{"title": "A Computational Model of Prefrontal Cortex Function", "book": "Advances in Neural Information Processing Systems", "page_first": 141, "page_last": 148, "abstract": null, "full_text": "A  Computational Model of Prefrontal \n\nCortex Function \n\nTodd S.  Braver \nDept.  of Psychology \nCarnegie  Mellon  Univ. \nPittsburgh,  PA  15213 \n\nJonathan D.  Cohen \nDept.  of Psychology \nCarnegie  Mellon  Univ . \nPittsburgh ,  PA  15213 \n\nDavid Servan-Schreiber \n\nDept.  of Psychiatry \nUniv .  of Pittsburgh \nPittsburgh, PA  15232 \n\nAbstract \n\nAccumulating  data  from  neurophysiology  and  neuropsychology \nhave  suggested  two  information processing roles  for  prefrontal cor(cid:173)\ntex  (PFC):  1)  short-term  active  memory;  and  2)  inhibition.  We \npresent  a  new  behavioral  task  and  a  computational model  which \nwere  developed  in  parallel.  The  task  was  developed  to probe  both \nof these  prefrontal  functions  simultaneously,  and  produces  a  rich \nset  of behavioral  data  that  act  as  constraints  on  the  model.  The \nmodel is  implemented in continuous-time, thus providing a  natural \nframework  in  which  to study  the  temporal dynamics of processing \nin the task.  We show how the model can be used to examine the be(cid:173)\nhavioral  consequences of neuromodulation in PFC . Specifically, we \nuse  the model to make novel and testable predictions regarding the \nbehavioral performance of schizophrenics,  who  are hypothesized  to \nsuffer  from  reduced  dopaminergic tone  in  this brain  area. \n\n1 \n\nIntroduction \n\nPrefrontal  cortex  (PFC)  is  an  area  of the  human  brain  which  is  significantly  ex(cid:173)\npanded  relative  to other  animals.  There  is  general  consensus  that  the  PFC is  cen(cid:173)\ntrally involved  in  higher cognitive  activities such  as  planning, problem solving  and \nlanguage.  Recently, the PFC has been associated with two specific  information pro(cid:173)\ncessing  mechanisms:  short-term  active  memory and  inhibition .  Active  memory is \nthe capacity of the nervous system to maintain information in the form of sustained \nactivation  states  (e.g. ,  cell  firing)  for  short  periods  of time.  This  can  be  distin(cid:173)\nguished  from  forms  of memory  that  are  longer  in  duration  and  are  instantiated  as \n\n\f142 \n\nTodd S.  Braver,  Jonathan  D.  Cohen,  David Servan-Schreiber \n\nmodified values of physiological parameters (e.g., synaptic strength).  Over the last \ntwo  decades,  there have been  a  large number of neurophysiological studies focusing \non  the  cellular basis of active  memory in  prefrontal  cortex.  These studies have re(cid:173)\nvealed neurons in PFC that fire selectively to specific stimuli and response patterns, \nand that  remain active  during a  delay  between  these.  Investigators such  as  Fuster \n(1989)  and  Goldman-Rakic (1987) have argued from this data that PFC maintains \ntemporary information needed to guide behavioral responses through sustained pat(cid:173)\nterns of neural activity.  This hypothesis is  consistent  with behavioral findings from \nboth animal and human lesion studies,  which suggest that PFC is required for tasks \ninvolving delayed responses  to  prior stimuli (Fuster,  1989; Stuss &  Benson,  1986). \nIn  addition  to  its  role  in  active  memory,  many  investigators  have  focused  on  the \ninhibitory functions  of PFC.  It has  been  argued  that  PFC representations  are  re(cid:173)\nquired  to  overcome  reflexive  or  previously  reinforced  response  tendencies  in  order \nto mediate a  contextually appropriate - but otherwise  weaker - response  (Cohen & \nServan-Schreiber,  1992).  Clinically, it has been observed that lesions to PFC are of(cid:173)\nten associated with a syndrome of behavioral disinhibition, in  which  patients act in \nimpulsive and often socially inappropriate ways (Stuss &  Benson,  1986).  