{"title": "A Hippocampal Model of Recognition Memory", "book": "Advances in Neural Information Processing Systems", "page_first": 73, "page_last": 79, "abstract": "", "full_text": "A Hippocampal Model of Recognition Memory \n\nRandall C. O'Reilly \n\nDepartment of Psychology \n\nUniversity of Colorado at Boulder \n\nCampus Box 345 \n\nBoulder, CO 80309-0345 \n\noreilly@psych.colorado.edu \n\nKenneth A. Norman \n\nDepartment of Psychology \n\nHarvard University \n33 Kirkland Street \n\nCambridge, MA 02138 \n\nnonnan@wjh.harvard.edu \n\nJames L. McClelland \n\nDepartment of Psychology and \n\nCenter for the Neural Basis of Cognition \n\nCarnegie Mellon University \n\nPittsburgh, PA  15213 \njlm@cnbc.cmu.edu \n\nAbstract \n\nA  rich  body  of data  exists  showing that  recollection of specific  infor(cid:173)\nmation makes an important contribution to  recognition memory,  which \nis distinct from the contribution of familiarity, and is not adequately cap(cid:173)\ntured by existing unitary memory models. Furthennore, neuropsycholog(cid:173)\nical evidence indicates that recollection is sub served by the hippocampus. \nWe  present a  model,  based  largely  on  known features  of hippocampal \nanatomy and physiology, that accounts for the following key character(cid:173)\nistics of recollection:  1) false recollection is rare (i.e., participants rarely \nclaim to recollect having studied nonstudied items), and 2) increasing in(cid:173)\nterference leads to less recollection but apparently does not compromise \nthe quality of recollection (i.e., the extent to which recollected infonna(cid:173)\ntion veridically reflects events that occurred at study). \n\n1 \n\nIntroduction \n\nFor nearly 50 years, memory researchers have known that our ability to remember specific \npast episodes depends critically on the hippocampus.  In this paper, we describe our initial \nattempt to use a mechanistically explicit model of hippocampal function to explain a wide \nrange of human memory data. \n\nOur understanding of hippocampal function from a computational and biological perspec-\n\n\f74 \n\nR.  C.  0 'Reilly, K.  A. Norman and 1.  L  McClelland \n\ntive is based on our prior work (McClelland, McNaughton, &  O'Reilly,  1995; O'Reilly & \nMcClelland,  1994).  At the broadest level,  we think that the hippocampus exists in part to \nprovide a memory system which can learn arbitrary information rapidly without suffering \nundue amounts of interference.  This memory system sits on top of, and works in conjunc(cid:173)\ntion with, the neocortex, which learns slowly over many experiences, producing integrative \nrepresentations of the relevant statistical features of the environment. The hippocampus ac(cid:173)\ncomplishes rapid, relatively interference-free learning by using relatively non-overlapping \n(pattern separated) representations.  Pattern separation occurs as a result of 1) the sparse(cid:173)\nness of hippocampal representations (relative to  cortical representations), and 2) the fact \nthat hippocampal units are sensitive to conjunctions of cortical features -\ngiven two cor(cid:173)\ntical  patterns  with  50%  feature  overlap,  the  probability that  a  particular conjunction of \nfeatures will be present in both patterns is much less than 50%. \n\nWe propose that the hippocampus produces a relatively high-threshold, high-quality recol(cid:173)\nlective response to  test items.  The response  is  \"high-threshold\" in the sense that studied \nitems sometimes trigger rich  recollection (defined as  \"retrieval of most or all  of the  test \nprobe's features from memory\") but lures never trigger rich recollection.  The response is \n\"high-quality\" in the sense that,  most of the  time,  the recollection signal consists of part \nor all  of a  single  studied pattern, as  opposed to a  blend of studied  patterns.  The  high(cid:173)\nthreshold,  high-quality nature  of recollection can be  explained in  terms  of the  conjunc(cid:173)\ntivity of hippocampal representations:  Insofar as recollection is a function of whether the \nfeatures of the test probe were encountered together at study,  lures (which contain many \nnovel feature conjunctions, even if their constituent features  are familiar) are unlikely to \ntrigger rich recollection;  also,  insofar as  the hippocampus stores feature conjunctions (as \nopposed to  individual  features),  features  which  appeared together at  study are  likely  to \nappear together at test.  