A Hippocampal Model of Recognition Memory 
Randall C. O'Reilly 
Department of Psychology 
University of Colorado at Boulder 
Campus Box 345 
Boulder, CO 80309-0345 
oreillypsych.colorado.edu 
Kenneth A. Norman 
Department of Psychology 
Harvard University 
33 Kirkland Street 
Cambridge, MA 02138 
nonnanwjh.harvard. edu 
James L. McClelland 
Department of Psychology and 
Center for the Neural Basis of Cognition 
Carnegie Mellon University 
Pittsburgh, PA 15213 
jlmcnbc.cmu.edu 
Abstract 
A rich body of data exists showing that recollection of specific infor- 
mation makes an important contribution to recognition memory, which 
is distinct from the contribution of familiarity, and is not adequately cap- 
tured by existing unitary memory models. Furthermore, neuropsycholog- 
ical evidence indicates that recollection is subserved by the hippocampus. 
We present a model, based largely on known features of hippocampal 
anatomy and physiology, that accounts for the following key character- 
istics of recollection: 1) false recollection is rare (i.e., participants rarely 
claim to recollect having studied nonstudied items), and 2) increasing in- 
terference leads to less recollection but apparently does not compromise 
the quality of recollection (i.e., the extent to which recollected informa- 
tion veridically reflects events that occurred at study). 
1 Introduction 
For nearly 50 years, memory researchers have known that our ability to remember specific 
past episodes depends critically on the hippocampus. In this paper, we describe our initial 
attempt to use a mechanistically explicit model of hippocampal function to explain a wide 
range of human memory data. 
Our understanding of hippocampal function from a computational and biological perspec- 
74 R. C. O'Reilly, K. A. Norman and J. L. McClelland 
tive is based on our prior work (McClelland, McNaughton, & O'Reilly, 1995; O'Reilly & 
McClelland, 1994). At the broadest level, we think that the hippocampus exists in part to 
provide a memory system which can learn arbitrary information rapidly without suffering 
undue amounts of interference. This memory system sits on top of, and works in conjunc- 
tion with, the neocortex, which learns slowly over many experiences, producing integrative 
representations of the relevant statistical features of the environment. The hippocampus ac- 
complishes rapid, relatively interference-free learning by using relatively non-overlapping 
(pattern separated) representations. Pattern separation occurs as a result of 1) the sparse- 
ness of hippocampal representations (relative to cortical representations), and 2) the fact 
that hippocampal units are sensitive to conjunctions of cortical features -- given two cor- 
tical patterns with 50% feature overlap, the probability that a particular conjunction of 
features will be present in both patterns is much less than 50%. 
We propose that the hippocampus produces a relatively high-threshold, high-quality recol- 
lective response to test items. The response is "high-threshold" in the sense that studied 
items sometimes trigger rich recollection (defined as "retrieval of most or all of the test 
probe's features from memory") but lures never trigger rich recollection. The response is 
"high-quality" in the sense that, most of the time, the recollection signal consists of part 
or all of a single studied pattern, as opposed to a blend of studied patterns. The high- 
threshold, high-quality nature of recollection can be explained in terms of the conjunc- 
tivity of hippocampal representations: Insofar as recollection is a function of whether the 
features of the test probe were encountered together at study, lures (which contain many 
novel feature conjunctions, even if their constituent features are familiar) are unlikely to 
trigger rich recollection; also, insofar as the hippocampus stores feature conjunctions (as 
opposed to individual features), features which appeared together at study are likely to 
appear together at test. Importantly, in accordance with dual-process accounts of recog- 
nition memory (Yonelinas, 1994; Jacoby, Yonelinas, & Jennings, 1996), we believe that 
hippocampally-driven recollection is not the sole contributor to recognition memory per- 
formance. Rather, extensive evidence exists that recollection is complemented by a "fall- 
back" familiarity signal which participants consult when rich recollection does not occur. 
The familiarity signal is mediated by as-yet unspecified areas (likely including the parahip- 
pocampal temporal cortex: Aggleton & Shaw, 1996; Miller & Desireone, 1994). 