This syn(cid:173)\ndrome has often been cited  as evidence that PFC plays an important role inhibiting \nbehaviors  which  are  compelling but socially inappropriate. \nWhile  the  involvement  of PFC  in  both  active  memory  and  inhibition  is  generally \nagreed  upon,  computational models can  play  an important role in providing mech(cid:173)\nanisms by  which  to  explain  how  these  two  information processing  functions  arise. \nThere  are  several  computational models now  in  the  literature  which  have  focused \non  either  the  active memory (Zipser,  1991),  or inhibitory  (Levine &  Pruiett,  1989) \nfunctions  of  PFC,  or  both  functions  together  (Dehaene  &  Changeux,  1989;  Co(cid:173)\nhen &  Servan-Schreiber,  1992).  These models have been instrumental in explaining \nthe role of PFC in a  variety of behavioral tasks  (e.g.,  the Wisconsin Card Sort  and \nStroop).  However,  these  earlier  models are  limited by  their  inability to fully  cap(cid:173)\nture the dynamical processes underlying active memory and inhibition.  Specifically, \nnone  of the simulations have  been  tightly  constrained  by  the  temporal parameters \nfound  in the behavioral tasks (e.g., durations of stimuli, delay periods,  and response \nlatencies).  This limitation is  not found solely in  the models,  but is  also  a feature  of \nthe behavioral tasks  themselves.  The tasks  simulated were  not  structured  in  ways \nthat could facilitate a  dynamical analysis of processing. \nIn  this paper  we  address  the limitations of the previous  work  by  describing  both a \nnew behavioral task and a computational model of PFC. These have been developed \nin  parallel  and,  together,  provide  a  useful  framework  for  exploring  the  temporal \ndynamics of active memory and inhibition and their consequences for  behavior.  We \nthen go on to describe how this framework can be used  to examine neuromodulatory \neffects  in PFC,  which  are believed  to playa critical role in both normal functioning \nand  in psychiatric disorders,  such  as schizophrenia. \n\n2  Behavioral Assessment of Human PFC  Function \n\nWe  have developed  a  task paradigm which  incorporates  two  components central  to \nthe function  of prefrontal  cortex  - short-term  active memory and  inhibition - and \nthat can be used  to study  the  dynamics of processing.  The task is  a  variant of the \ncontinuous performance  test  (CPT),  which  is  commonly used  to study  attention in \n\n\fA  Computational Model of Prefrontal Cortex Function \n\n143 \n\nbehavioral  and  clinical  research.  In  a standard  version  of the  task  (the  CPT-AX), \nletters are presented one at a  time in  the middle of a computer screen.  Subjects  are \ninstructed to press  the target button to the letter X (probe stimulus) but only when \nit is  preceded  by an A (the cue stimulus).  In previous versions of the CPT, subjects \nonly  responded  on  target  trials.  In  the  present  version  of the  task,  a  two  response \nforced-choice  procedure  is  employed; on  non-A-X  trials  subjects  are  asked  to press \nthe non-target button.  This procedure allows for response  latencies  to be evaluated \non every  trial, thus  providing more information about  the  temporal dimensions of \nprocessing  in the task. \nTwo  additional  modifications  were  made  to  the  standard  paradigm  in  order  to \nmaximally engage  PFC  activity.  The  memory function  of PFC  is  tapped  by  ma(cid:173)\nnipulating the  delay  between  stimuli.  In  the  CPT-AX,  the  prior stimulus  (cue  or \nnon-cue)  provides  the context  necessary  to decide how  to respond  to the probe let(cid:173)\nter.  However,  with  a  short  delay  (750  msec.),  there  is  little  demand  on  memory \nfor  the  prior  stimulus.  