting random connectivity additional topographic constraints network already constrained receive particular types inputs huge largely problem biology cognitive neuroscience understand biological structure complex sequence interactions genetic chemical surface etc significant portion initial brain kind process development brain greater degree subject influences experience know specific examples detail early visual system example random noise coming retina plays important role neurons line set initial configuration ends learning begins undoubtedly understanding brain gets initial configuration critical understanding later effects learning typically simplified control processing learning inputs presented controlled fashion networks activations reset input phases input plus minus learning constrained operate appropriate information real system simple events defined pre activation occurs controlled tests simulations show activation important rapid successful learning needs explored context spiking models generally exhibit prior states hysteresis code rate models constant output activation interconnected units existing activation patterns activation important discrete spiking models biological reality potential implementation explored well take issue phases learning learning generally later section argued exploring aspects biological complexity greater detail strong argument thinking neuron simpler biology suggest summarize neuron shouldnt able take advantage further complexity brain noisy needs robust simple graded signals actually useful neurons output signal receive inputs isnt complex processing inputs useful form single output shown detailed spike timing cortical neurons seem convey meaningful information keep mind cases simpler things work better currently technique images large number smaller images serve pixels larger image close images looks random images step back eyes larger overall picture simpler describe overall image picture describing properties individual component images similarly eyes biological details produces simpler relevant picture neural computation obviously take pieces put right look simple models text focused cortex including hippocampus exception way thalamus basal ganglia reality course cortex operates context large number brain areas including areas computational models including fact generally determined means complex specialized neuron types easily terms patterns weight values otherwise generic unit model case cortex cortical model simply learning playing dominant role producing useful model difficulty modeling additional brain areas raw increase computational complexity involved adding additional brain areas model missing brain areas seen providing basic inputs cortex effects nature processing cortex explored role dopamine context prefrontal active maintenance learning addition number similarly important roles example ability cortical neurons remain particular task function task performance variables important neurotransmitter regulation states versus part possible ignore extent assume cortex completely focused task hand richer complete model cognition require factors play critical role human performance psychological task situations argued brain kind similar self structure coarse models multiple brain areas relatively units same principles fine models individual columns large number units number good reasons believe generally true multiple models approach test scaling assumptions test well detailed fine model coarse clearly coarse model represent small fraction information compared fine overall dynamics relevant behavioral characteristics ultimately limitation computational power impossible fine models multiple brain areas computers computational rapidly example computer power basically period book written increasingly possible implement large fine models test simplified scaled models provide reasonable approximation realistic presented biological mechanisms implement driven error task learning showed models learn basis mechanisms models simply imposed necessary minus plus phase structure required learning occur provided target outcome patterns output layer open challenge demonstrate expectation outcome representations actually arise naturally simple motor perceptual system operating simulated environment particularly perception outcome happens same layers represented expectation challenge address question system phase plus perform learning resolution challenges further modeling work evidence brain kind switching phase proposed underlie driven error learning well signals signal learn evidence comes updating spatial representations parietal cortex function neurons actually new input coding function motor same neurons update reflect actual input coding result motor specifically parietal neurons represent see movement movement update reflect actually see movement actual representations clear expectation outcome difference drive learning issue learn related dopamine examined role driving learning based differences expected obtained reward requires dopamine signal occur difference expected versus obtained outcome reward todo add finally body evidence suggests cortex involved detecting errors seems likely brain area plays important role driven error learning exact role remains temporally extended tasks planning relationship production system kinds models etc initial important issues described clearly work needs done important question higher level cognition necessarily requires scale large models important aspects captured smaller models words higher level cognition happens large critical cortex gets going basic principles scaled models clearly believe latter least partially true number problems neural network models past successfully existing models generally constitute major challenges discussed problems represent ongoing challenges field continue pay attention address future models neural network models theoretical big parameters nice here add additional work statistical learning provided purely theoretical account tradeoff rapid arbitrary slow integrative learning related complementary roles hippocampus neocortex learning memory arguments work based general principles particular truly theoretical principles demonstrate required assumptions etc argued theoretical importance