{"title": "A Superadditive-Impairment Theory of Optic Aphasia", "book": "Advances in Neural Information Processing Systems", "page_first": 66, "page_last": 72, "abstract": "", "full_text": "A Superadditive-Impairment Theory \n\nof Optic Aphasia \n\nMichael C. Mozer \n\nDept.  of Computer Science \n\nUniversity of Colorado \n\nMark Sitton \n\nDept.  of Computer Science \n\nUniversity of Colorado \n\nBoulder;  CO  80309-0430 \n\nBoulder;  CO  80309-0430 \n\nMartha Farah \n\nDept. of Psychology \n\nUniversity of Pennsylvania \n\nPhila.,  PA  19104-6196 \n\nAbstract \n\nAccounts  of  neurological  disorders  often  posit  damage  to  a  specific \nfunctional  pathway of the brain.  Farah (1990) has  proposed an  alterna(cid:173)\ntive  class  of explanations  involving  partial  damage  to  multiple  path(cid:173)\nways. We explore this explanation for optic aphasia, a disorder in which \nsevere  perfonnance  deficits  are  observed  when  patients  are  asked  to \nname  visually  presented objects, but  surprisingly,  performance  is  rela(cid:173)\ntively  nonnal  on  naming  objects  from  auditory  cues  and  on  gesturing \nthe appropriate use  of visually  presented objects.  We  model  this  highly \nspecific deficit through partial damage to two pathways-one that maps \nvisual  input to  semantics, and  the other that maps semantics  to  naming \nresponses.  The  effect  of  this  damage  is  superadditive,  meaning  that \ntasks  which  require  one  pathway  or the  other show  little  or  no  perfor(cid:173)\nmance  deficit,  but the  damage  is  manifested  when  a  task  requires  both \npathways (i.e.,  naming  visually  presented objects).  Our model explains \nother  phenomena  associated  with  optic  aphasia,  and  makes  testable \nexperimental predictions. \n\nNeuropsychology is the study of disrupted cognition resulting from  damage to functional \nsystems in the brain. Generally, accounts of neuropsychological disorders posit damage to \na  particular  functional  system  or  a  disconnection  between  systems.  Farah  (1990)  sug(cid:173)\ngested an alternative class of explanations for  neuropsychological disorders:  partial  dam(cid:173)\nage  to  multiple  systems,  which  is  manifested  through  interactions  among  the  loci  of \ndamage. We explore this explanation for the neuropsychological disorder of optic aphasia. \nOptic aphasia, arising from unilateral left posterior lesions, including occipital cortex \nand  the splenium of the corpus callosum (Schnider,  Benson, &  Scharre,  1994),  is  marked \nby  a deficit  in  naming  visually  presented objects,  hereafter referred  to  as  visual naming \n(Farah,  1990).  However,  patients  can  demonstrate  recognition  of  visually  presented \nobjects nonverbally,  for  example, by gesturing the appropriate use  of an object or sorting \nvisual  items  into  their  proper  superordinate  categories  (hereafter,  visual  gesturing). \nPatients  can  also  name  objects  by  nonvisual  cues  such  as  a  verbal  definition  or  typical \nsounds made by  the objects (hereafter, auditory naming). The highly specific nature of the \ndeficit rules out an explanation in terms of damage to a single pathway in a standard model \nof visual naming (Figure  1), suggesting that a more complex model  is required, involving \n\n\fA Superadditive-Impainnent Theory of Optic Aphasia \n\n67 \n\nFIGURE  1.  A  standard  box-and-arrow \nmodel  of visual  naming. The  boxes  denote \nlevels  of  representation,  and  the  arrows \ndenote pathways mapping from one level of \nrepresentation  to  another.  Although  optic \naphasia cannot be explained by damage to \nthe  vision-to-semantics  pathway  or  the \nsemantics-ta-naming \nFarah \n(1990) proposed an explanation in terms of \npartial damage to both pathways (the X's). \n\npathway, \n\nvisual \n\nauditory \n\nmUltiple semantic systems or multiple pathways to visual  naming. However, a mere parsi(cid:173)\nmonious  account  is  suggested  by  Farah  (1990):  Optic  aphasia  might  arise  from  partial \nlesions to two pathways in  the  standard model-those connecting visual  input to seman(cid:173)\ntics,  and semantics to naming-and the effect of damage  to  these  pathways is superaddi(cid:173)\ntive,  meaning that tasks  which require only one of these  pathways (e.g.,  visual gesturing, \nor auditory naming) will be  relatively unimpaired, whereas tasks requiring both pathways \n(e.g., visual naming) will show a significant deficit. \n\n1  A MODEL OF SUPERADDITIVE IMPAIRMENTS \nWe  present a computational model of the superadditive-impairment theory of optic apha(cid:173)\nsia by elaborating the architecture of Figure  1. The architecture has four pathways:  visual \ninput  to  semantics  (V~S), auditory  input  to  semantics  (A~S), semantics  to  naming \n(S~N), and semantics to gesturing (S~G). Each pathway acts as an associative memory. \nThe critical property of a pathway that is required to explain optic aphasia is a speed-accu(cid:173)\nracy  trade off:  The initial  output of a pathway  appears  rapidly,  but it  may  be  inaccurate. \nThis \"quick and dirty\" guess is refined over time, and the pathway output asymptotically \nconverges on the best interpretation of the input. \n\nWe  implement  a  pathway  using  the  architecture  suggested  by  Mathis  and  Mozer \n(1996).  In  this  architecture, inputs are mapped  to their best interpretations by  means of a \ntwo-stage  process  (Figure  2).  First,  a  quick,  one-shot mapping is  performed  by  a  multi(cid:173)\nlayer feedforward connectionist network to transform the input directly to its correspond(cid:173)\ning  output.  This  is  followed  by  a  slower  iterative  clean-up  process  carried  out  by  a \nrecurrent  attractor network.  This  architecture  shows  a speed-accuracy  trade off by  virtue \nof the .assumption  that  the  feed forward  mapping  network  does  not  have  the  capacity  to \nproduce exactly  the  right output to  every  input, especially  when  the  inputs are corrupted \nby noise or are otherwise incomplete. Consequently, the clean up stage is  required to pro(cid:173)\nduce a sensible interpretation of the noisy output of the mapping network. \n\nFully distributed attractor networks have been used  for  similar purposes (e.g., Plaut \n&  Shallice,  1993).  For  simplicity,  we  adopt  a  localist -attractor  network  with  a  layer  of \nstate  units  and  a  layer of radial  basis  function  (RBF)  units,  one  RBF  unit  per attractor. \nEach RBF or attractor unit measures the distance of the current state to the attractor that it \nrepresents. The activity of attractor unit i, aj, is: \n\nFIGURE 2. Connectionist implementa(cid:173)\ntion of a processing pathway. The path(cid:173)\nway  consists  of  feedforward  mapping \nnetwork followed by a  recurrent clean(cid:173)\nup or attractor network. Circles denote \nconnectionist  processing  units  and \narrows  denote  connections  between \nunits or between layers of units. \n\nclean up network \n\nmapping network \n\n\f68 \n\nM.  C.  Mozer,  M. Sitton and M.  Farah \n\n(1) \n\n(2) \n\nwhere set) is the state unit activity vector at time t,  ~i is  the vector denoting the location of \nattractor i,  and  ~i is the strength of the attractor. The strength detennines the region of the \nstate  space over which an attractor will exert its  pull,  and  also the rate  at  which  the  state \nwill  converge to the attractor. The state units receive input from  the  mapping network and \nfrom the attractor units and are updated as  follows: \n\ns;(t)  = d;(t)e;(t) + (I - d;(t\u00bb  La/t -I )~j; \n\nj \n\n(3) \n\nwhere sift) is the activity of state unit i at time t,  e;  is the ith output of the  mapping net, ~j; \nis the ith element of attractor j, and d;  is  given by \n\nd .(t)  = h[I __ e;(_t-..,.-_l)] \n\nej(t) \n\nI \n\n(4) \n\nwhere h[.]  is  a linear threshold function  that bounds  activity  between -1  and +1,  ej  IS  a \nweighted time average of the ith output of the mapping net, \nej(t)  = cxej(t) + (I -CX)e;(t-I) \n\n(5) \n\nIn all simulations, cx = .02. \n\nThe activity of the  state  units  are governed  by  two  forces-the external  input from \nthe  feedforward  net  (first  tenn  in  Equation  3)  and  the  attractor  unit  activities  (second \ntenn). The parameter d; acts as a kind of attentional mechanism that modulates the relative \ninfluence of these two forces.  The basic idea is that when  the input coming from  the map(cid:173)\nping  net is  changing,  the  system should be  responsive  to  the  input and  should  not yet be \nconcerned  with  interpreting the  input.  In  this case,  the  input is  copied straight through to \nthe state units and hence dj  should have a value close to  I. When the input begins to stabi(cid:173)\nlize, however,  the  focus  shifts to  interpreting the input and following  the dynamics of the \nattractor network. This shift corresponds  to  d j  being  lowered  to  zero.  The weighted  time \naverage in the update rule for dj  is what allows for the smooth transition of the function  to \nits  new  value.  