Our account differs substantially from most other computational and mathematical models 
of recognition memory. Most of these models compute the "global match" between the 
test probe and stored memories (e.g., Hintzman, 1988; Gillund & Shiffrin, 1984); recollec- 
tion in these models involves computing a similarity-weighted average of stored memory 
patterns. In other memory models, recollection of an item depends critically on the extent 
to which the components of the item's representation were linked with that of the study 
context (e.g., Chappell & Humphreys, 1994). Critically, recollection in all of these models 
lacks the high-threshold, high-quality character of recollection in our model. This is most 
evident when we consider the effects of manipulations which increase interference (e.g., 
increasing the length of the study list, or increasing inter-item similarity). As interference 
increases, global matching models predict increasingly "blurry" recollection (reflecting the 
contribution of more items to the composite output vector), while the other models predict 
that false recollection of lures will increase. In contrast, our model predicts that increasing 
interference should lead to decreased correct recollection of studied test probes, but there 
should be no concomitant increase in "erroneous" types of recollection (i.e., recollection 
of details which mismatch studied test probes, or rich recollection of lures). This predic- 
tion is consistent with the recent finding that correct recollection of studied items decreases 
with increasing list length (Yonelinas, 1994). Lastly, although extant data certainly do not 
contradict the claim that the veridicality of recollection is robust to interference, we ac- 
knowledge that additional, focused experimentation is needed to definitively resolve this 
issue. 
A Hippocampal Model of Recognition Memory 75 
b) 
Figure 1: The model. a) Shows the areas and connectivity, and the corresponding columns within 
the Input, EC, and CA1 (see text). b) Shows an example activity pattern. Note the sparse activity in 
the DG and CA3, and intermediate sparseness of the CAl. 
2 Architecture and Overall Behavior 
Figure 1 shows a diagram of our model, which contains the basic anatomical regions of 
the hippocampal formation, as well as the entorhinal cortex (EC), which serves as the 
primary cortical input/output pathway for the hippocampus. The model as described below 
instantiates a series of hypotheses about the structure and function of the hippocampus and 
associated cortical areas, which are based on anatomical and physiological data and other 
models as described in O'Reilly and McClelland (1994) and McClelland et al. (1995), but 
not elaborated upon significantly here. 
The Input layer activity pattern represents the state of the EC resulting from the presentation 
of a given item. We assume that the hippocampus stores and retrieves memories by way 
of reduced representations in the EC, which have a correspondence with more elaborated 
representations in other areas of cortex that is developed via long-term conical learning. 
We further assume that there is a rough topology to the organization of EC, with different 
cortical areas and/or sub-areas represented by different slots, which can be thought of as 
representing different feature dimensions of the input (e.g., color, font, semantic features, 
etc.). Our EC has 36 slots with four units per slot; one unit per slot was active (with each 
unit representing a particular "feature value"). Input patterns were constructed from pro- 
totypes by randomly selecting different feature values for a random subset of slots. There 
are two functionally distinct layers of the EC, one which receives input from cortical ar- 
eas and projects into the hippocampus (superficial or ECi,), and another which receives 
projections from the CA1 and projects back out to the cortex (deep or ECoa, t). While the 
representations in these layers are probably different in their details, we assume that they 
are functionally equivalent, and use the same representations across both for convenience. 
JCin projects to three areas of the hippocampus: the dentate gyrus (DG), area CA3, and 
area CA 1. The storage of the input pattern occurs through weight changes in the feedfor- 
ward and recurrent projections into the CA3, and the CA3 to CA1 connections. The CA3 
and CA1 contain the two primary representations of the input pattern, while the DG plays 
an important but secondary role as a pattern-separation enhancer for the CA3. 
The CA3 provides the primary sparse, pattern-separated, conjunctive representation de- 
scribed above. This is achieved by random, partial connectivity between the EC and CA3, 
and a high threshold for activation (i.e., sparseness), such that the few units which are acti- 
vated in the CA3 (5% in our model) are those which have the most inputs from active EC 
units. The odds of a unit having such a high proportion of inputs from even two relatively 
similar EC patterns is low, resulting in pattern separation (see O'Reilly & McClelland, 
76 R. C. O'Reilly, K. A. Norman and J. L. McClelland 
1994 for a much more detailed and precise treatment of this issue, and the role of the DG 
in facilitating pattern separation). While these CA3 representations are useful for allowing 
rapid learning without undue interference, the pattern-separation process eliminates any 
systematic relationship between the CA3 pattern and the original EC pattern that gave rise 
to it. Thus, there must be some means of translating the CA3 pattern back into the language 
of the EC. The simple solution of directly associating the CA3 pattern with the correspond- 
ing EC pattern is problematic due to the interference caused by the relatively high activity 
levels in the EC (around 15%, and 25% in our model). For this reason, we think that the 
translation is formed via the CA1, which (as a result of long-term learning) is capable of 
expanding EC representations into sparser patterns that are more easily linked to CA3, and 
then mapping these sparser patterns back onto the EC. 