This  is  supported  by  evidence  that  PFC  lesions  have  been \nshown  to  have  no  effect  on  performance  when  there  is  only a  short  delay  (Stuss  & \nBenson,  1986).  With a  longer delay  (5000 msec.), however,  it becomes necessary  to \nmaintain a representation of the prior stimulus in order for  it to be used  as context \nfor  responding  to  the  current  one.  The  ability  of the  PFC  to  sustain  contextual \nrepresentations over the delay period can be determined behaviorally by comparing \nperformance on short  delay  trials  (50%)  against  those  with long delays  (50%). \nThe  inhibitory  function  of  PFC  is  probed  by  introducing  a  prepotent  response \ntendency  that must be overcome to respond  correctly.  This tendency  is  introduced \ninto the  task  by  increasing  the frequency  of target  trials  (A  followed  by  X).  In  the \nremaining trials,  there  are  three  types  of distractors:  1)  a  cue  followed  by  a  non(cid:173)\ntarget probe letter (e.g. , A-Y); 2)  a non-cue followed  by the target probe letter (e.g., \nB-X);  and a  non-cue followed  by a  non-target probe letter  (e.g.,  B-Y).  Target trials \noccur  70%  of the  time,  while  each  type  of distract or  trial  occurs  only  10%  of the \ntime.  The frequency  of targets  promotes  the  development  of a strong  tendency  to \nrespond  to the  target  probe letter  whenever  it occurs ,  regardless  of the identity  of \nthe cue  (since  a  response  to the  X  itself is  correct  7 out of 8 times). \nThe ability to inhibit this  response  tendency  can  be examined by  comparing accu(cid:173)\nracy  on  trials  when  the  target  occurs  in  the  absence  of the  cue  (B-X  trials) ,  with \nthose made when  neither the cue  nor target occurs (i.e.,  B-Y trials, which provide a \nmeasure of non-specific  response  bias and random responding).  Trials in  which  the \ncue  but  not  the  target  probe  appears  (A-Y  trials)  are  also  particularly interesting \nwith  respect  to  PFC  function.  These  trials  measure  the  cumulative  influence  of \nactive  representations  of context  in  guiding  responses.  In  a  normally functioning \nsystem,  context  representations  should  stabilize  and  increase  in  strength  as  time \nprogresses.  Thus,  it  is  expected  that  A- Y  accuracy  will  tend  to  decrease  for  long \ndelay  trials relative  to short  ones . \nAs  mentioned  above,  the  primary  benefit  of  this  paradigm  is  that  it  provides  a \nframework  in  which  to simultaneously probe the inhibitory and  memory functions \nassociated  with  PFC.  This  is  supported  by  preliminary  neuroimaging  data  from \nour  laboratory  (using  PET)  which  suggests  that  PFC  is,  in fact,  activated  during \nperformance of the  task.  Although  it is  simple in structure,  the task also generates \na  rich  set  of behavioral  data.  There  are four  stimulus  conditions  crossed  with  two \ndelay  conditions  for  which  both  accuracy  and  reaction  time  performance  can  be \n\n\f144 \n\nTodd S.  Braver, Jonathan  D. Cohen,  David Servan-Schreiber \n\n100 \n\n90 \n\n80 \n\n70 \n\n60 \n\n750 \n\n650 \n\n! 550 \n\" .Ii 1 450 \n\n..: \n\n350 \n\n250 \n\nAccurac, (Short Delay) \n\nAccurac, (Long Delay) \n\nV  V \n\n1- MODEL  (Ace) J \n\n.DATA  (Ace) \n\nRT(ShortDelay) \n\nRT (Long Delay) \n\nI  09 ~ '''J!.'  -~~-~ \n\n, \n-~~ \n\n, \n, \n\nAX  AY \n\nBX \n\nBY \n\nTrial Condition \n\nAX  AY \n\nBX \n\nBY \n\nTrial Condition \n\nMODEL (Correct) \n\n--- MODEL (Incorrect) \n\n.. DATA  (Correct) \n'V DATA  (Incorrect) \n\nFigure  1:  Subjecl  beha.viora.1  da.la.  with  model  performa.nce  s uperimposed .  Top  Panels:  Acc ura.cy  a.c ross \nboth  dela.ys  in  a.1I  four  condilion s.  Bottom  Panels:  Rea.ction  times  for  both  correc t  a.nd  incorrec t  res pon se s  in \n\na.1I  conditions .  