{"title": "Attentional Modulation of Human Pattern Discrimination Psychophysics Reproduced by a Quantitative Model", "book": "Advances in Neural Information Processing Systems", "page_first": 789, "page_last": 795, "abstract": null, "full_text": "Attentional  Modulation of Human Pattern \nDiscrimination Psychophysics Reproduced \n\nby a  Quantitative  Model \n\nLaurent Itti, Jochen Braun,  Dale K.  Lee and Christof Koch \n\nCalifornia Institute  of Technology,  Pasadena,  CA  91125,  U.S.A. \n\n{itti,  achim,  jjwen,  koch}Oklab.caltech.edu \n\nComputation &  Neural  Systems,  MSC  139-74 \n\nAbstract \n\nWe  previously  proposed  a  quantitative model  of early  visual  pro(cid:173)\ncessing  in  primates,  based  on  non-linearly interacting visual filters \nand statistically efficient  decision.  We  now  use  this model to inter(cid:173)\npret  the observed  modulation of a  range of human psychophysical \nthresholds  with  and  without  focal  visual  attention.  Our  model  -\ncalibrated  by  an  automatic fitting  procedure  - simultaneously  re(cid:173)\nproduces  thresholds for  four  classical  pattern discrimination tasks, \nperformed while attention was engaged by another concurrent task. \nOur model then predicts that the seemingly complex improvements \nof certain  thresholds,  which  we  observed  when  attention  was  fully \navailable for  the  discrimination  tasks,  can  best  be  explained  by  a \nstrengthening  of competition among early  visual filters. \n\n1 \n\nINTRODUCTION \n\nWhat happens  when  we  voluntarily focus  our  attention  to  a  restricted  part  of our \nvisual  field?  Focal  attention  is  often  thought  as  a gating  mechanism,  which  selec(cid:173)\ntively  allows  a  certain  spatial  location  and  and  certain  types  of visual  features  to \nreach  higher  visual  processes.  We  here  investigate  the  possibility  that  attention \nmight have a  specific  computational modulatory effect  on  early  visual  processing. \n\nWe  and others  have  observed  that focal  visual attention can  modulate human psy(cid:173)\nchophysical thresholds for  simple pattern discrimination tasks  [7,  8,  5]  When atten(cid:173)\ntion  is  drawn  away  from  a  task,  for  example  by  \"cueing\"  [12]  to  another  location \nof  the  display,  or  by  a  second,  concurrent  task  [1,  7,  8],  an  apparently  complex \npattern of performance degradation is  observed:  For  some  tasks,  attention has  lit(cid:173)\ntle  or  no  effect  on  performance  (e.g.,  detection  of luminance increments),  while for \n\n\f790 \n\nL. ltti, J.  Braun, D.  K.  Lee and C.  Koch \n\nother  tasks,  attention  dramatically improves  performance  (e.g.,  discrimination  of \norientation).  Our specific findings  with dual-task psychophysics are detailed below. \n\nThese observations have been  paralleled by electrophysiological studies of attention. \nIn  the awake macaque, neuronal responses  to attended stimuli can be 20%  to 100% \nhigher than to otherwise  identical unattended stimuli.  This has been demonstrated \nin visual cortical areas VI, V2,  and V4  [16,  11, 10,9] when the animal discriminates \nstimulus orientation, and  in  areas  MT and  MST when  the animal discriminates the \nspeed  of stimulus motion  [17].  Even  spontaneous firing  rates  are 40%  larger  when \nattention  is  directed  at  a  neuron's  receptive  field  [9].  Whether  neuronal  responses \nto attended stimuli are merely enhanced  [17]  or whether they are also more sharply \ntuned  for  certain  stimulus  dimensions  [16]  remains  controversial.  