Combinatorial structures and processing in Neural Blackboard Architectures

Transcription

1 Combiaorial srucures ad processig i Neural Blackboard Archiecures Frak a der Velde CTIT, CPE-BMS Uiersiy of Twee, Eschede IOP, Leide The Neherlads f.aderelde@uwee.l Marc de Kamps Isiue for Arificial Ielligece ad Biological Sysems Uiersiy of Leeds Leeds, Uied Kigdom M.deKamps@leeds.ac.uk Absrac We discuss ad illusrae Neural Blackboard Archiecures (NBAs) as he basis for ariable bidig ad combiaorial processig he brai. We focus o he NBA for seece srucure. NBAs are based o he oio ha cocepual represeaios are i siu, hece cao be copied or raspored. Noel combiaorial srucures ca be formed wih hese represeaios by embeddig hem i NBAs. We discuss ad illusrae he mai characerisics of his form of combiaorial processig. We also illusrae he NBA for seece srucures by simulaig eural aciiy as foud i recely repored iracraial brai obseraios. Furhermore, we will show how he NBA ca accou for ambiguiy resoluio ad garde pah effecs i seece processig. 1 Iroducio We oulie ad illusrae he implemeaio ad processig of combiaorial srucures i Neural Blackboard Archiecures (or NBAs). Firs, we will discuss he ideas o which NBAs are based. The, we will illusrae seece represeaio ad processig i he NBA for seece srucure [1]. Nex, we will simulae eural aciaio paers ha arise i he NBA durig seece processig ad relae i o recely repored iracraial brai obseraios. Fially, we will illusrae ad discuss how he NBA hadles liguisic pheomea such as uproblemaic ambiguiies ad garde pah seeces [2]. 2 I siu cocep represeaios The deelopme of Neural Blackboard Archiecures is based o he oio ha represeaios i he brai, such as represeaios of coceps, are i siu, i lie wih he eural assemblies proposed by Hebb [3]. Figure 1 (based o [4]) illusraes his wih he cocep ca. As Hebb argued, a eural assembly deelops oer ime as euros ha process iformaio abou a cocep become iercoeced. These euros could be locaed i differe pars of he brai, depedig o he aure of he uderlyig cocep. So, i case of a cocep like ca, hey will cosis of euros i he isual ad audiory corex, ioled i he percepio of cas. They could also cosis of euros ioled i acios relaed o cas (e.g., prooucig he word ca). Cocep represeaios wih eural assemblies could be disribued, bu pars of he assembly could also cosis of more local represeaios. They could cosis of euros ioled i processig iformaio abou he coceps, bu also of euros ioled i acios relaed o coceps. This eails ha boh icomig ad ougoig coecios deermie he aure of a cocep represeaio [5]. Furhermore, ulike he origial assemblies as described by Hebb, coceps as ieded here 1 Copyrigh 2015 for his paper by is auhors. Copyig permied for priae ad academic purposes.