Importantly, in  accordance with dual-process accounts of recog(cid:173)\nnition memory (Yonelinas,  1994; Jacoby,  Yonelinas,  &  Jennings,  1996),  we  believe that \nhippocampally-driven recollection is  not the  sole contributor to  recognition memory per(cid:173)\nformance.  Rather.  extensive evidence exists that recollection is complemented by a \"fall(cid:173)\nback\" familiarity signal which participants consult when rich recollection does not occur. \nThe familiarity signal is mediated by as-yet unspecified areas (likely including the parahip(cid:173)\npocampal temporal cortex:  Aggleton & Shaw,  1996; Miller & Desimone,  1994). \n\nOur account differs substantially from most other computational and mathematical models \nof recognition memory.  Most of these  models compute the  \"global match\" between the \ntest probe and stored memories (e.g .\u2022 Hintzman, 1988; Gillund & Shiffrin, 1984); recollec(cid:173)\ntion in these models involves computing a similarity-weighted average of stored memory \npatterns. In other memory models, recollection of an item depends critically on the extent \nto  which the  components of the item's representation were linked with that of the  study \ncontext (e.g., Chappell &  Humphreys, 1994). Critically, recollection in all of these models \nlacks the high-threshold, high-quality character of recollection in our model.  This is most  -\nevident when we  consider the effects  of manipulations which increase  interference (e.g., \nincreasing the length of the study list.  or increasing inter-item similarity).  As interference \nincreases, global matching models predict increasingly \"blurry\" recollection (reflecting the \ncontribution of more items to the composite output vector), while the other models predict \nthat false recollection of lures will increase. In contrast, our model predicts that increasing \ninterference should lead to decreased correct recollection of studied test probes, but there \nshould be no concomitant increase in \"erroneous\" types of recollection (i.e.,  recollection \nof details which mismatch studied test probes, or rich recollection of lures).  This predic(cid:173)\ntion is consistent with the recent finding that correct recollection of studied items decreases \nwith increasing list length (Yonelinas,  1994). Lastly, although extant data certainly do not \ncontradict the  claim that  the  veridicality of recollection is  robust  to  interference,  we  ac(cid:173)\nknowledge that additional,  focused experimentation is needed to  definitively resolve this \nissue. \n\n\fA Hippocampal Model of Recognition Memory \n\n75 \n\n/ \n\n,\\ \n'---;----; \n--- --\n\nI \nC~-:;~ _  1 \nL __ oL::..J' \n\n/\" \n\n-\n\n'-\n\nFigure  I:  The model.  a) Shows the areas and  connectivity, and  the corresponding columns  within \nthe Input,  EC. and CAl  (see text).  b) Shows an  example activity pattern.  Note the sparse activity in \nthe DG and CA3, and intermediate sparseness of the CAL \n\n2  Architecture and Overall Behavior \n\nFigure  I  shows a  diagram of our model,  which contains the basic  anatomical regions of \nthe  hippocampal  formation,  as  well  as  the  entorhinal  cortex  (EC),  which  serves  as  the \nprimary cortical input/output pathway for the hippocampus. The model as described below \ninstantiates a series of hypotheses about the structure and function of the hippocampus and \nassociated cortical areas, which are based on anatomical and physiological data and other \nmodels as described in O'Reilly and McClelland (1994) and McClelland et al. (1995), but \nnot elaborated upon significantly here. \n\nThe Input layer activity pattern represents the state of the EC resulting from the presentation \nof a  given item.  We  assume that the hippocampus stores and retrieves memories by way \nof reduced representations in the EC, which have a correspondence with more elaborated \nrepresentations in other areas of cortex that is  developed via  long-term cortical  learning. \nWe further assume that there is a rough topology to the organization of EC, with different \ncortical areas and/or sub-areas represented by different slots,  which can be thought of as \nrepresenting different feature dimensions of the input (e.g_, color, font,  semantic features, \netc.).  Our EC has 36 slots with four units per slot; one unit per slot was active (with each \nunit representing a particular \"feature value\").  