Ba.rs  repre sent  sta.nda.rd  error  of  mea.s ure ment for  the  empirica.l  da.ta.. \n\nmeasured.  Figure  1 shows  data gathered  from  36  college-age  subjects  performing \nthis  task. \nIn  brief,  we  found  that:  1)  Accuracy  was  relatively  unchanged  in  the  long  delays \ncompared to the short, demonstrating that active memory was  adequately support(cid:173)\ning  performance;  2)  A-Y  accuracy,  however,  did  slightly  decrease  at  long  delays, \nreflecting  the  normal  build-up  of context  representations  over  time;  3)  Accuracy \non  B-X  trials  was  relatively  high,  supporting  the  assumption  that  subjects  could \neffectively  use  context representations  to inhibit prepotent  responses; 4)  A  distinct \npattern  emerged  in  the  latencies  of correct  and  incorrect  responses ,  providing  in(cid:173)\nformation on  the temporal dynamics of processing  (i .e. , responses  to A-Y  trials are \nslow  on correct  trials and fast  on incorrect  ones; the pattern is reversed for  B-X  tri(cid:173)\nals) .  Taken together,  the data provides specific,  detailed  information about normal \nPFC functioning,  which  act  as  constraints on the  development  and  evaluation of a \ncomputational model. \n\n3  A  Computational Model of the CPT-AX \n\nWe  have developed  a recurrent network  model which  produces  detailed information \nregarding  the  temporal course  of processing  in  the  CPT-AX  task.  The  network  is \ncomposed  of three  modules:  an  input  module, a  memory  module,  and  an  output \nmodule.  The  memory  module  implements  the  memory  and  inhibitory  functions \nbelieved  to be carried  out by  PFC. Figure 2 shows  a  diagram of the  model. \nEach  unit in the input module represents  a  different stimulus condition: A,  B,  X & \n\n\fA  Computational Model of Prefrontal Cortex Function \n\n145 \n\nOUTPUT LAYER \n\n~~L0~ \n\nINPUT LAYER \n\nFigure  2:  A  diagram  of  the  CPT\u00b7AX  model. \n\nY.  Units  in the input module make excitatory  connections  on the response  module, \nboth directly  and indirectly through the memory module.  Lateral inhibition within \neach  layer produces competition for  representations.  Activity from the cue stimulus \nflows  to  the  memory  module,  which  is  responsible  for  maintaining a  trace  of the \nrelevant  context  in  each  trial.  Units  in  the  memory  module  have  self-excitatory \nconnections,  which  allow  for  the  activity  generated  by  the  cue  to  be  sustained  in \nthe absence of input.  The recurrent connectivity utilized by each unit in this module \nis  assumed  to  be  a  simpler,  but formally  equivalent  analogue  of a  fully  connected \nrecurrent  cell  assembly.  Further,  Zipser  (1991)  has used  this type of connectivity to \nproduce  temporal  activity  patterns  which  are  highly  similar to  the  firing  patterns \nof neurons  in  memory-associated  areas  of cortex,  such  as  PFC.  Activity  from  the \ninput and memory modules is  integrated  in  the output  module.  The output of this \nmodule determines  whether  a  target  (T)  or non-target  (N)  response  is  made. \nTo simulate the CPT-AX  task  we  have  purposefully  kept  the  network  architecture \nand size  as simple as possible in order  to maximize the model's interpretability.  We \nhave therefore not attempted to simulate neural information processing in a neuron(cid:173)\nby-neuron manner.  Rather, the populations of a few  units are seen  as capturing the \ninformation  processing  characteristics  of much  larger  populations  of real  neurons. \nIn  this  way,  it  is  possible  to  capture  the  stochastic,  distributed,  and  dynamical \nproperties of real neural networks  with small and  analytically tractable simulations. \nThe simulation is  run in  a  temporally continuous framework in  which  processing is \ngoverned  by  the following  difference  equation: \n\nwhere \n\n1 \n\n(1 ) \n\n(2) \n\nis  the  state  of unit j,  Ij  is  the  total input  to j , dt  is  the  time-step  of integration,  'Y \nis  the gain  and  f3  is  the  bias.  