2 pe is paw has Seece blackboard acio emoio ca Phoology blackboard Percepio blackboards Oher blackboards Figure 1: I siu cocep represeaio (based o [4]). are o oly associaie eural srucures. They also icorporae relaios, as illusraed wih he relaios is pe ad has paw i Figure 1. The assembly or web-like srucure of a cocep represeaio eails ha coceps represeaios are i siu [4]. Tha is, whereer a cocep is aciaed i always cosiss of he aciaio of he assembly of ha cocep or a par of i. I his iew, i is o possible o make a copy of a cocep represeaio ad sore i elsewhere i he brai o, for example, creae a combiaorial srucure wih i or use i i a reasoig process. I his sese, i siu cocep represeaios are fudameally differe from symbols i symbol archiecures [6]. Due o he i siu aure of coceps represeaios, hey are also grouded i percepio or acio (depedig o he aure of he cocep). The i siu aure of euroal cocep represeaios imposes cosrais o he way hey ca be combied i higher-leel cogiie combiaorial srucures, such as seeces. I siu cocep represeaio cao be copied or raspored o cosruc combiaorial srucures. Isead, as illusraed i Figure 1, hey are embedded i eural blackboards o form combiaorial srucures. We assume ha here will be specific eural blackboards for specific forms of combiaorial srucures, ad cocep represeaios will be embedded i a umber of hem depedig o he aure of ha cocep ad he role i plays i specific combiaorial srucures. For example, he cocep ca will be embedded i a phoology blackboard, formig he word ca based o familiar phoemes. I will also be embedded i a seece blackboard, eeded o form seece srucures. I could also be embedded i a percepio blackboard uderlyig he combiaorial aure of isual cogiio [7]. Oher blackboards, e.g., eeded for sequeial processig or reasoig, could be icluded as well. The eural blackboards allow he cosrucio of (poeially oel) combiaorial srucures based o (familiar) i siu cocep represeaios, usig forms of ariable bidig (depedig o he specific blackboard). They aim o saisfy he followig characerisics: The cocep represeaios remai i siu (ad hus grouded) ee whe hey are a par of (oel) combiaorial srucures. The combiaorial srucures or pars hereof are coe addressable, ee wih oel combiaorial srucures (of familiar cosiues). The ariable bidig of coceps i blackboards creaes he coecio pahs eeded o produce behaior (likig sesory ad moor aciaio), ee wih (oel) combiaorial srucures. We will discuss ad illusrae hese characerisics usig he seece blackboard. 3 Seece eural blackboard Figure 2 illusraes he srucure of he seece Soe (he) dog i he Neural Blackboard Archiecure for seece srucure [1]. Here, we assume ha he words ad dog are familiar 2

3 Seece blackboard Nx Sy Vz Srucure assemblies Codiioal coecios: W1 Nu /so/ // /e/ Phoology blackboard dog Word assemblies Figure 2: Seece represeaio i a seece eural blackboard. cosiues bu Soe is a ew ame, o heard before. The represeaios of ad dog cosis of eural assembly srucures ( word assemblies ), i lie wih he cocep ca i Figure 1. For simpliciy, we represe hem wih oals, bu he full coecio srucure is always implied. The seece blackboard allows he creaio of a (emporal) coecio srucure ha bids he word assemblies i lie wih he relaios expressed by he seece. To his ed, he NBA cosiss of srucure assemblies such as N x ad N u i Figure 2 (represeig ous) ad S y (seece) ad V z (erb). The word assemblies ca bid o srucure assemblies of heir kid, ad srucure assemblies ca bid o each oher i lie wih he seece srucure. The bidig process resuls from specific codiioal coecios (oulied below). As oed i [8], humas hae he abiliy o iclude ew words ad ames i (ee oel) seeces srucures, exemplified here wih he ame Soe. Howeer, he