For certain constructions of the  function  d,  Zemel and Mozer (in  prepara(cid:173)\ntion) have proven convergence of the algorithm to an attractor. \n\nApart from  speed-accuracy trade off,  these dynamics have another important conse(cid:173)\n\nquence for the present model, particularly with respect to cascading pathways. If pathway \nA feeds into pathway S, such as V~S feeding into S~N, then the state unit activities of A \nact as  the  input to  S. Because these activities change over time  as  the  state approaches a \nwell-fonned state, the dynamics of pathway S can be quite complex as  it is forced to deal \nwith an unstable input. This property is important in explaining several phenomena associ(cid:173)\nated with optic aphasia. \n\n1.1  PATTERN GENERATION \nPatterns were constructed for each of the five  representational spaces: visual and auditory \ninput,  semantic, name and gesture responses.  Each representational space  was  arbitrarily \nmade  to  be  200  dimensional.  We  generated  200  binary-valued  (-1,+1)  patterns  in  each \nspace, which were meant to correspond to known entities of that representational domain. \nFor the visual, auditory, and semantic spaces, patterns were partitioned into 50 simi(cid:173)\n\nlarity  clusters  with  4 \u00b7siblings per cluster.  Patterns  were  chosen  randomly  subject  to  two \nconstraints: patterns in different clusters had to be at least 80\u00b7 apart, and siblings had to be \nbetween 25\u00b7 and 50\u00b7 apart.  Because similarity of patterns  in  the  name and gesture spaces \nwas irrelevant to  our modeling,  we  did  not impose a similarity structure on  these  spaces. \n\n\fA Superadditive-Impainnent Theory of Optic Aphasia \n\n69 \n\nInstead,  we  generated  patterns  in  these  spaces  at  random  subject  to  the  constraint  that \nevery pattern had to be at least 60\u00b7 from every other. \n\nAfter generating patterns in each of the representational spaces, we established arbi(cid:173)\n\ntrary  correspondences  among  the  patterns  such  that  visual  pattern  n.  auditory  pattern n, \nsemantic pattern n, name pattern n,  and gesture pattern n all represented the same concept. \nThat  is,  the  appropriate  response  in  a  visual-naming  task  to  visual  pattern  n  would  be \nsemantic pattern n and name pattern n. \n\n1.2  TRAINING PROCEDURE \nThe  feedforward  networks  in  the  four  pathways  (V-\"7S,  A-\"7S,  S-\"7N,  and  S-\"70)  were \nindependently  trained on all  200 associations using back propagation.  Each of these  net(cid:173)\nworks contained a single  hidden  layer of 150 units, and all  units  in  the  network used the \nsymmetric activation function  to give activities in the range [-1,+1]. The amount of train(cid:173)\ning  was chosen such that performance on the  training examples was not perfect;  usually \nseveral elements in the output would be erroneous-i.e., have the wrong sign-and others \nwould  not be  exactly correct-i.e., -lor + 1.  This was  done  to embody  the  architectural \nassumption  that  the  feedforward  net  does  not  have  the  capacity  to  map  every  input  to \nexactly the right output, and hence, the clean-up process is required. \n\nTraining was not required for the clean-up network. Due to the localist representation \nof attractors in the clean-up network, it was trivial to hand wire each clean-up net with the \n200 attractors  for  its  domain,  along  with  one  rest-state  attractor.  All  attractor strengths \nwere initialized to the same value,  ~= 15, except the rest-state attractor, for which P=5. The \nrest-state  attractor required a  lower strength so that even  a  weak external  input would be \nsufficient to kick the attractor network out of the rest state. \n\n1.3  SIMULATION METHODOLOGY \n\nAfter  each  pathway  had  been  trained,  the  model  was  damaged  by  \"lesioning\"  or \nremoving a  fraction y of the connections in  the  V -\"7S  and  S-\"7N  mapping  networks. The \nlesioned connections were chosen at random and an equal fraction  was removed from the \ntwo pathways. The clean-up nets were not damaged. The architecture was damaged a total \nof 30 different times, creating 30 simulated patients who  were tested on each of the four \ntasks and on all  200 input patterns for a  task.  The results we report come from averaging \nacross simulated patients and  input patterns. Responses were detennined after the system \nhad been given sufficient time to relax into a name or gesture attractor, which was taken to \nbe the response.  