Input patterns were constructed from pro(cid:173)\ntotypes by randomly selecting different feature values for a random subset of slots.  There \nare two functionally distinct layers of the EC,  one which receives input from cortical ar(cid:173)\neas and projects into the hippocampus (superficial or ECin ),  and another which receives \nprojections from the CAl and projects back out to  the cortex (deep or ECout ).  While the \nrepresentations in these layers are probably different in their details,  we assume that they \nare functionally equivalent, and use the same representations across both for convenience. \nECin  projects to three areas of the hippocampus:  the dentate gyrus (DO), area CA3, and \narea CAL  The storage of the input pattern occurs through weight changes in the feedfor(cid:173)\nward and recurrent projections into the CA3, and the CA3  to CAl connections.  The CA3 \nand CAl contain the two primary representations of the input pattern, while the DO plays \nan important but secondary role as a pattern-separation enhancer for the CA3. \n\nThe  CA3  provides  the primary  sparse,  pattern-separated,  conjunctive representation de(cid:173)\nscribed above.  This is achieved by random, partial connectivity between the EC and CA3, \nand a high threshold for activation (i.e., sparseness), such that the few units which are acti(cid:173)\nvated in the CA3  (5% in our model) are those which have the most inputs from active EC \nunits.  The odds of a unit having such a high proportion of inputs from even two relatively \nsimilar EC  patterns  is  low,  resulting  in  pattern  separation (see  O'Reilly &  McClelland, \n\n\f76 \n\nR.  C.  O'Reilly, K.  A. Norman and 1.  L.  McClelland \n\n1994 for a much more detailed and precise treatment of this issue, and the role of the DO \nin facilitating pattern separation).  While these CA3  representations are useful for allowing \nrapid  learning without undue  interference,  the pattern-separation process  eliminates any \nsystematic relationship between the CA3  pattern and the original EC pattern that gave rise \nto it.  Thus, there must be some means of translating the CA3 pattern back into the language \nof the EC. The simple solution of directly associating the CA3 pattern with the correspond(cid:173)\ning EC pattern is problematic due to the interference caused by the relatively high activity \nlevels in the  EC  (around 15%,  and 25%  in our model).  For this reason,  we think that the \ntranslation is formed via the CAl, which (as a result of long-term learning) is  capable of \nexpanding EC representations into sparser patterns that are more easily linked to CA3, and \nthen mapping these sparser patterns back onto the EC. \n\nOur CAl  has  separate  representations of small combinations of slots (labeled columns); \ncolumns can be arbitrarily combined to reproduce any valid EC representation.  Thus, rep(cid:173)\nresentations  in  CAl  are  intermediate  between the  fully  conjunctive CA3,  and  the  fully \ncombinatorial EC.  This is  achieved in our model by training a single CAl  column of 32 \nunits with slightly less than 10% activity levels to be able to reproduce any combination of \npatterns over 3 ECin  slots (64 different combinations) in a corresponding set of3 ECout \nslots.  The resulting weights are replicated across columns covering the entire EC (see Fig(cid:173)\nure  la).  The cost of this scheme is that more CAl units are required (32 vs  12 per column \nin the EC), which is nonetheless consistent with the relatively greater expansion of this area \nrelative to other hippocampal areas as a function of cortical size. \n\nAfter learning, our model recollects studied items by simply reactivating the original CA3, \nCAl and ECout  patterns via facilitated weights.  With partial or noisy input patterns (and \nwith  interference),  these  weights  and  two  forms  of recurrence  (the  \"short  loop\"  within \nCA3,  and the  \"big loop\" out to  the  EC  and back through the  entire hippocampus) allow \nthe hippocampus to bootstrap its way into recalling the complete original pattern (pattern \ncompletion).  If the EC  input pattern corresponds to a nonstudied pattern, then the weights \nwill not have been facilitated for this particular activity pattern,  and the CAl  will  not be \nstrongly driven by the CA3.  Even if the ECin  activity pattern corresponds to two compo(cid:173)\nnents that were previously studied, but not together (see below), the conjunctive nature of \nthe CA3 representations will minimize the extent to which recall occurs. \n\nRecollection is operationalized as successful recall of the test probe.  