abiliy o iclude oel words i seeces is resriced o ew words based o familiar phoemes. I he NBA his abiliy resuls from a phoology blackboard i which ew ames ca be formed based o familiar phoemes [9]. The phoology blackboard ieracs wih he seece blackboard, which allows he icorporaio of he ew ame i a seece srucure [10]. Here, Soe aciaes a se of phoemes (illusraed wih he pseudo phoemes so,, ad e), which bid o a srucure assembly i he phoology blackboard (W1). I ur, his srucure assembly bids o he seece srucure i he seece blackboard. Codiioal coecios X Corol circui i di WM circui ca Srucure assemblies Word assemblies dog Srucure assemblies: = Mai assemblies = Subassemblies Figure 3: Bidig i he seece eural blackboard. The bidig process i he NBA is illusraed i more deail i Figure 3 wih he seece ca dog. The codiioal coecios, here he red ad black hick lies, cosis of gaig circuis. I he NBA, each word assembly (e.g., of a ou) is coeced o a se of srucure assemblies of he same kid (all N i assemblies i he case of a ou) wih gaig circuis. (The umber of hese coecios ca be reduced by represeig words i he phoology blackboard firs, see [9].) I ur, each srucure assembly cosiss of a mai assembly, such as, ad (a se of) subassemblies, such as or. The coecio bewee a mai assembly ad a subassembly cosiss of a gaig circui as well. 3

4 Srucure assemblies of differe kids, such as ad, are coeced by heir subassemblies of he same kid. Here, heir (heme) subassemblies, which represes he fac ha a erb ca hae a heme (objec). This coecio also cosiss of a gaig circui. The gaig circuis operae by disihibiio (di), illusraed i Figure 3. So, whe is acie, i aciaes a euro (or euro populaio) X ad a ihibiory euro (or populaio) i. The laer ihibis X, which blocks he flow of aciaio. Bu whe i iself is ihibied (by euro or populaio di) aciaio ca flow from (ia X) o. Gaig circuis ca be disihibied (or aciaed ) i oe of wo differe ways. I case of gaig circuis bewee mai assemblies ad subassemblies he aciaio resuls from a exeral corol circui ha aciaes he di populaio. This is how syacical operaios affec bidig i he blackboard. A corol circui could hae recogized ha dog represe a erb ad a heme. I he aciaes all di populaios i he gaig circuis bewee all V i ad N j assemblies ad heir assemblies. As a resul, he acie V i ad N j will aciae heir subassembly. Gaig circuis bewee subassemblies ad bewee word ad mai assemblies are aciaed by (specific) workig memory (WM) populaios. A WM populaio remais acie for a while, afer iiial aciaio, by reerberaig aciaio i he populaio [11]. A acie WM populaio bids he assemblies o which i is coeced i he maer oulied below. 3.1 Variable bidig i a coecio pah Figure 4 illusraes how bidig is achieed i he NBA. The pars (a), (b) ad (c) illusrae he same bidig process wih icreasig deail. I (a), he bidig bewee he subassemblies of (or -) ad (-) i Figure 3 is repeaed. Figure 4b illusraes ha his bidig is based o a coecio marix, which cosiss of colums ad rows of coecio odes illusraed i Figure 4c. Yj - Xi X i Y i - WM i di di i Coecio marix ( = ihibiio) Coecio ode (a) (b) (c) Figure 4: Bidig by aciiy i a Coecio Marix (based o [1]). Each specific V i - ad N j - pair of subassemblies is iercoeced i a specific coecio ode, locaed i a coecio marix dedicaed o bidig V i - ad N j - subassemblies. I geeral, whe wo assemblies X i ad Y j (e.g., - ad -) are cocurrely acie i he processig of a seece, hey aciae a WM populaio i heir coecio ode by meas of a gaig circui, as illusraed i Figure 4c. I ur, he acie WM populaio disihibis a gaig circui by which aciaio ca flow from X i o Y j, ad aoher such circui, o show i (c), by which aciaio ca flow from Y j o X i. As log as heir WM populaio is acie, X i ad Y j are boud because aciaio will flow from oe o he oher wheeer oe of hem is (iiially) aciaed. The NBA allows ay ou o bid o ay erb i ay hemaic role usig dedicaed coecio marices. Also, he NBA has srucure assemblies ha ca bid o oher srucure assemblies, such as i Figure 3, or clause srucure assemblies ([1]). I his way, hierarchical seece srucures ca be represeed, such as relaie or compleme clauses. 4