Each response was classified as one of the following  mutually exclusive \nresponse types: correct, perseveration (response is the same as that produced on any of the \nthree  immediately preceding  trials),  visual error (the  visual  pattern  corresponding to  the \nincorrect response is a sibling of the visual pattern corresponding to the correct response), \nsemantic error, visual+semantic error, or other error. \n\n1.4  PRIMING MECHANISM \nPriming-the  increased  availability  of  recently  experienced  stimuli-has  been  found \nacross a  wide variety of tasks in  normal subjects. We included priming in our model as a \nstrengthening  (increasing  the  Pi  parameter)  of recently  visited  attractors  (see  Mathis  & \nMozer  1996,  for  details,  and  Becker,  Behrmann,  &  Moscovitch,  1993,  for  a  related \napproach). In the damaged model, this mechanism often gave rise to perseverations. \n\n2  RESULTS \nWe have examined the model's behavior as we varied the amount of damage, quantified by \nthe parameter y.  However, we report on the perfonnance of simulated patients with y = .30. \nThis intermediate amount of damage yielded no floor or ceiling effects, and also produced \nerror rates for  the visual-naming task in the  range of 30-40%, roughly the median perfor(cid:173)\nmance of patients in the literature. \n\n\f70 \n\nM.  C.  Mozer,  M.  Sitton and M.  Farah \n\nTABLE  1.  Error rate of the \ndamaged  model  on  various \ntasks. \n\ntask \n\naUOltory gesturing \nauditory naming \nvisual gesturing \nvisual naming \n\nerror rate \nU.U\"'/o \n0.5% \n8.7% \n36.8% \n\nTable  1 presents  the  error  rates  of the  model  on  four  tasks.  The  pattern  of errors \n\nshows a qualitative fit  to  human patient data.  The model produced no  errors on  the  audi(cid:173)\ntory gesturing task because the two component pathways (A~S and  S~G) were undam(cid:173)\naged.  Relatively few  errors were made on the auditory-naming and visual-gesturing tasks, \neach  of which  involved  one  damaged  pathway,  because  the  clean-up  nets  were  able  to \ncompensate for the damage. However,  the  error rate for the  visual-naming task was quite \nlarge, due t'O  damage on both of its component pathways (V~S and  S~N). The error rate \nfor  visual  naming cannot be accounted  for  by  summing the  effects  of the  damage  to  the \ntwo  component  pathways  because  the  sum  of the  error  rates  for  auditory  naming  and \nvisual  gesturing,  each  of which  involves  one of the  two  partially damaged  pathways,  is \nnearly  four  times  smaller.  Rather,  the  effects  of damage  on  these  pathways  interact,  and \ntheir interaction leads to superadditive impairments. \n\nWhen a visual pattern is presented to the model, it is mapped by the damaged V~S \n\npathway into a corrupted semantic representation which is then cleaned up. While the cor(cid:173)\nruption is  sufficiently minor that clean  up  will eventually succeed, the clean up process is \nslowed considerably by  the  corruption.  During the  period of time  in  which the  semantic \nclean-up network is  searching for the correct attractor, the corrupted semantic representa(cid:173)\ntion  is  nonetheless fed  into the damaged  S~N pathway. The combined effect of the (ini(cid:173)\ntially)  noisy  semantic  representation  serving  as  input  to  a  damaged  pathway  leads  to \ncorruption of the naming representation past the point where it can be cleaned-up properly. \nInteractions in the  architecture  are  inevitable,  and  are  not  merely  a consequence of \nsome arbitrary  assumption  that  is  built  into  our model.  To  argue  this  point,  we  consider \ntwo  modifications to  the  architecture that might eliminate the interaction  in  the damaged \nmodel.  First,  if we  allowed  the  V~S pathway  to  relax  into  a  well-formed  state  before \nfeeding  its output into the  S~N pathway, there would  be  little interaction-the effects of \nthe  damage  would  be  additive.  However,  cortical  pathways  do  not operate  sequentially, \none  stage  finishing  its computation and  then  turning on  the  next  stage.  Moreover,  in  the \nundamaged  brain,  such  a  processing  strategy  is  unadaptive,  as  cascading  partial  results \nfrom  one  pathway  to  the  next  can  speed  processing  without  the  introduction  of errors \n(McClelland,  1979).  Second,  the  interaction  might  be  eliminated  by  making  the  S~N \npathway continually responsive to  changes in  the output of the  V ~S pathway.  