Very  recently, \nfMRI studies  have shown  similar enhancement  (as measured  with  BOLD contrast) \nin  area  VI  of humans,  specifically  at  the  retinotopic  location  where  subjects  had \nbeen  instructed  to focus  their attention to  [2,  14]. \n\nAll of these observations directly address the issue of the  \"top-down\" computational \neffect  of attentional focusing  onto early  visual  processing stages.  This issue  should \nbe distinguished from that of the  \"bottom-up\" control of visual attention  [6],  which \nstudies  which  visual  features  are  likely  to  attract  the  attention  focusing  mecha(cid:173)\nnism (e.g.,  pop-out phenomena and studies of visual search).  Top-down attentional \nmodulation  happens  after  attention  has  been  focused  to  a  location  of the  visual \nfield,  and most probably involves the massive feedback  circuits which  anatomically \nproject from higher cortical  areas  back to early visual  processing  areas. \n\nIn  the  present  study,  we  quantify  the  modulatory  effect  of attention  observed  in \nhuman psychophysics using a  model of early visual processing.  The model is  based \non  non-linearly  interacting  visual  filters  and  statistically  efficient  decision  [4,  5]. \nAlthough  attention could  modulate virtually any visual  processing  stage  (e.g., the \ndecision stage,  which  compares internal responses  from different  stimuli), our basic \nhypothesis here - supported by electrophysiology and fMRI [16,11,10,17,9,2, 14](cid:173)\nis  that this modulation might happen  very early in the visual processing  hierarchy. \nGiven this basic  hypothesis,  we  investigate how  attention should affect  early visual \nprocessing  in order to quantitatively reproduce  the psychophysical  results. \n\n2  PSYCHOPHYSICAL  EXPERIMENTS \n\nCentral task: \n\nWe  measured  attentional modulation \nof  spatial  vision  thresholds  using  a \ndual-task  paradigm  [15,  7]:  At  the \ncenter  of the visual field,  a  letter dis(cid:173)\ncrimination  task  is  presented,  while \na  pattern  discrimination  task  is  si(cid:173)\nmultaneously  presented  at  a  random \nperipheral  location  (4 0  eccentricity). \nThe  central  task  consists  of discrim(cid:173)\ninating  between  five  letters  \"T\"  or \nfour  \"T\"  and  one  \"L\".  It  has  been \nshown  to  efficiently  engage  attention \n[7].  The  peripheral  task  is  chosen \nfrom  a  battery  of a  classical  pattern \ndiscrimination  tasks,  and  is  the  task \nof interest for this study.  Psychophys-\nical  thresholds  are measured for  two distinct conditions:  In the  \"fully attended\" \ncondition,  observers  are  asked  to  devote  their  entire  attention  to  the  peripheral \n\nthreshold measurement \n\n\fQuantitative Modeling of Attentional Modulation \n\n791 \n\ntask,  and  to  ignore  the  central  task  (while  still  fixating  the  center  of the  screen). \nIn  the  \"poorly attended\" condition,  observers  are  asked  to pay full  attention to \nthe central task  (and the blocks of trials for  which  performance for  the central  task \nfalls  below  a  certain cut-off are discarded). \n\nFour classical  pattern discrimination tasks were  investigated,  each  with  two volun(cid:173)\nteer  subjects  (average  shown  in  Figure  1),  similarly to  our  previous  experiments \n[7,  8].  Screen  luminance resolution  was  0.2%.  