This raises the basic \nproblem that the system needs to be able to distinguish between the EC out activation due to \nthe item input on ECin  (either directly or via the CAl), and that which is due to activation \ncoming  from  recall  in  the  CA3-CAl  pathway.  One  solution  to  this  problem,  which  is \nsuggested by autocorrelation histograms during reversible CA3  lesions (Mizumori et aI., \n1989),  is that the CA3  and CAl  are  1800  out of phase with respect to the theta rhythm. \nThus,  when  the  CA3  drives  the  CAl,  it  does  so  at a  point when  the  CAl  units  would \notherwise be silent, providing a means for distinguishing between EC and CA3 driven CA 1 \nactivation. We approximate something like this mechanism by simply turning off the ECin \ninputs to CAl during testing. We assess the quality of hippocampal recall by comparing the \nresulting ECout pattern with the ECin cue. The number of active units that match between \nECin  and ECout  (labeled C) indicates how much of the test probe was recollected.  The \nnumber of units that are active in EC out  but not in ECin  (labeled E) indicates the extent \nto which the model recollected an item other than the test probe. \n\n3  Activation and Learning Dynamics \n\nOur model is  implemented using the Leabra framework,  which provides a robust mecha(cid:173)\nnism for producing controlled levels of sparse activation in the presence of recurrent activa-\n\n\fA Hippocampal Model of Recognition Memory \n\n77 \n\ntion dynamics, and a simple, effective Hebbian learning rule (O'Reilly, 1996)1.  The activa(cid:173)\ntion function is a simple thresholded single-compartment neuron model with continuous-\nvalued spike rate output.  Membrane potential is updated by  dVd't(t)  =  T L:c gc (t)gc (Ec  -\nVm(t)),  with 3 channels (c) corresponding to:  e excitatory input;  lleak current; and i  in(cid:173)\nhibitory input.  Activation communicated to other cells is a simple thresholded function of \nthe membrane potential: Yj(t) = 1/ (1 + 'Y[v>n(:)-9J+)' As in the hippocampus (and cor(cid:173)\ntex), all principal weights (synaptic efficacies) are excitatory, while the local-circuit inhi(cid:173)\nbition controls positive feedback loops (i.e., preventing epileptiform activity) and produces \nsparse representations.  Leabra assumes that  the  inhibitory feedback has  an approximate \nset-point (i.e., strong activity creates compensatorially stronger inhibition, and vice-versa), \nresulting in roughly constant overall activity  levels,  with a firm upper bound.  Inhibitory \ncurrent is  given by gi  = g~+l + q(gr  - g~+l)' where 0  <  q  <  1 is  typically .25,  and \n8  L:. 9c9c(Ec-8) \nfor the UnIts With the k  th and k + 1 th highest excitatory mputs. \n9  = \nA simple, appropriately normalized Hebbian rule is used in Leabra:  f).wij  =  XiYj  - YjWij, \nwhich  can be seen as  computing the expected value  of the  sending unit's activity  condi(cid:173)\ntional on the receiver's activity (if treated like a binary variable active with probability Yj): \nWij  ~ (xiIYj}p'  This is essentially the same rule used in standard competitive learning or \nmixtures-of-Gaussians. \n\n. '  . \n\nct\u00b7 8-Ei \n\n. .  \n\n4  Interference and List-Length, Item Similarity \n\nHere,  we demonstrate that the hippocampal recollection system degrades with increasing \ninterference in a way that preserves its essential high-threshold, high-quality nature.  Fig(cid:173)\nure 2 shows the effects of list length and item similarity on our C and E  measures.  Only \nstudied items appear in the high C, low E  comer representing rich recollection.  As length \nand similarity increase,  interference results in decreased C  for studied items (without in(cid:173)\ncreased E), but critically there is no change in responding to new items. Interference in our \nmodel arises from the reduced but nevertheless extant overlap between representations in \nthe hippocampal system as a function of item similarity and number of items stored.  To the \nextent that increasing numbers of individual CA3 units are linked to mUltiple contradictory \nCAl representations, their contribution is reduced, and eventually recollection fails.  