The continuous framework  is  preferable  to  a  discrete \nevent-based  one  in  that  it  allows  for  a  plausible  way  to scale  events  appropriately \nto  the  exact  temporal  specifications  of the  task  (i.e.,  the  duration  of stimuli  and \nthe  delay  between  cue  and  probe).  In  addition,  the  continuous  character  of the \nsimulation naturally  provides  a  framework  for  inferring  the  reaction  times  in  the \nvarious  conditions. \n\n\f146 \n\nTodd S.  Braver,  Jonathan  D. Cohen,  David Servan-Schreiber \n\n4  Simulations of Behavioral Performance \n\nWe  used  a  continuous  recurrent  generalization  of backpropagation  (Pearlmutter, \n1989)  to  train  the  network  to perform  the  CPT-AX.  All  of the  connection  weights \nwere  developed entirely by the  training procedure,  with the constraint that that all \nself and between layer weights were forced  to be positive and all within layer weights \nwere  forced  to  be  negative.  Training consisted  of repeated  presentation  of each  of \nthe 8 conditions in the task (A-X,A-Y,B-X,B-Y, at both long and short delays), with \nthe  presentation frequency  of each  condition  matching that of the  behavioral task . \nWeights were  updated  after  the  presentation  of each  trial, biases  ({3)  were  fixed  at \n-2.5,  and dt  was  set at  0.1.  The network  was  trained  deterministically; completion \nof training occurred  when  network  accuracy  reached  100% for  each  condition. \nFollowing training,  weights  were fixed.  Errors  and reaction  time distributions  were \nthen  simulated by  adding  zero-mean  Gaussian  noise  to  the  net  input  of each  unit \nat  every  time step  during  trial  presentation.  A  trial  consisted  of the  presentation \nof the  cue  stimulus,  a  delay  period  and  then  the  probe  stimulus.  As  mentioned \nabove, the duration of these events was  appropriately scaled to match the temporal \nparameters  of the  task  (e.g.,  300  msec.  duration  for  cue  and  probe  presentation, \n750  msec.  for  short  delays,  5000  msec.  for  long  delays).  A  time constant  (1\")  of 50 \nmsec.  was used for simulation in the network.  This scaling factor provided sufficient \ntemporal resolution  to  capture  the  relationship  between  the  two  task  delays  while \nstill permitting a  tractable way  of simulating the events . \nResponses  were  determined  by  noting which  output unit  reached  a  threshold  value \nfirst following presentation of the probe stimulus.  Response latency  was determined \nby  calculating  the  number  of  time  steps  taken  by  the  model  to  reach  threshold \nmultiplied by  the time constant 1\".  To facilitate comparisons  with the experimental \nreaction times, a constant k was added to all values produced .  This parameter might \ncorrespond  to  the  time required  to  execute  a  motor response.  The  value  of k  was \ndetermined  by  a  least  mean squares  fit  to  the  data.  1000  trials  of each  condition \nwere  run  in  order  to  obtain  a  reliable  estimate  of  performance  under  stochastic \nconditions.  The standard  deviation of the  noise  distribution (0')  and  the  threshold \n(T) of the response  units  were  adjusted  to produce  the  best fit  to the subject  data. \nFigure  1 compares the  results  of the simulation against  the behavioral data. \nAs  can  be  seen  in  the figure,  the  model provides  a  good fit  to  the  behavioral  data \nin  both  the  pattern  of accuracy  and  reaction  times.  