Screen  luminance  varied  from  1 to \n90cd/m2  (mean  45cd/m2),  room  illumination  was  5cd/m2  and  viewing  distance \n80cm.  The Yes/No (present/absent)  paradigm was used  (one stimulus presentation \nper trial).  Threshold  (75%  correct  peformance)  was  reached  using  a  staircase  pro(cid:173)\ncedure, and computed through a  maximum-likelihood fit  of a Weibull function  with \ntwo degrees  of freedom  to the  psychometric curves. \n\nExp. 2: Orientation discrimination \n\nf 20 J 15 \n\n.,  10 \nc: ,g \nS  5 \n~  ~~~~~ \n\n0.8 \n\n0.6 \nContrast \n\nMask contrast \n\n0.4 \n\n:2 \n.B \nII) \nII> \n\u00a3; \ni  0.2 \n.to c: \n8 \n\n\u00b0 \n\n:2 \n0 \nI: \n\ne \n\n-6 \n!0.2 \nc: \n0 \n() \n\n\u00b0 \u00b0 \n\nFigure  1:  Psychophysical  data  and  model  fits  using  the  parameters  from  Table  1 \n(P=poorly  and  F=fully  attended).  Gray  curves:  Model  predictions  for  fully  attended \ndata,  using  the poorly  attended parameters,  except  for  -y  =  2.9  and {)  = 2.1  (see  Results). \n\nExpo  1  measured  increment  contrast  discrimination threshold:  The  observer  dis(cid:173)\ncriminates between  a  4cpd  (cycles  per degree)  stochastic oriented mask [7]  at fixed \ncontrast, and the same mask plus a low-contrast sixth-derivative-of-Gaussian (D6G) \nbar;  threshold  is  measured  for  bar contrast  [8].  Expo  2  measured  orientation dis(cid:173)\ncrimination  thresholds:  The  observer  discriminates  between  a  vertical  and  tilted \ngrating  at  4cpd;  threshold  for  the  angle  difference  is  measured.  In  addition,  two \ncontrast  masking tasks  were  investigated  for  their  sensitivity  to  non-linearities  in \nvisual  processing.  A 4cpd stochastic mask  (50% contrast)  was  always present,  and \nthreshold  was  measured  for  the contrast  of a  vertical  superimposed  D6G  bar.  In \nExpo  3,  the  orientation  of the  masker  was  varied  and  its  spatial frequency  fixed \n(4cpd),  while  in  Expo  4  the spatial period of the masker was  varied and  its orien(cid:173)\ntation vertical.  Our  aim was to investigate very  dissimilar tasks, in particular with \nrespect  to the decision  strategy  used  by  the observer. \n\nUsing the dual-task paradigm, we found  mixed attentional effects on psychophysical \nthresholds,  including the appearance of a  more pronounced  contrast discrimination \n\n\f792 \n\nL.  ltti, J.  Braun,  D.  K.  Lee and C. Koch \n\n\"dipper\"  in  Exp.  1, substantial improvement of orientation thresholds  in  Exp.  2, \nand reduced  contrast elevations due  to masking in  Exps. 3-4 (also see  [7,  8]). \n\n3  MODEL \n\nThe  model  consists  of  three \nsuccessive  stages  [4,  5]. \nIn \nthe  first  stage,  a  bank  of \nGabor-like  linear  filters  ana(cid:173)\nlyzes  a  fixed  location  of the \nvisual  scene.  Here,  a  single(cid:173)\nscale  model  composed  of  12 \npairs  of filters  in  quadrature \nphase,  tuned  for  orientations \no E e evenly  spanning  1800 , \nwas  sufficient  to  account  for \nthe  data  (although  a  multi(cid:173)\nscale  model  may  account  for \na  wider range of psychophysical thresholds).  The linear filters  take  values  between \n0.0 and 100.0, then multiplied by  a gain factor A  (one of the ten free  parameters of \nthe model),  and to which  a small background activity  f.  is  added. \nIn  the  second  stage,  filters  non-linearly  interact  as  follows:  (1)  Each  unit receives \nnon-linear  self-excitation,  and  (2)  each  unit  receives  non-linear  divisive  inhibition \nfrom a pool of similarly-tuned units:  With E8  being the linear response from a unit \ntuned for  orientation 0,  the pooled  response  R8  is  given  by: \n\nwhere  W8(O')  = e -\n\n(/1'_/1)2 \n2E~ \n\nis  a  Gaussian  weighting function  centered  around  0,  and  1J  a  positive  constant  to \naccount  for  background  activity  in  the  pooling stage.  