Our CA1 has separate representations of small combinations of slots (labeled columns); 
columns can be arbitrarily combined to reproduce any valid EC representation. Thus, rep- 
resentations in CA1 are intermediate between the fully conjunctive CA3, and the fully 
combinatorial EC. This is achieved in our model by training a single CA1 column of 32 
units with slightly less than 10% activity levels to be able to reproduce any combination of 
patterns over 3 EC'i, slots (64 different combinations) in a corresponding set of 3 EC'o,t 
slots. The resulting weights are replicated across columns coveting the entire EC (see Fig- 
ure !a). The cost of this scheme is that more CA1 units are required (32 vs 12 per column 
in the EC), which is nonetheless consistent with the relatively greater expansion of this area 
relative to other hippocampal areas as a function of cortical size. 
After learning, our model recollects studied items by simply reactivating the original CA3, 
CA1 and EC'ot patterns via facilitated weights. With partial or noisy input patterns (and 
with interference), these weights and two forms of recurrence (the "short loop" within 
CA3, and the "big loop" out to the EC and back through the entire hippocampus) allow 
the hippocampus to bootstrap its way into recalling the complete original pattern (pattern 
completion). If the EC input pattern corresponds to a nonstudied pattern, then the weights 
will not have been facilitated for this particular activity pattern, and the CAI will not be 
strongly driven by the CA3. Even if the EC'i, activity pattern corresponds to two compo- 
nents that were previously studied, but not together (see below), the conjunctive nature of 
the CA3 representations will minimize the extent to which recall occurs. 
Recollection is operationalized as successful recall of the test probe. This raises the basic 
problem that the system needs to be able to distinguish between the ECot activation due to 
the item input on ECi, (either directly or via the CA1), and that which is due to activation 
coming from recall in the CA3-CA1 pathway. One solution to this problem, which is 
suggested by autocorrelation histograms during reversible CA3 lesions (Mizumori et al., 
1989), is that the CA3 and CA1 are 180  out of phase with respect to the theta rhythm. 
Thus, when the CA3 drives the CA1, it does so at a point when the CA1 units would 
otherwise be silent, providing a means for distinguishing between EC and CA3 driven CA 1 
activation. We approximate something like this mechanism by simply turning off the ECin 
inputs to CA1 during testing. We assess the quality ofhippocampal recall by comparing the 
resulting EC'o,t pattern with the EUi, cue. The number of active units that match between 
EUi, and EUo,t (labeled C) indicates how much of the test probe was recollected. The 
number of units that are active in ECo,t but not in ECi, (labeled E) indicates the extent 
to which the model recollected an item other than the test probe. 
3 Activation and Learning Dynamics 
Our model is implemented using the Leabra flamework, which provides a robust mecha- 
nism for producing controlled levels of sparse activation in the presence of recurrent activa- 
A Hippocampal Model of Recognition Memory 77 
tion dynamics, and a simple, effective Hebbian learning rule (O'Reilly, 1996) . The activa- 
tion function is a simple threshtided single-compartment neuron model with continuous- 
valued spike rate output. Membrane potential is updated by 
dt = r Yc gc(t)(Ec - 
Vm(t)), with 3 channels (c) corresponding to: e excitatory input; I leak current; and i in- 
hibitory input. Activation communicated to other cells is a simple thresholded function of 
the membrane potential: yj(t) = 1/(1 + -[v,()-o]+ )' As in the hippocampus (and cor- 
tex), all principal weights (synaptic efficacies) are excitatory, while the local-circuit inhi- 
bition controls positive feedback loops (i.e., preventing epileptiform activity) and produces 
sparse representations. Leabra assumes that the inhibitory feedback has an approximate 
set-point (i.e., strong activity creates compensatorially stronger inhibition, and vice-versa), 
resulting in roughly constant overall activity levels, with a finn upper bound. Inhibitory 
current is given by gi = g+l + q(g - g+), where 0 < q < I is typically .25, and 
gO y' :.: 9''2(E-O) 
= e-E, for the units with the k th and k + 1 th highest excitatory inputs. 