5 A gie word assembly ca bid o differe srucure assemblies a he same ime, allowig he creaio of seece srucures i which words are repeaed (e.g., ca ca). Ye, word assemblies remai is siu i his way, so words i seece srucures are always accessible ad grouded. The role of he NBA is hus o esablish a coecio pah for ay seece srucure ha ca be formed i a laguage. Alhough i has bee quesioed wheher his would be possible [8], a aailable coecio pah is i fac a ecessary codiio for ay model of (euroal) ariable bidig. Tha is, wheeer wo euroal represeaios are boud i a fucioal way, here is some kid of (direc or idirec) coecio pah bewee hose represeaios. This pah is eeded o produce (meaigful) behaior based o he fucioal bidig a had. For example, o aswer quesios abou he relaio expressed by he bidig [10]. The abiliy of he brai o esablish such coecio pahs deries from is small-word like coecio srucure [12]. The NBA ideed has such a srucure [1]. I he NBA he coecio marices are gie. Bu i a pilo simulaio [13] we iesigaed he possible deelopme of such a coecio srucure. The simulaios idicaed ha a srucure like a coecio marix could arise durig deelopme, based o iiially radom aciaio ad coecio paers. Bu, of course, each ode i he coecio marix would also hae o hae he circuis as illusraed i Figure 4c for he bidig process o occur. 3.2 Coe addressable seece srucures The role of he NBA o (emporarily) creae a coecio pah eeded for he producio of behaior is illusraed i Figure 5. This figure illusraes how he NBA ca aswer (bidig) quesios (for a simulaio, see [1]). S2 Bill kows? N4 V2 N3 Bill kows Mary Joh Figure 5: Coe addressable seece srucures ad compeiio i he NBA i aswerig (bidig) quesios. Grey odes are acie. The srucure of he seece Bill kows Joh, sored i he NBA, ca be used o aswer a (bidig) quesio like Wha does Bill kow? (or Bill kows?). The quesio aciaes he assemblies for Bill ad kows ad proides iformaio ha Bill is he subjec of kows. This resuls i he aciaio of he seece srucure Bill kows. The quesio also asks for he heme of he relaio Bill kows, which ca be used o aciae he gaig circuis for he heme () subassemblies. This will iiiae a flow of aciaio from o ad hus o Joh i Figure 5, producig he aswer o he quesio. The NBA ca correcly produce he aswer ee whe, for example, Bill is a cosiue of oher seeces as well, as i Bill Mary. The represeaio of Bill is he same i siu word assembly i boh seeces (which direcly deries from he oio of i siu represeaio). As oed aboe, hey are disiguished i he NBA by he bidig wih differe ou mai assemblies (here, s N3). The quesio Bill kows? aciaes Bill, which i ur aciaes boh ad N3 o which i is boud. Because he quesio eails ha Bill is he subjec, his resuls i he aciaio of boh ad S2. Howeer, he quesio also ideifies kows as he erb ad o. This resuls i he aciaio of which ihibis he aciaio of V2, because mai assemblies of he same kid ihibi each oher i he NBA [1]. I his way, askig for he heme produces he correc aswer Joh (boud o kows) isead of he icorrec aswer Mary (boud o ). 5