This stage  is  inspired  from \nHeeger's  model of gain control  in cat VI  [3,  4].  Our formulation, in  which  none  of \nthe  parameters  is  given  a  particular value,  however  allows  for  multiple outcomes, \nto  be  determined  by  fitting  the  model  to  our  psychophysical  data:  A  sigmoidal \n(S  > 0, I  > d')  as  well  as  simple  power-law  (S  = 0)  or  even  linear  (! = 1, d'  = 0) \ncontrast  response  characteristic  could  emerge,  the  responses  could  be  saturating \n(, = d')  or not (, i=  d'),  and the inhibitory pool size  (~8) could be broad or  narrow. \nBecause  striate  neurons  are  noisy,  physiological  noise  is  assumed  in  the  model  at \nthe outputs of the second  stage.  The noise level is chosen  close  to what is  typically \nobserved  in cortical  pyramidal cells,  and modeled by  Gaussian  noise  with  variance \nequal to mean taken to some power  a  determined  by fitting. \n\nBecause  the decision stage - which  quantitatively relates  activity in the population \nof pooled  noisy  units to  behavioral discrimination performance - is  not fully  char(cid:173)\nacterized  in  humans,  we  are  not  in  a  position  to model  it  in  any  detail.  Instead, \nwe  trained  our  subjects  (for  2-3  hours  on  each  task),  and  assume  that  they  per(cid:173)\nform close  to  an  \"optimal detector\".  Such  optimal detector  may be  characterized \nin  a  formal manner,  using  Statistical Estimation Theory  [4,  5].  We  assume that  a \nbrain  mechanism exists,  which,  for  a  given  stimulus presentation,  builds  an  inter(cid:173)\nnal  estimate of some stimulus attribute (  (e.g.,  contrast,  orientation, period).  The \ncentral  assumption of our decision  stage  is  that this brain mechanism will  perform \nclose  to  an  unbiased  efficient  statistic T,  which  is  the  best  possible  estimator of ( \n\n\fQuantitative Modeling of Attentional Modulation \n\n793 \n\ngiven  the  noisy  population  response  from  the  second  stage.  The  accuracy  (vari(cid:173)\nance)  with  which  T  estimates  (  can  be  computed  formally,  and  is  the  inverse  of \nthe  Fisher Information with respect  to (  [13,  4].  Simply put,  this means that, from \nthe  first  two  stages  of the  model  alone,  we  have  a  means  of computing  the  best \npossible estimation performance for  (, and consequently,  the best  possible discrim(cid:173)\nination  performance  between  two  stimuli  with  parameters  (1  and  (2  [4,  5].  Such \nstatistically efficient  decision  stage is  implementable as  a  neural  network  [13]. \n\nThis decision  stage  provides a  unified  framework for  optimal discrimination in  any \nbehavioral situation, and eliminates the need for task-dependent assumptions about \nthe  strategy  used  by  the  observers  to  perform  the  task in  a  near  optimal manner. \nOur model allows for a quantitative prediction of human psychophysical thresholds, \nbased  on  a crude simulation of the physiology of primary visual cortex  (area VI). \n\n4  RESULTS \n\nAll parameters in the model were automatically adjusted in order to best fit  the psy(cid:173)\nchophysical  data from  all  experiments.  