The  model  not  only  matches \nthe qualitative pattern of errors  and reaction times but produces very similar quan(cid:173)\ntitative results  as  well.  The match between  model and experimental results  is  par(cid:173)\nticularly striking when  it is  considered  that there  are  a  total of 24  data points that \nthis  model is  fitting,  with  only 4 free  parameters  (O',T,1\",k).  The model's ability to \nsuccessfully  account for  the pattern of behavioral performance  provides  convincing \nevidence  that it  captures  the essential  principles  of processing  in  the  task.  We  can \nthen feel  confident  in not  only  examining normal processing,  but  also  in  extending \nthe model to  explore  the effects  of specific  disturbances  to processing  in  PFC . \n\n5  Behavioral Effects of Neuromodulation in  PFC \n\nIn a previous meeting of this conference a simulation of a simpler version of the CPT \nwas  discussed  (Servan-Schreiber,  Printz,  &  Cohen,  1990).  In  this  simulation  the \n\n\fA  Computational Model of Prefrontal Cortex Function \n\n147 \n\nAccuracy (Short Delay) \n\nAccuracy (Long Delay) \n\n.... CJ \n~ ... ... Q \nU .... \n== \nCJ ... ~ =-\n\n~ \n\n100 \n\n90 \n\n80 \n\n70 \n\n60 \n\nI , , \n\nI \nI \nI \nI \n, \nI \n, \nI \n, \n~ \n\nAX \n\nAY \n\nBX \n\nBY \n\nAX  AY \n\nBX \n\nBY \n\n-\nMODEL  (Normal Gain) \n-- - MODEL  (Reduced Gain) \n\n.DATA \n\n(Controls) \n\nFigure 3:  Comparision  of of model  performance with  normal and  redu ced  gain .  The graph illustrates ~he effec~ \nof  reducing  gain  in  the  memory  layer  on  task  performance.  In  the  baseline  network  \"1=1 ,  in  ~he reduced-gain \nnetwork  \"1=0.8. \n\neffects of system-wide changes in catecholaminergic tone were  captured by  changing \nthe gain (-r)  parameter of network units.  Changes in gain are thought correspond  to \nthe action of modulatory neurotransmitters in modifying the responsivity of neurons \nto  input signals  (Servan-Schreiber  et  aI. , 1990;  Cohen &  Servan-Schreiber,  1992). \nThe  current  simulation  of the  CPT  offers  the  opportunity  to  explore  the  effects \nof neuromodulation on  the  information processing  functions  specific  to  PFC.  The \ntransmitter  dopamine  is  known  to  modulate  activity  in  PFC,  and  manipulations \nto  prefrontal  dopamine  have  been  shown  to  have  effects  on  both  memory-related \nneuronal activity and behavioral performance (Sawaguchi &  Goldman-Rakic, 1991). \nFurthermore,  it has  been  hypothesized  that  reductions  of the neuromodulatory ef(cid:173)\nfects  of  dopamine  in  PFC  are  responsible  for  some  of the  information processing \ndeficits  seen  in  schizophrenia.  To simulate the  behavior of schizophrenic  subjects, \nwe  therefore  reduce  the gain ('Y)  of units in  the  memory module of the  network. \nWith reduced  gain in the memory module, there  are striking changes in the model's \nperformance of the  task.  As  can be seen  in  Figure  3,  in  the short  delay  conditions \nthe  performance  of the  reduced-gain  model  is  relatively  similar to  that  of control \nsubjects  (and  the  intact  model).  However,  at long  delays ,  the  reduced-gain  model \nproduces  a  qualitatively  different  pattern  of performance.  In  this  condition,  the \nmodel has a high B-X error rate but a low A-Y error rate, a pattern which is opposite \nto  that  seen  in the  control  subjects.  This  double  dissociation  in  performance  is  a \nrobust  effect  of the  reduced-gain  simulation  (i.e. ,  it  seems  relatively  uninfluenced \nby  other parameter adjustments) . \nThus,  the  model  makes  clear-cut  predictions  which  are  both  novel  and  highly \ntestable.  Specifically,  the  model  predicts  that:  1)  Differences  in  performance  be-\n\n\f148 \n\nTodd S.  Braver,  lonatMn D.  