6 Figure 5 illusraes he role of a coecio pah i producig behaior. The quesio proides sesory iformaio resulig i he aciaio of Bill kows. The aswer (Joh) is produced by a moor program ha would be aciaed by he i siu word assembly of Joh. If here were o a coecio pah ha would lik he sesory iformaio (Bill kows) o he moor aciaio (Joh) i is difficul o see how his behaior could be produced. Ye, his kid of behaior is possible for ay arbirary bidig of words i (poeially oel) seece srucures or ee ew words i seece srucures, as i Figure 2. I each of hese cases, he behaior of aswerig a quesio (probig for relaio iformaio or bidig) depeds o coecig sesory iformaio o moor aciaio, i a maer ha saisfies he bidig relaios expressed by he relaio a had. The NBA produces his coecio pah due o is small-world like coecio srucure ad is abiliy o bid arbirary relaios by he bidig process oulied aboe. Figure 5 also illusraes he role of coe addressabiliy i he NBA. Due o he i siu aure of word (cocep) represeaio, he quesio Bill kows? direcly aciaes he word assemblies Bill ad kows, Because hey are boud o he seece srucure i he NBA (i.e., because word represeaios remai i siu i ay seece srucure) he aciaio of Bill ad kows direcly aciaes he parial seece srucure Bill kows. I seems ha he srucure of a quesio iself idicaes ha seece srucures are coe addressable. Afer all, he quesio Bill kows? preses he words Bill ad kows, idicaes ha Bill is he subjec of kows ad idicaes ha he respose should be he heme. All of his iformaio is precisely he iformaio eeded o direcly aciae he parial seece srucure of Bill kows (isead of oher parial srucures such as Bill ) ad o iiiae he flow of aciaio o produce he aswer Joh. I is difficul o see how his process could be more direc ad hus faser ha i his way. Gie he eoluioary pressure o fas ad ye meaigful behaior, i seems ha he NBA fis hese cosrais i a opimal way. 4 Seece processig i he NBA Figure 6 illusraes how seece processig ca proceed i he NBA [14]. Here, seece processig resuls from he ieracio bewee a corol ework ad he aciaio i he eural blackboard. Figure 6 illusraes he processig of ca dog preseed i Figure 3. ca ca ca N--S S- V- V- N- dog ca N V V T dog N (a) (b) (c) Figure 6: Seece processig i he NBA (based o [14]). Seece processig is icremeal i he NBA. The ipu o he corol ework cosiss of word iformaio (e.g. Nou, Verb) ad predicios based o he aciaio i he blackboard ha arise durig processig. Based o is ipu, he corol ework has leared o aciae he gaig circuis i he NBA eeded o creae he seece srucure. I (a) ca is recogized as a ou ad he ework aciaes he bidig bewee ad wih heir subassemblies. I also aciaes he subassembly of, which i ur aciaes a predicio ode V, predicig he occurrece of a erb of he mai clause. I (b) he corol ework, based o he combiaio of he predicio ode V ad he occurrece of he erb, aciaes he subassembly of, which resuls i he bidig of S ad V. I also aciaes he subassembly of ad a predicio ode of a heme (T), 6

7 aicipaig a objec of he erb. I (c), he combiaio of he T ode ad he ou dog resuls i he bidig of ad wih heir subassemblies. I [14] we raied a corol ework wih a se of raiig seeces wih differe syacic forms (e.g., subjec, heme, relaie clause, compleme clause). The ework could he corol a subsaially larger se of seeces i which hese syacic forms were combied i differe ways. For example, he corol ework could produce he correc bidig operaios i seeces wih arbirary deep (ceer) embeddigs. Howeer, he bidig process i he NBA sars o break dow wih hese seeces afer oe or wo embeddigs, due o compeiios wih he coecio marices ioled [1, 14]. Hece, he NBA could accou for boh compeece ad performace aspecs of laguage processig. 