Then, the \nrate of convergence of the V ~S pathway would be irrelevant to determining the eventual \noutput of the  S~N pathway.  However,  because the output of the  S~N pathway depends \nnot only on its input but its internal state (the state of the clean-up net), one cannot design \na pathway that is continually responsive to changes in the input and is also able to clean up \nnoisy  responses. Thus, the  two modifications one might consider to eliminate the interac(cid:173)\ntions in  the damaged model seriously weaken the computational power of the undamaged \nmodel. We  therefore conclude that the framework of our model makes it difficult to avoid \nan interaction of damage in two pathways. \n\nA subtle yet significant aspect of the model's performance is that the error rate on the \nvisual-gesturing task  was  reliably higher than  the error rate on  the  auditory-naming task, \ndespite the fact that each task made use of one damaged pathway, and the  pathways were \ndamaged to the same degree. The difference in performance is due to the fact that the dam(cid:173)\naged pathway for the  visual-gesturing task is  the first in a cascade of two, while the dam(cid:173)\naged  pathway  for  the  auditory-naming  task  is  the  second.  The  initially  noisy  response \nfrom  a  damaged  pathway  early  in  the  system  propagates  to  subsequent  pathways,  and \n\n\fA Superadditive-Impairment Theory of Optic Aphasia \n\n71 \n\nalthough  the  damaged  pathway  will  eventually  produce  the  correct response,  this  is  not \nsufficient to ensure that subsequent pathways will do so as well. \n\n2.1  DISTRIBUTION OF ERRORS FOR VISUAL OBJECT NAMING \n\nFigure 2 presents the  model's error distribution for the  visual-naming  task.  Consis(cid:173)\ntent with the patient data (Farah,  1990), the model produces many more semantic and per(cid:173)\nseveration  errors  than  by  chance.  The  chance  error  proportions  were  computed  by \nassuming that if the correct response was not made, then all other responses had an equal \nprobability of being chosen. \n\nTo understand the predominance of semantic errors, consider the effect of damage to \nthe V ~S pathway. For relatively small amounts of damage, the mapping produced will be \nclose  to  the  correct  mapping.  \"Close\"  here  means  that  the  Euclidean  distance  in  the \nsemantic  output space between  the correct and  perturbed  mapping  is  small.  Most of the \ntime,  minor  perturbation  of the  mapping  will  be  compensated  for  by  the  clean-up  net. \nOccasionally, the perturbation will  land the model in a different attractor basin, and a dif(cid:173)\nferent response will be made. However, when the wrong attractor is selected, it will be one \n\"close\" to the correct attractor, i.e., it will likely be a sibling in the same pattern cluster as \nthe correct attractor.  In the case of the V ~S pathway, the siblings of the correct attract or \nare  by  definition  semantically  related. A  semantic  error  will  be  produced  by  the  model \nwhen a sibling semantic attractor is chosen, and then this pattern is correctly mapped to a \nnaming response in  the  S~N pathway. \n\nIn addition to semantic errors, the other frequent error type in visual naming is perse(cid:173)\nverations. The priming mechanism is responsible for the significant number of persevera(cid:173)\ntions,  although  in  the  unlesioned  model,  it  facilitates  processing  of  repeated  stimuli \nwithout producing perseverations. \n\nJust as important as  the presence of perseverative and semantic errors is the absence \nof visual  errors,  a  feature  of optic  aphasia  that  contrasts  sharply  with  visual  agnosia \n(Farah,  1990). The same mechanisms explain  why  the  rate  of visual  errors  is  close to its \nchance  value  and  why  visual+semantic  errors  are  above  chance.  Visual-naming  errors \noccur because there  is  an error either in  the  V ~S or S~N mappings, or both.  Since the \nerroneous outputs of these  pathways show  a strong  tendency  to  be  similar to  the  correct \noutput,  and  because  semantic  and  name  similarity  does  not  imply  visual  similarity  (the \npatterns were paired randomly), visual errors should only occur by chance. When a visual \nerror does occur, though, there is a high probability that the error is also semantic because \nof the strong bias that already exists toward producing semantic errors. This is the reason \nwhy  more  visual+semantic  errors  occur than  by  chance and  why  the  proportion of these \n\n\u2022 \n\n\u2022 \n\nactual \n\ncI1ance \n\nFIGURE 3. Distribution \nof error types  made  by \nmodel on the V ~N task \n(black  bars)  relative  to \nchance (grey bars). \n\nvisual \n\nvis+sem \nError type \n\npe~Yendwe \n\n\f72 \n\nM.  