As for \nthe frequently obtained finding that decreased recollection of studied items is accompanied \nby an increase in overall false  alarms, we think this results from subjects being forced to \nrely more on the (less reliable) fallback familiarity mechanism. \n\n5  Conjunctivity and Associative Recognition \n\nNow,  we consider what happens when lures are  constructed by recombining elements of \nstudied patterns  (e.g.,  study ''window-reason'' and \"car-oyster\",  and test  with  \"window(cid:173)\noyster\").  One recent study found that participants are much more likely to claim to recol(cid:173)\nlect studied pairs than re-paired lures (Yonelinas,  1997). Furthermore, data from this study \nis consistent with the idea that re-paired lures sometimes trigger recollection of the stud(cid:173)\nied word pairs that were re-combined to  generate the lure;  when this happens (assuming \nthat each word occurred in only one pair),  the participant can confidently reject the  lure. \nOur simulation data is  consistent with these findings:  For studied word pairs,  the model \n(richly) recollected both pair components 86% of the time.  As for re-paired lures, both pair \ncomponents were never recalled together, but  16% of the time the model recollected one \nof the pair components,  along with the component that it was paired with at  study.  The \n\nI Note that the version of Leabra described here is an update to the cited version, which is currently \n\nbeing prepared for publication. \n\n\f78 \n\nR.  C.  O'Reilly, K.  A. Nonnan and 1. L  McClelland \n\nFigure 2:  Effects of list length  and similarity on  recollection perfonnance.  Responses can  be cat(cid:173)\negorized according  to  the  thresholds  shown,  producing  three regions:  rich  recollection (RR),  weak \nrecollection  (WR),  and misrecollection  (MR).  Increasing  list length  and  similarity  lead  to  less  rich \nrecollection of studied items (without increasing misrecollection for these items), and  do not signifi(cid:173)\ncantly affect the model's responding to lures. \n\nmodel responded in a  similar fashion to pairs consisting of one studied word and a  new \nword (never recollecting both pair components together, but recollecting the old item and \nthe item it was paired with at  study  13% of the time).  Word pairs consisting of two new \nitems failed to trigger recollection of even a single pair component.  Similar findings were \nobtained in our simulation of the (Hintzman, Curran, &  Oppy,  1992) experiment involving \nrecombinations of word and plurality cues. \n\n6  Discussion \n\nWhile the results presented above have dealt with the presentation of complete probe stim(cid:173)\nuli for recognition memory tests, our model is obviously capable of explaining cued recall \nand related phenomena such as source or context memory by virtue of its pattern comple(cid:173)\ntion abilities.  There are a  number of interesting issues that this  raises.  For example,  we \npredict that  successful  item  recollection will  be highly correlated with the  ability  to  re(cid:173)\ncall additional information from the study episode, since both rely on the same underlying \nmemory.  Further,  to the extent that elderly adults form less distinct encodings of stimuli \n(Rabinowitz &  Ackerman,  1982), this explains both their impaired recollection on recog(cid:173)\nnition tests (Parkin &  Walter,  1992) and their impaired memory for contextual (\"source\") \ndetails (Schacter et aI.,  1991). \n\nIn summary, existing mathematical models of recognition memory are most likely incorrect \nin assuming that recognition is performed with one memory system.  Global matching mod(cid:173)\nels may provide a good account of familiarity-based recognition, but they fail to account for \nthe contributions of recollection to recognition, as discussed above.  For example, global \nmatchil).g  models predict that lures which are similar to  studied items will always trigger \na stronger signal than dissimilar lures;  as  such,  these models can not account for the fact \nthat sometimes subjects can reject similar lures with high levels of confidence (due, in our \nmodel, to recollection ofa similar studied item; Brainerd, Reyna, & Kneer, 1995; Hintzman \net aI.,  1992). Further, global matching models confound the signal for the extent to which \nindividual components of the test probe were present at all during study, and signal for the \n\n\fA Hippocampal Model of Recognition Memory \n\n79 \n\nextent to which they occurred together.  