A  multidimensional downhill simplex with \nsimulated annealing  overhead  was  used  to minimize the root-mean-square distance \nbetween  the  quantitative  predictions  of the  model  and  the  human  data  [4].  The \nbest-fit  parameters  obtained  independently  for  the  \"fully  attended\"  and  \"poorly \nattended\"  conditions are reported  in  Table 1.  The model's simultaneous fits  to our \nentire  dataset  are  plotted  in  Figure 1  for  both  conditions.  After  convergence  of \nthe fitting procedure,  a measure of how  well constrained each  parameter was by the \ndata was computed as follows:  Each parameter was systematically varied around its \nbest-fit  value,  in  0.5%  steps,  and  the fitting  error  was  recomputed;  the  amplitude \nby  which  each  parameter could  be varied  before  the fitting error increased by more \nthan  10%  of its  optimum is  noted  as  a  standard  deviation  in  Table  1.  A  lower \ndeviation indicates that the parameter is  more strongly constrained by the dataset. \n\nTable 1.  Model parameters for  both attentional conditions. \n\nSymbol \n\nName \nLinear  gaint \nActivity-independent  inhibitiont \nExcitatory  exponent \nInhibitory  exponent \nNoise  exponent \nBackground  activity,  linear  stage \nBackground  activity,  pooling  stage \nSpatial period  tuning  width X \nOrientation  tuning  width X \nOrientation  pooling  width X \nt Dynamic range  of linear  filters  is  [\u20ac \nx  For clarity,  FWHM  values  are given  rather  than  17  values  (FWHM  =  2I7J2ln(2\u00bb. \n\nfully  attended \nl.7 \u00b1 0.2 \n14.1 \u00b1 2.3 \n3.36 \u00b1 0.02 \n2.48 \u00b1 0.02 \nl.34 \u00b1 0.07 \nl.13 \u00b1 0.35 \n0.18 \u00b1 0.05 \n0.85 \u00b1 0.06  oct.  0.85 \u00b1 0.09  oct . \n26\u00b0 \u00b1 2.4\u00b0 \n48\u00b0  \u00b1 25\u00b0 \n\nA \nS \n'Y \n6 \na \nf \n7] \n(r>. \n(r8 \n~8 \n\npoorly attended \n8.2 \u00b1 0.9 \n10l.5 \u00b1 16.6 \n2.09 \u00b1 0.01 \nl.51 \u00b1 0.02 \n1.39 \u00b1 0.08 \n1.25 \u00b1 0.60 \n0.77 \u00b1 0.11 \n\n38\u00b0  \u00b1 5.5\u00b0 \n50\u00b0  \u00b1 26\u00b0 \n\n...  100.0  X  A + 4 \n\nAlthough no  human bias was  introduced during the fitting  procedure,  interestingly, \nall of the model's internal  parameters reached  physiologically plausible best-fit  val(cid:173)\nues,  such  as,  for  example,  slightly  supra-Poisson  noise  level  (a  ~ 1.35),  ~ 30\u00b0 \norientation  tuning  FWHM  (full-width  at  half-maximum),  and  ~ 0.85  octave  spa(cid:173)\ntial  period tuning FWHM.  Some of the internal characteristics  of the  model which \nmore closely  relate to the putative underlying physiological mechanisms are shown \nin  Figure 2. \n\n\f794 \n\na \n\nTransducer function \n\np \n\n0.4 \n0.8 \nContrast \n\n0.8 \n\nb \n\n0.8 \n\n5.o.s \nc: \n~0.4 \nen \n\n0.2 \n\nL. ltti, J.  Braun,  D.  K.  Lee and C.  Koch \n\nOrientation tuning \n\nC \n\nOrientation pooling \n\n0.8 \n\n..c: \n~o.s \ne \nCi5  0.4 \n\nF \n\n0  -40 \n\n-20 \n20 \nOrientation (deg) \n\n0 \n\n40 \n\nFigure 2:  Internals of the model.  (a) The response function  of individual  units to contrast \nwas  sigmoidal under full  (F) and almost linear under poor (P) attention.  (b) Native linear \norientation  tuning  was  broader  under  poor  (NP)  than  full  (NF)  attention,  but  it  was \nsharpened in both cases by pooling  (PP=pooled poor, and PF=pooled full  attention).  (c) \nThere was  no  difference  in  orientation  pooling  width under poor (P) or full  (F)  attention. \nUsing  poorly  attended parameters,  except  for  -y  =  2.9 and  ~ =  2.1  (grey  curves),  yielded \nsteep non-linear  contrast response,  and intermediary  tuning  (same  width as  NF). \n\nIn  Table  1,  attention  had  the following  significant  effects  on  the  model's  param(cid:173)\neters:  1)  Both  pooling exponents  (-y, d)  were  higher;  2)  the  tuning  width  (0\"/1)  was \nnarrower;  3)  the linear gain  (A)  and associated  activity-independent  inhibition  (5) \nwere  lower;  and  4)  the  background  activity  of the  pooling  stage  was  lower.  This \nyielded  increased  competition  between  filters:  The  network  behaved  more  like  a \nwinner-take-all under full  attention, and more like a linear network  of independent \nunits  under  poor  attention.  While  the  attentional  modulation  of \"d and  0\"/1  are \neasy  to interpret,  its effect  on  the A, 5  and 'fJ  is  more difficult  to understand. \nConsequently,  we  conducted  a  further  automatic  fit,  which,  starting  from  the \n\"poorly  attended\"  parameters,  was  only  allowed  to  alter, and  d  to fit  the  \"fully \nattended\"  data.  The motivation for  not varying 0\"/1  was  that we  observed significant \nsharpening of the tuning induced by higher exponents \"d (Figure 2) .  Also,  slight \nchanges  in  the difference  ,  - d can easily produce large changes in  the overall gain \nof the system, hence  compensating for  changes  in  A, 5  and 'fJ .  (We  however  do  not \nimply here that 0\"/1,  A, 5  and 'fJ  are redundant parameters; there is only a small range \naround  the  best-fit  point over  which, and d can  compensate for  variations in  the \nother parameters,  without dramatically impairing the quality of fit) . \nAlthough  the  new  fit  was  not  as  accurate  as  that  obtained  with  all  parameters \nallowed  to  vary,  it  appeared  that  a  simple modification  of the  pooling  exponents \nwell  captured  the  effect  of attention  (Figure  1).  Hence,  the  \"poorly  attended\" \nparameters  of Table  1  well  described  the  \"poorly  attended\"  data,  and  the  same \nparameters except for, = 2.9 and d = 2.1  well described  the  \"fully attended\"  data. \nA variety of other simple parameter modifications were also tested,  but none except \nfor  the pooling exponents  (-y,o)  could fully  account for the attentional modulation. \nThese  modifications  include:  Changes  in  gain  (obtained  by  modifying  A  only,  , \nonly, or d only), in tuning (0\"/1),  in the extent ofthe inhibitory pool (E/I),  and in the \nnoise  level  (a).  A  more systematic study,  in  which  all  possible  parameter subsets \nare successively  examined, is currently  in progress in our laboratory. \n\n5  DISCUSSION and  CONCLUSION \n\nAt  the  basis  of our  results  is  the  hypothesis  that  attention  might  modulate  the \nearlier  rather  than  the  later  stages  of  visual  processing.  We  found  that  a  very \n\n\fQuantitative Modeling of Attentional Modulation \n\n795 \n\nsimple,  prototypical,  task-independent  enhancement  of the  amount of competition \nbetween  early  visual  filters  accounts  well  for  the  human  data.  This  enhancement \nresulted from increases  in  parameters 'Y  and 5 in  the  model,  and was  paralleled by \nan increase  in contrast gain and a  sharpening in  orientation tuning.  