A simple, appropriately normalized Hebbian rule is used in Leabra: Awij -- xiyj - yjw O, 
which can be seen as computing the expected value of the sending unit's activity condi- 
tional on the receiver's activity (if treated like a binary variable active with probability y): 
wij  (xi [y)p. This is essentially the same rule used in standard competitive learning or 
mixtures-of-Gaussians. 
4 Interference and List-Length, Item Similarity 
Here, we demonstrate that the hippocampal recollection system degrades with increasing 
interference in a way that preserves its essential high-threshold, high-quality nature. Fig- 
ure 2 shows the effects of list length and item similarity on our C and E measures. Only 
studied items appear in the high C, low E comer representing rich recollection. As length 
and similarity increase, interference results in decreased C for studied items (without in- 
creased E), but critically there is no change in responding to new items. Interference in our 
model arises from the reduced but nevertheless extant overlap between representations in 
the hippocampal system as a function of item similarity and number of items stored. To the 
extent that increasing numbers of individual CA3 units are linked to multiple contradictory 
CA 1 representations, their contribution is reduced, and eventually recollection fails. As for 
the frequently obtained finding that decreased recollection of studied items is accompanied 
by an increase in overall false alarms, we think this results from subjects being forced to 
rely more on the (less reliable) fallback familiarity mechanism. 
5 Conjunctivity and Associative Recognition 
Now, we consider what happens when lures are constructed by recombining elements of 
studied patterns (e.g., study "window-reason" and "car-oyster", and test with "window- 
oyster"). One recent study found that participants are much more likely to claim to recol- 
lect studied pairs than re-paired lures (Yonelinas, 1997). Furthermore, data from this study 
is consistent with the idea that re-paired lures sometimes trigger recollection of the stud- 
ied word pairs that were re-combined to generate the lure; when this happens (assuming 
that each word occurred in only one pair), the participant can confidently reject the lure. 
Our simulation data is consistent with these findings: For studied word pairs, the model 
(richly) recollected both pair components 86% of the time. As for re-paired lures, both pair 
components were never recalled together, but 16% of the time the model recollected one 
of the pair components, along with the component that it was paired with at study. The 
Note that the version of Leabra described here is an update to the cited version, which is currently 
being prepared for publication. 
78 R. C. O'Reilly, K. A. Norman and J. L. McClelland 
'co oooo o 0 8 o 08% 
cwgll 
Figure 2: Effects of list length and similarity on recollection performance. Responses can be cat- 
egorized according to the thresholds shown, producing three regions: rich recollection (RR), weak 
recollection (WR), and misrecollection (MR). Increasing list length and similarity lead to less rich 
recollection of studied items (without increasing misrecollection for these items), and do not signifi- 
cantly affect the model's responding to lures. 
model responded in a similar fashion to pairs consisting of one studied word and a new 
word (never recollecting both pair components together, but recollecting the old item and 
the item it was paired with at study 13% of the time). Word pairs consisting of two new 
items failed to trigger recollection of even a single pair component. Similar findings were 
obtained in our simulation of the (Hintzman, Curran, & Oppy, 1992) experiment involving 
recombinations of word and plurality cues. 
6 Discussion 
While the results presented above have dealt with the presentation of complete probe stim- 
uli for recognition memory tests, our model is obviously capable of explaining cued recall 
and related phenomena such as source or context memory by virtue of its pattern comple- 
tion abilities. There are a number of interesting issues that this raises. For example, we 
predict that successful item recollection will be highly correlated with the ability to re- 
call additional information from the study episode, since both rely on the same underlying 
memory. Further, to the extent that elderly adults form less distinct encodings of stimuli 
(Rabinowitz & Ackerman, 1982), this explains both their impaired recollection on recog- 
nition tests (Parkin & Walter, 1992) and their impaired memory for contextual ("source") 
details (Schacter et al., 1991). 