We  believe that these  signals may be separable, \nwith recollection (implemented by the hippocampus) showing sensitivity to conjunctions \nof features, but not the occurrence of individual features, and familiarity (implemented by \ncortical regions) showing sensitivity to component occurrence but not co-occurence.  This \ndivision of labor is consistent with recent data showing that familiarity does not discrimi(cid:173)\nnate well between studied item pairs and lures constructed by conjoining items from two \ndifferent studied pairs (so long as the pairings are truly novel) (Yonelinas,  1997), and with \nthe point, set forth by (McClelland et aI.,  1995), that catastrophic interference would occur \nif rapid  learning (required to  learn feature  co-occurrences) took place in  the  neocortical \nstructures which generate the familiarity signal. \n\n7  References \n\nAggleton, J.  P., &  Shaw, C. (1996).  Amnesia and recognition memory:  are-analysis of psychometric \n\ndata.  Neuropsychologia, 34, 51. \n\nBrainerd,  C.  J.,  Reyna,  V.  F.,  &  Kneer,  R.  (1995).  False-recognition  reversal:  When  similarity  is \n\ndistinctive.  Journal of Memory and Language, 34,  157-185. \n\nChappell, M., &  Humphreys, M.  S.  (1994).  An auto-associative neural network  for sparse represen(cid:173)\n\ntations:  Analysis and application to models of recognition and cued recall.  Psychological Review, \n101,  103-128. \n\nGillund, G., & Shiffrin, R. M. (1984).  A retrieval model for both recognition and recall. Psychological \n\nReview, 91, 1-67. \n\nHintzman, D. L. (1988). Judgments of frequency and recognition memory in a multiple-trace memory \n\nmodel.  Psychological Review, 95, 528-551. \n\nHintzman,  D.  L.,  Curran,  T.,  &  Oppy,  B.  (1992).  Effects of similiarity and  repetition  on memory: \nRegistration without learning. Journal of Experimental Psychology:  Learning.  Memory.  and Cog(cid:173)\nnition,  18, 667-680. \n\nJacoby,  L.  L.,  Yonelinas,  A.  P.,  &  Jennings,  J.  M.  (1996).  The  relation  between  conscious  and \nIn  J.  D.  Cohen,  &  J.  W. \nunconscious  (automatic)  influences:  A  declaration  of independence. \nSchooler (Eds.), Scientific approaches to the question of consciousness (pp.  13-47). Hi1lsdale, NJ: \nLawrence Erbaum Associates. \n\nMcClelland,  J.  L.,  MCNaughton,  B.  L.,  &  O'Reilly, R.  C.  (1995).  Why  there  are  complementary \nlearning systems in  the hippocampus  and  neocortex:  Insights from  the successes and failures  of \nconnectionst models of learning and memory.  Psychological Review, 102,419-457. \n\nMiller, E. K., & Desimone, R. (1994).  Parallel neuronal mechanisms for short-term memory. Science, \n\n263, 520--522. \n\nMizumori, S. J. Y., McNaughton, B. L., Barnes, C. A., &  Fox, K. B. (1989).  Preserved spatial coding \nin hippocampal CAl  pyramidal cells during reversible suppression ofCA3c output:  Evidence for \npattern completion in hippocampus.  Journal of NeuroSCience,  9( II), 3915-3928. \n\nO'Reilly, R. C. (1996).  The leabra model of neural interactions and learning in the neocortex.  PhD \n\nthesis, Carnegie Mellon University, Pittsburgh, PA, USA. \n\nO'Reilly, R. C., &  McClelland, J.  L. (1994).  Hippocampal conjunctive encoding, storage, and recall: \n\nAvoiding a tradeoff.  Hippocampus, 4(6), 661-682. \n\nParkin, A. J., & Walter, B. M. (1992). Recollective experience, normal aging, and frontal dysfunction. \n\nPsychology and Aging,  7,290--298. \n\nRabinowitz, J.  C., &  Ackerman, B. P.  (1982).  General encoding of episodic events by elderly adults. \n\nIn  F.  I. M. C.  S.  Trehub (Ed.), Aging and cognitive processes. Plenum Publishing Corporation. \n\nSchacter,  D.  L.,  Kaszniak,  A. W.,  Kihlstrom,  J.  F.,  &  Valdiserri,  M.  (1991).  The relation  between \n\nsource memory and aging.  Psychology and Aging, 6, 559-568. \n\nYonelinas,  A.  P.  (1994).  Receiver-operating characteristics in recognition memory:  Evidence for  a \ndual-process model. Journal of Experimental Psychology:  Learning.  Memory.  and Cognition,  20, \n1341-1354. \n\nYonelinas,  A.  P.  (1997).  Recognition  memory  ROCs  for  item  and  associative  information:  The \n\ncontribution of recollection and familiarity.  Memory and Cognition,  25,747-763. \n\n\f", "award": [], "sourceid": 1415, "authors": [{"given_name": "Randall", "family_name": "O'Reilly", "institution": null}, {"given_name": "Kenneth", "family_name": "Norman", "institution": null}, {"given_name": "James", "family_name": "McClelland", "institution": null}]}