network models issues analyzed terms general principles apply kind learning mechanism based statistical properties environment models understood terms principles todo example comes provided implemented model effects regularity frequency learning reaction time number models produced form relatively implementation shows effect clear principled understanding increasingly rare field improve important maintain focus understanding models performance terms set theoretical principles particular issue closely related previous book times planning writing algorithm leabra possibly write textbook based completely new algorithm level true leabra new algorithm widely specific version leabra here didnt exist final form well writing book implementation general level essential principles define leabra algorithm number times different single algorithm entire book computational cognitive neuroscience particular algorithm algorithm connection saying collection principles important combined appropriate fashion help understand wide range phenomena ion channels language processing higher level cognitive planning distinction implementation leabra general principles based upon discussed similar distinctions clearly came side balance multiple levels analysis feel same way here results obtained book say specific implementation general principles found models largely successful conclude implementation principles pretty good undoubtedly better performance way obtained specific cases different implementation different set principles time know obvious alternative algorithms sets principles provide good fit wide range performance ones chosen book emerge adopt subsequent book related points traditional neural network models typically backpropagation networks learning learned emphasized people neural networks solving practical problems todo number points here brain complex dynamic system likely easy restricting models easy understand likely good approach number reasons believe standard backpropagation algorithm interpretation typically highly learning weights themselves relevant aspects task discussed length showed hebbian model learning provides generally useful bias advantage producing constrained easily weights advantages hebbian constraint interpretation text particularly good examples object recognition model reading tense past language models resulting models provide nice balance computational power models currently computationally weak localist models related focus number free parameters neural network models argument parameters models learn showing learn impossible know parameters crucial learning essential emphasize challenge modeling neural networks basic nature free parameters neural networks people fail generally people say train network true interesting aspects models based aspects network performance generalization new problems response damage example fact network behaves certain way damaged due fact trained way specifically trained perform correctly observed resulting performance basic computational properties model damaged further learning shapes weight parameters model researchers precise control learning precise understood well set principles shape networks weights interaction environment understanding principles network learns apparent complexity model reduced relatively simple application small set principles important aspect models included text generally exactly same set standard parameters models differ parameters differences based principled parameter example main parameters models activity level hidden layers saw activity level important tradeoff learning specific information individual patterns sparse activations versus integrating patterns distributed activations neural networks etc learning terms representations automatically produce generalization object recognition model see showed networks form systematic internal representations enable good generalization introduce completely novel unit network know people novel combinations existing representations via distributed represent novel items care real humans mechanistic models tell real binding comparison multiple time hierarchical structure etc section main contributions computational approach rest cognitive neuroscience areas science computational models theoretical obvious involved advantages theory theory cognitive neuroscience empirical approaches empirical researchers found common sense theoretical approaches simple box process models general computational modeling field things weve tried text demonstrate specialized theoretical computational approach necessary complexity brain behavior show approach actually practical general theories important sense data here examples ways computational models sense data sense means showing fits thing observe things understand basic force memory systems pfc deal reward executive control complicated models help principles division labor learning interactivity etc number constructs cognitive neuroscience psychology single isolated viewed terms brain true things attention memory working memory neural network models help field whole constructs mechanistic constructs fit better underlying biology example saw attention emerge interaction brain areas inhibitory activation constraints overall process multiple constraint satisfaction operating inhibitory constraints representations situation notion attention proposed non network neural perspective idea neural network perspective considerable support memory good example constructs priming semantic priming considered separate viewing mechanistic computational basis see important distinctions types priming weight versus activation based see priming typically typically identified term long effects corresponding based weight mechanism whereas semantic priming typically identified transient based activation effects based activation based weight semantic priming recently demonstrated case semantic priming researchers working neural network perspective need introduce new box new finding disengage inhibition working memory active maintenance top activation attention executive control etc difficult see gradual development spatially invariant representations work implemented model sentence model complex emergent representations capture lots things semantics via word occurrence