4.1 Simulaig euroal aciaio durig seece processig Brai aciaio obsered wih fmri idicaes ha euroal aciaio icreases for loger ad more complex seeces [15]. Obseraios wih iracraial elecrodes proided more deailed iformaio [16]. I paricular, whe (sub) phrases i a seece are formed (or merged i liguisic erms) euroal aciiy firs icreases bu he briefly decreases. Figure 7 shows he aciiy i he NBA whe he seece Te sad sude of Bill Gaes will moe from [16] is processed (akig Bill Gaes as a sigle ou). Figure 7 (lef) shows he basic srucure of he seece i he NBA. Figure 7 (righ) shows he sum of all aciiy (Toal) i he NBA, ad he sum aciiies of he mai assemblies (MA), he subassemblies (SA), ad he WM populaios i he coecio marices. All assemblies ad gaig circuis are modeled wih Wilso Cowa populaio dyamics [17]. MA ad SA assemblies hae reerberaig aciiy as well (uil hey are ihibied). Aciaio alues are ormalized o fi i oe plo. sudes AX1 will Toal m p NM1 e a PP1 of Adj1 p sad B-Gaes a moe Te sad sudes of B-G will moe WM MA SA (ms) Figure 7: NBA aciiy i processig Te sad sude of Bill-Gaes will moe. The sum of all aciiy resembles he paers obsered i [16]. For example, afer Te sad sudes (formig a phrase) aciaio icreases bu he drops, as wih oher phrases i he seece. This is i paricular due o he subassemblies, which are ihibied by feedback from heir coecio marix whe a bidig has succeeded. Bu oerall aciaio icreases wih seece legh ad complexiy [15]. Figure 8 illusraes he eural aciiy i he NBA whe i resoles a uproblemaic ambiguiy (UPA) preseed i [2]. Here, i is UPA13a (Bird foud i room died) s UPA13b (Bird foud i room debris). The NBA resoles ambiguiies by a compeiio bewee coflicig WM populaios [18]. The NBA srucure of boh seeces ad he coflics (red dashed lies) are illusraed i he figure (lef). The assumpio is ha UPA13a is he basis ierpreaio. So, foud is he erb of a relaie clause boud o bird. Simulaio shows ha his icremeal seece processig proceeds wihou problems. So, he processig of UPA13b sars i he same way. The figure shows he idiidual WM aciaios obaied wih UPA13b, which reeal he bidigs achieed (he sum of his aciiy would resul i a plo as i Figure 7). Whe bird () is preseed i bids o, idicaed by he black lie - represeig he WM bidig populaio. Afer he iiial aciaio by he word, his populaio remais acie a a susaied leel, so he bidig remais iac durig (ad afer) 7

8 bird c C1 died V2 foud c p N3 debris i PP1 p - C1- C1- PP1- -PP1 - -N3 room Bird foud i room debris (ms) Figure 8: NBA aciiy i processig Bird foud i room debris. he seece processig. The combiaio bird foud is ierpreed as foud () beig he erb of a relaie clause boud o (as i UPA13a). This bidig succeeds, as idicaed by he WM populaios C1- (gree lie) ad C1- (red lie). The bidigs of foud i he room proceed wihou problems, idicaed by he WM populaios -PP1 (red dash) ad PP1- (blue dash). Howeer, he las word debris proides a differe ierpreaio of foud. Isead of he erb of a relaie clause i is he mai erb of he seece. This ierpreaio iiiaes he aciaio of he WM populaio -N3 (gree dash), which represes he bidig of debris as heme of foud. I also iiiaes he aciaio of he WM populaio - (blue lie) which represes he bidig of foud as he mai erb. This bidig is i coflic wih he bidig C1- (red lie). The compeiio bewee he WM populaios resuls i he declie of he aciaio of C1-, which resoles he bidig coflic. Hece, ambiguiy resoluio i he NBA resuls from dyamic compeiios i he archiecure [18]. Aleraiely, UPA13b could be he basis ierpreaio. The foud is he mai erb direcly. I ha case died causes a coflic i UPA13a. Howeer, simulaios show ha his coflic cao be resoled because blocks he bidig of V2 (died) o. Also, cao bid aymore as erb o a relaie clause of bird (), because he aciaios of ad hae bee or are ihibied by oher mai assemblies ( ad V2 respeciely). Hece, his ambiguiy cao be resoled by he NBA. Howeer, his ambiguiy is similar o he classic garde pah seece The horse raced pas de bar fell [19], which idicaes ha he NBA ca accou for boh ambiguiy resoluio ad garde pah effecs [18]. 