In summary, existing mathematical models of recognition memory are most likely incorrect 
in assuming that recognition is performed with one memory system. Global matching mod- 
els may provide a good account of familiarity-based recognition, but they fail to account for 
the contributions of recollection to recognition, as discussed above. For example, global 
matchh3. g models predict that lures which are similar to studied items will always trigger 
a stronger signal than dissimilar lures; as such, these models can not account for the fact 
that sometimes subjects can reject similar lures with high levels of confidence (due, in our 
model, to recollection of a similar studied item; Brainerd, Reyna, & Kneer, 1995; Hintzman 
et al., 1992). Further, global matching models confound the signal for the extent to which 
individual components of the test probe were present at all during study, and signal for the 
A Hippocampal Model of Recognition Memory 79 
extent to which they occurred together. We believe that these signals may be separable, 
with recollection (implemented by the hippocampus) showing sensitivity to conjunctions 
of features, but not the occurrence of individual features, and familiarity (implemented by 
cortical regions) showing sensitivity to component occurrence but not co-occurence. This 
division of labor is consistent with recent data showing that familiarity does not discrimi- 
nate well between studied item pairs and lures constructed by conjoining items from two 
different studied pairs (so long as the pairings are truly novel) (Yonelinas, 1997), and with 
the point, set forth by (McClelland et al., 1995), that catastrophic interference would occur 
if rapid learning (required to learn feature co-occurrences) took place in the neocortical 
structures which generate the familiarity signal. 
7 References 
Aggleton, J. P., & Shaw, C. (1996). Amnesia and recognition memory: a re-analysis of psychometric 
data. Neuropsychologia, 34, 51. 
Brainerd, C. J., Reyna, V. F., & Kneer, R. (1995). False-recognition reversal: When similarity is 
distinctive. Journal of Memory and Language, 34, 157-185. 
Chappell, M., & Humphreys, M. S. (1994). An auto-associative neural network for sparse represen- 
tations: Analysis and application to models of recognition and cued recall. Psychological Review, 
101, 103-128. 
Gillund, G., & Shiffrin, R. M. (1984). A retrieval model for both recognition and recall. Psychological 
Review, 91, 1-67. 
Hintzman, D. L. (1988). Judgments of frequency and recognition memory in a multiple-trace memory 
model. Psychological Review, 95, 528-551. 
Hintzman, D. L., Curran, T., & Oppy, B. (1992). Effects of similiarity and repetition on memory: 
Registration without leaming. Journal of Experimental Psychology: Learning, Memory, and Cog- 
nition, 18, 667-680. 
Jacoby, L. L., Yonelinas, A. P., & Jennings, J. M. (1996). The relation between conscious and 
unconscious (automatic) influences: A declaration of independence. In J. D. Cohen, & J. W. 
Schooler (Eds.), Scientific approaches to the question of consciousness (pp. 13-47). Hillsdale, N J: 
Lawrence Erbaum Associates. 
McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why them are complementary 
learning systems in the hippocampus and neocortex: Insights from the successes and failures of 
connectionst models of learning and memory. Psychological Review, 102, 419-457. 
Miller, E. K., & Desimone, R. (1994). Parallel neuronal mechanisms for short-term memory. Science, 
263, 520-522. 
Mizumori, S. J. Y., McNaughton, B. L., Bames, C. A., & Fox, K. B. (1989). Preserved spatial coding 
in hippocampal CA1 pyramidal cells during reversible suppression of CA3c output: Evidence for 
pattern completion in hippocampus. Journal of Neuroscience, 9(11), 3915-3928. 
O'Reilly, R. C. (1996). The leabra model of neural interactions and learning in the neocortex. PhD 
thesis, Camegie Mellon University, Pittsburgh, PA, USA. 
O'Reilly, R. C., & McClelland, J. L. (1994). Hippocampal conjunctive encoding, storage, and recall: 
Avoiding a tradeoff. Hippocampus, 4(6), 661-682. 
Parkin, A. J., & Walter, B. M. (1992). Recollective experience, normal aging, and frontal dysfunction. 
Psychology and Aging, 7, 290-298. 
Rabinowitz, J. C., & Ackerman, B. P. (1982). General encoding of episodic events by elderly adults. 
In F. I. M. C. S. Trehub (Ed.), Aging and cognitive processes. Plenum Publishing Corporation. 
Schacter, D. L., Kaszniak, A. W., Kihlstrom, J. F., & Valdiserri, M. (1991). The relation between 
source memory and aging. Psychology and Aging, 6, 559-568. 
Yonelinas, A. P. (1994). Receiver-operating characteristics in recognition memory: Evidence for a 
dual-process model. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 
1341-1354. 
Yonelinas, A. P. (1997). Recognition memory ROCs for item and associative information: The 
contribution of recollection and familiarity. Memory and Cognition, 25, 747-763. 