Although it is \nnot  possible  from  our data to  rule out  any  attentional  modulation at  later  stages, \nour  hypothesis  has  recently  received  experimental  support  that  attention  indeed \nmodulates early visual processing  in  humans  [2,  14]. \n\nMore  psychophysical experiments are  needed  to investigate attentional modulation \nat  later  processing  stages.  For  example,  it  might  be  possible  to  study  the  effect \nof attention  on  the  decision  stage  by  manipulating  attention  during  experiments \ninvolving decision  uncertainty.  In  the  absence  of such  results,  we  have  attempted \nin  our  experiments  to  minimize  the  possible  impact  of attention  on  later  stages, \nby using only simple stimulus patterns devoid of conceptual or emotional meaning, \nsuch  as to involve as little as possible the more cognitive stages of visual processing. \n\nOur finding  that attention  may increase  the  amount of competition between  early \nvisual filters  is  accompanied  by  an enhancement  of the  gain and sensitivity  of the \nfilters,  and  by  a  sharpening  of their  tuning  properties.  The existence  of two  such \nprocessing  states  - one,  more sensitive  and selective  inside  the  focus  of attention, \nand  the  other,  more  broadly-tuned  and  non-specific  outside  - can  be  justified  by \nat  least  two  observations:  First,  the  higher  level  of  activity  in  attended  neurons \nconsumes more energy,  which  may not  be desirable over the entire extent  of visual \ncortices.  Second,  although less  efficient  for  fine  discriminations,  the  broadly-tuned \nand non-specific state may have greater ability at catching unexpected,  non-specific \nvisual  events. \nIn  this  perspective,  this  state  would  be  desirable  as  an  input  to \nbottom-up, visual  alerting mechanisms, which  monitor the rest  of our visual  world \nwhile  we  are focusing  on  a  specific  task  requiring  high focal  accuracy. \n\nAcknowledgements \n\nThis research  was supported  by  ONR and  NSF  (Caltech  ERG). \n\nReferences \n[1]  Bonnel AM,  Stein  JF,  Bertucci  P.  Q  J  Exp  Psychol fA} 1992;44(4):601-26 \n[2]  Gandhi  SP,  Heeger DJ,  Boynton  GM.  Inv  Opht  Vis  Sci  (ARVO'98) 1998;39(4):5194 \n[3]  Heeger DJ.  Vis  Neurosci 1992;9:181-97 \n[4]  Itti  L,  Braun  J,  Lee  DK,  Koch  C.  Proc  NIPS *97 {in  press) \n[5]  Itti L,  Koch  C,  Braun  J.  Inv  Opht  Vis  Sci  (Proc  ARVO'98) 1998;39(4):2934 \n[6]  Koch  C,  Ullman  S.  Hum  NeurobioI1985;4:219-27 \n[7]  Lee  DK,  Koch C,  Braun  J.  Vis  Res  1997:37(17):2409-18 \n[8]  Lee  DK, Koch C,  Itti L,  Braun J .  Inv Opht  Vis  Sci (Proc  ARVO'98) 1998;39(4):2938 \n[9]  Luck SJ, Chelazzi  L,  Hillyard  SA,  Desimone  R.  J  Neurophysio/1997;77{I):24-42 \n[10]  Maunsell  JH.  Science 1995;270(5237)764-9 \n[11]  Motter BC.  J  NeurophysioiI993;70(3):909-19 \n[12]  Nakayama K,  Mackeben  M.  Vis  Res  1989;29(11):1631-47 \n[13]  Pouget  A,  Zhang  K,  Deneve  S,  Latham PE.  Neur  Comp  1998;10:373-401 \n[14]  Somers  DC,  et  al.  lnv  Opht  Vis  Sci  (Proc  ARVO'98) 1998;39(4):5192 \n[15]  Sperling  G,  Melchner  MJ.  Science  1978;202:315-8 \n[16]  Spitzer  H,  Desimone  R,  Moran  J.  Science 1988;240(4850):338-40 \n[17]  Treue  S,  Maunsell  JH.  Nature 1996;382(6591):539-41 \n\n\f", "award": [], "sourceid": 1547, "authors": [{"given_name": "Laurent", "family_name": "Itti", "institution": null}, {"given_name": "Jochen", "family_name": "Braun", "institution": null}, {"given_name": "Dale", "family_name": "Lee", "institution": null}, {"given_name": "Christof", "family_name": "Koch", "institution": null}]}