Cohen,  David Servan-Schreiber \n\ntween  control  and  schizophrenic  subjects  will  be  most  apparent  at  long  delays ;  2) \nSchizophrenics  will  perform significantly  worse  than  control subjects  on  B-X  trials \nat long  delays;  3)  Schizophrenics  will  perform significantly  better than  control sub(cid:173)\njects on A-Y  trials at long delays.  This last prediction is especially interesting given \nthe  fact  that  tasks  in  which  schizophrenics  show  superior  performance  relative  to \ncontrols  are  relatively  rare  in experimental research. \nFurthermore, the model not only makes predictions regarding schizophrenic behav(cid:173)\nioral performance, but also offers  explanations as  to their mechanisms.  Analyses of \nthe  trajectories  of activation states  in  the  memory module reveals  that  both of the \ndissociations  in  performance  are  due  to  failures  in  maintaining representations  of \nthe context set  up  by  the cue stimulus.  Reducing gain in the memory module blurs \nthe distinction  between  signal and noise , and causes  the  context  representations  to \ndecay  over  time.  As  a  result,  in  the  long  delay  trials ,  there  is  a  higher  probability \nthat the model will  show  both failures  of inhibition (more B-X  errors)  and memory \n(less  A- Y  errors) . \n\n6  Conclusions \n\nThe results of this paper show how a computational analysis of the temporal dynam(cid:173)\nics  of PFC information processing  can  aid  in  understanding  both  normal  and  dis(cid:173)\nturbed behavior.  We  have developed  a behavioral task which simultaneously probes \nboth the inhibitory and active memory functions of PFC. We have used  this task in \ncombination with  a  computational model to explore  the effects  of neuromodulatory \ndysfunction, making specific predictions regarding schizophrenic performance in the \nCPT-AX.  Confirmation of these  predictions  now  await further  testing. \n\nReferences \nCohen,  J.  &  Servan-Schreiber,  D.  (1992).  Context , cortex, and dopamine:  A  connectionist \n\napproach  to  behavior  and  biology  in  schizophrenia.  Psychological Review,  99 , 45- 77. \n\nDehaene,  S.  &  Changeux,  J.  (1989).  A  simple  model  of  prefrontal  cortex  function  III \n\ndelayed-response  tasks.  Journal of Cognitive  Neuroscience,  1 (3),  244- 261. \n\nFuster,  J.  (1989).  The  prefrontal cortex.  New  York:  Raven  Press. \nGoldman-Rakic,  P.  (1987).  Circuitry of primate prefrontal cortex and  regulation  of behav(cid:173)\n\nior  by  representational  memory.  In  F.  Plum  (Ed .),  Handbook of physiology-the nervous \nsystem,  v.  Bethesda,  MD:  American  Physiological  Society,  373-417. \n\nLevine,  D.  &  Pruiett,  P.  (1989).  Modeling  some  effects  of frontal  lobe  damage:  novelty \n\nand  perseveration.  Neural  Networks, 2 ,  103-116. \n\nPearlmutter,  B.  (1989).  Learning  state  space  trajectories  in  recurrent  neural  networks. \n\nNeural  Computation,  1 , 263-269. \n\nSawaguchi,  T.  &  Goldman-Rakic,  P.  (1991).  D1  dopamine  receptors in  prefrontal  cortex: \n\nInvolvement  in  working  memory.  Science, 251 , 947-950. \n\nServan-Schreiber,  D.,  Printz,  H.,  &  Cohen,  J.  (1990).  The  effect  of  catecholamines  on \nperformance:  From unit  to system behavior.  In  D.  Touretzky  (Ed.),  Neural information \nprocessing systems 2. San  Mateo,  GA:  Morgan  Kaufman ,  100-108. \n\nStuss,  D.  &  Benson ,  D.  (1986) .  The  frontal  lobes.  New  York:  Raven  Press. \nZipser,  D.  (1991).  Recurrent  network model  of the neural  mechanism of short-term active \n\nmemory.  Neural  Computation,  3,179- 19.3. \n\n\f", "award": [], "sourceid": 1020, "authors": [{"given_name": "Todd", "family_name": "Braver", "institution": null}, {"given_name": "Jonathan", "family_name": "Cohen", "institution": null}, {"given_name": "David", "family_name": "Servan-Schreiber", "institution": null}]}