5 Coclusios Neural Blackboard Archiecures (NBAs) ca accou for ariable bidig ad (oel) combiaorial processig i eural erms. They also saisfy a umber of characerisics of brai represeaio ad processig. Cocepual represeaios are always i siu, also whe hey are a par of a combiaorial srucure. Hece, combiaorial srucures are coe addressable. This allows heir direc aciaio based o ipu iformaio. NBAs also proide he coecio pahs ha are eeded o accou for he producio of behaior. The seece NBA ca accou for obsered paers of euroal aciaio durig seece processig. I ca also hadle liguisics pheomea such as ambiguiy resoluio ad ca accou for garde pah seeces by he dyamic compeiios i he archiecure. Ackowledgmes The work of Frak a der Velde was fuded by he projec CoCreTe. The projec CoCreTe ackowledges he fiacial suppor of he Fuure ad Emergig Techologies (FET) programme wihi he Seeh Framework Programme for Research of he Europea Commissio, uder FET gra umber The research of Marc de Kamps has receied fudig from he Europea Uio Seeh Framework Programme (FP7/ ) uder gra agreeme o (HBP) (Ref: Aricle II.30. of he Gra Agreeme). 8

9 Refereces 1] a der Velde, F. & de Kamps, M. (2006). Neural blackboard archiecures of combiaorial srucures i cogiio. Behaioral ad Brai Scieces, 29, [2] Lewis, R. L. (1993). A Archiecurally-based Theory of Huma Seece Comprehesio. Thesis Caregie Mello Uiersiy, Pisburgh, PA. [3] Hebb, D. O. (1949). The orgaisaio of behaiour. New York: Wiley. [4] a der Velde, F & de Kamps, M. (2011). Composiioal coeciois srucures based o i siu grouded represeaios. Coecio Sciece, 23, [5] a der Velde, F (2015). Commuicaio, coceps ad groudig. Neural eworks, 62, [6] Newell, A. (1990). Uified heories of cogiio. Cambridge, MA: Harard Uiersiy Press. [7] Va der Velde, F. & de Kamps, M. (2001). From kowig wha o kowig where: Modelig objec-based aeio wih feedback disihibiio of aciaio. Joural of Cogiie Neurosciece, 13(4), [8] Feldma, J. (2013). The eural bidig problem(s). Cogiie Neurodyamics, 7, [9] Va der Velde, F. & de Kamps, M. (2006). From eural dyamics o rue combiaorial srucures (reply o commearies). Behaioral ad Brai Scieces, 29, [10] a der Velde, F. & de Kamps, M. (2015). The ecessiy of coecio srucures i eural models of ariable bidig. Cogiie eurodyamics, 9, DOI /s [11] Ami, D. (1995). The Hebbia paradigm reiegraed: Local reerberaios as ieral represeaios. Behaioral ad Brai Scieces, 18, [12] Shaaha, M. (2010). Embodime ad he ier life. Oxford: Oxford Uiersiy Press. [13] a der Velde, F. & de Kamps, M. (2011). Deelopme of a coecio marix for producie grouded cogiio. IEEE Ieraioal Coferece o Deelopme ad Learig (ICDL) (Vol 2), 1-6. DOI: /DEVLRN [14] a der Velde, F & de Kamps, M. (2010). Learig of corol i a eural archiecure of grouded laguage processig. Cogiie Sysems Research, 11, [15] Pallier, C., Deauchelle, A-D. & Dehaee, S. (2011). Corical represeaio of he cosiue srucure of seeces. PNAS, 108, [16] Nelso, M. J., El Karoui, I., Ragaraja, V., Pallier, C., Parizi, J., Cohe, L., Naccache, L. & Dehaee, S. (2014). Cosiue srucure represeaios reealed wih iracraial daa. Sociey for Neurosciece Aual Meeig. Washigo, DC, USA. [17] Wilso, H. R. & Cowa, J. D (1972). Exciaory ad ihibiory ieracios i localized populaios of model euros. Biophysical Joural, 12, [18] a der Velde, F. & de Kamps, M. (2015). Ambiguiy resoluio i a Neural Blackboard Archiecure for seece srucure. I Tarek R. Besold & Kai-Uwe Kühberger (eds.), Proceedigs of he KI 2015 Workshop o Neural-Cogiie Iegraio. Dresde, Germay. [19] Beer, T. G. (1970). The cogiie basis for liguisic srucures. I Hayes, J. R., (ed.) Cogiio ad he Deelopme of Laguage. New York: Wiley. 9