C.  Mozer,  M. Sitton and M  Farah \n\nerrors is only slightly less than the proportion of visual errors. \n\nPlaut and  Shallice  (1993)  have  proposed  a connectionist  model  to  account  for  the \ndistribution of errors  made by  optic  aphasics.  Although  their model  was  not designed to \naccount for any of the other phenomena associated with the disorder, it has much in com(cid:173)\nmon  with  the  model  we  are  proposing.  Unlike  our  model,  however,  theirs  requires  the \nassumption  that  visually  similar objects  also  share  semantic  similarity.  This  assumption \nmight  be  questioned,  especially  because  our model  does  not  require  this  assumption  to \nproduce the correct distribution of error responses. \n\n3  DISCUSSION \nIn  demonstrating  superadditive  effects  of damage,  we  have  offered  an  account  of optic \naphasia  that  explains  the  primary  phenomenon:  severe  impairments  in  visual  naming  in \nconjunction with relatively spared performance on naming from verbal description or ges(cid:173)\nturing the appropriate use of a visually presented object. The model also explains the dis(cid:173)\ntribution  of errors  on  visual  naming.  Although  we  did  not  have  the  space  in  this  brief \nreport to  elaborate,  the  model  accounts  for  several  other distinct characteristics of optic \naphasia, including the tendency of patients to \"home in\" on the correct name for a visually \npresented object when  given  sufficient time,  and a positive correlation  between  the error \nrates on naming and gesturing responses to a visual object (Sitton, Mozer, &  Farah,  1998). \nFurther,  the  model  makes  several  strong  predictions  which  have  yet  to  be  tested  experi(cid:173)\nmentally. One such prediction, which was apparent in the results presented earlier, is that a \nhigher error rate should be observed on visual gesturing than on auditory naming when the \ntasks are equated for difficulty, as our simulation does. \n\nMore generally,  we  have  strengthened  the  plausibility of Farah's (1990)  hypothesis \n\nthat  partial  damage  to  two  processing  pathways  may  result  in  close-to-normal  perfor(cid:173)\nmance on  tasks  involving one  pathway or the other while yielding  a severe performance \ndeficit on  tasks involving  both damaged pathways.  The  superadditive-impairment theory \nthus may  provide a more parsimonious account of various disorders that were previously \nbelieved to require more complex architectures or explanations. \n\n4  ACKNOWLEDGMENTS \nThis research was supported by  grant 97-18 from  the McDonnell-Pew Program in Cogni(cid:173)\ntive Neuroscience. \n\n5  REFERENCES \nBecker, S., Behrmann, M., &  Moscovitch, K. (1993). Word priming in attractor networks. \nProceedings of the Fifteenth Annual Conference of the  Cognitive Science Society (pp. \n231-236). Hillsdale, NJ:  Erlbaum. \n\nFarah, M. J.  (1990). Visual agnosia. Cambridge, MA: MIT PresslBradford Books. \nMathis, D.  W., &  Mozer, M. C. (1996). Conscious and unconscious perception: A compu(cid:173)\ntational theory. In G. Cottrell (Ed.), Proceedings of the Eighteenth Annual Conference \nof the Cognitive Science Society (pp.  324-328). Hillsdale, NJ:  Erlbaum. \n\nMcClelland, J. L. (1979).  On the time relations of mental processes: An examination of \n\nsystems of processes in cascade. Psychological Review, 86, 287-330. \n\nPlaut, D., & Shallice, T.  (1993). Perseverative and semantic influences on visual object \n\nnaming errors in optic aphasia: A connectionist approach. Journal of Cognitive Neuro(cid:173)\nscience, 5,89-112. \n\nSchnider, A., Benson, D.  E, and Scharre, D.  W.  (1994). Visual agnosia and optic aphasia: \n\nAre they anatomically distinct? Cortex, 30, 445-457. \n\nSitton, M., Mozer, M. C., & Farah, M. (1998). Diffuse lesions in a modular connectionisl \n\narchitecture: An account of optic aphasia. Manuscript submitted for publication. \n\n\f", "award": [], "sourceid": 1469, "authors": [{"given_name": "Michael", "family_name": "Mozer", "institution": null}, {"given_name": "Mark", "family_name": "Sitton", "institution": null}, {"given_name": "Martha", "family_name": "Farah", "institution": null}]}