Modeling Hippocampal and Neocortical Contributions to

Modeling Hippocampal and Neocortical Contributionsto Recognition Memory: A Complementary Learning

Systems Approach

Kenneth A. Norman and Randall C. O’Reilly

[email protected], [email protected]

in press, Psychological Review

29th August, 2002

This version may differ in minor ways from the final published version

Abstract

We present a computational neural network model of recognition memory based on the biological structures ofthe hippocampus and medial temporal lobe cortex (MTLC), which perform complementary learning functions. Thehippocampal component of the model contributes to recognition by recalling specific studied details. MTLC can notsupport recall, but it is possible to extract a scalar familiarity signal from MTLC that tracks how well the test itemmatches studied items. We present simulations that establish key qualitative differences in the operating characteris-tics of the hippocampal recall and MTLC familiarity signals, and we identify several manipulations (e.g., target-luresimilarity, interference) that differentially affect the two signals. We also use the model to address the stochastic rela-tionship between recall and familiarity (i.e., are they independent), and the effects of partial vs. complete hippocampallesions on recognition.

Contents

Introduction 4Dual-Process Controversies . . . . . . . . . . . 4Cognitive Neuroscience Approaches to

Recognition Memory . . . . . . . . . . . 5Summary: Combining the Approaches . . . . . 6Precis of Modeling Work . . . . . . . . . . . . 6

Complementary Learning Systems . . . . 6A Dual-Process Model of Recognition . . 7Part I: Basic Network Properties . . . . . 7Part II: Applications to Behavioral Phe-

nomena . . . . . . . . . . . . . . 8A Note on Terminology . . . . . . . . . . . . . 10Modeling Framework . . . . . . . . . . . . . . 10The Cortical Model . . . . . . . . . . . . . . . 10

Familiarity as Sharpening . . . . . . . . . 11The Hippocampal Model . . . . . . . . . . . . 12

The Hippocampal Recall Measure . . . . 13Simulation Methodology . . . . . . . . . . . . 13

Basic Methods . . . . . . . . . . . . . . 14Simulating YN and FC Testing . . . . . . 14

Part I: Basic Network Properties 15

Simulation 1: Pattern Separation 15

Simulation 2: Nature of the Underlying Distribu-tions 15Limits on Diagnosticity: High Average Overlap 16

Simulation 3: YN Related Lure Simulations 17Methods . . . . . . . . . . . . . . . . . . 17Results . . . . . . . . . . . . . . . . . . . 17

Sources of Variability 18Other Sources of Variability . . . . . . . . . . . 19Sources of Variability: Summary . . . . . . . . 19

Part II: Applications To Behavioral Phenomena 19

Simulation 4: Lure Relatedness and Test FormatInteractions 20Forced-Choice Testing and Covariance . . . . . 20

FC Simulations Using the Basic CorticalModel . . . . . . . . . . . . . . . 20

Cortical and Hippocampal SimulationsWith Encoding Variability . . . . 21

Methods . . . . . . . . . . . . . . . . . . 21Cortical FC Results . . . . . . . . . . . . 21Hippocampal FC Results . . . . . . . . . 21YN Results . . . . . . . . . . . . . . . . 21

Tests of the Model’s Predictions . . . . . . . . 22

Simulation 5: Associative Recognition and Sensi-tivity to Conjunctions 23Methods . . . . . . . . . . . . . . . . . . . . . 23YN Results . . . . . . . . . . . . . . . . . . . 23Effects of Test Format . . . . . . . . . . . . . . 24

Methods . . . . . . . . . . . . . . . . . . . . . 24Results . . . . . . . . . . . . . . . . . . . . . . 24Tests of the Model’s Predictions . . . . . . . . 25

Simulation 6: Interference and List Strength 26General Principles of Interference in our Models 26List Strength Simulations . . . . . . . . . . . . 26

Methods . . . . . . . . . . . . . . . . . . 27Results . . . . . . . . . . . . . . . . . . . 27

Interference in the Hippocampal Model . . . . 28Interference in the Cortical Model . . . . . . . 28

Differentiation . . . . . . . . . . . . . . . 29Boundary Conditions on the Null LSE . . 29

Tests of the Model’s Predictions . . . . . . . . 29Self-Report Measures . . . . . . . . . . . 30Lure Relatedness . . . . . . . . . . . . . 30

List Length Effects . . . . . . . . . . . . . . . 31Fast Weights . . . . . . . . . . . . . . . . 31

Simulation 7: The Combined Model and the In-dependence Assumption 32Combined Model Architecture . . . . . . . . . 32Effects of Encoding Variability . . . . . . . . . 33Interference Induced Decorrelation . . . . . . . 33

Simulation 8: Lesion Effects in the CombinedModel 34Partial Lesion Simulation Methods . . . . . . . 34Hippocampal Lesions . . . . . . . . . . . . . . 34MTLC Lesions . . . . . . . . . . . . . . . . . 35Relevant Data and Implications . . . . . . . . . 35

General Discussion 36Implications for Theories of How Recall Con-

tributes to Recognition . . . . . . . . . . 36Implications for Theories of How the Hip-

pocampus (vs. MTLC) Contributes toRecognition . . . . . . . . . . . . . . . . 37

Comparison with Abstract ComputationalModels of Recognition . . . . . . . . . . 37

Comparison with Other Neural Network Mod-els of Hippocampal and Cortical Contri-butions to Recognition . . . . . . . . . . 38Models of Hippocampus . . . . . . . . . 38Models of Neocortical Contributions to

Recognition Memory . . . . . . . 39Future Directions . . . . . . . . . . . . . . . . 40

Alternate Dependent Measures . . . . . . 41Conclusion . . . . . . . . . . . . . . . . . . . 41

Acknowledgments 41

Appendix A: Algorithm Details 41Pseudocode . . . . . . . . . . . . . . . . . . . 41Point Neuron Activation Function . . . . . . . 42k-Winners-Take-All Inhibition . . . . . . . . . 42Hebbian Learning . . . . . . . . . . . . . . . . 43Weight Contrast Enhancement . . . . . . . . . 43

Norman & O’Reilly 3

Appendix B: Hippocampal Model Details 43

Appendix C: Basic Parameters 44

References 45

4 Modeling Hippocampal and Neocortical Contributions to Recognition

Introduction

Memory can be subdivided according to func-tional categories (e.g., declarative vs. procedural mem-ory; Squire, 1992b; Cohen & Eichenbaum, 1993), andaccording to neural structures (e.g., hippocampally-dependent vs. non-hippocampally-dependent forms ofmemory). Various attempts have been made to alignthese functional and neural levels of analysis; for exam-ple, Squire (1992b) and others have argued that declar-ative memory depends on the medial temporal lobe,whereas procedural memory depends on other corticaland subcortical structures. Recently, we and our col-leagues have set forth a computationally explicit the-ory of how hippocampus and neocortex contribute tolearning and memory (the Complementary Learning Sys-tems model; O’Reilly & Rudy, 2001; McClelland, Mc-Naughton, & O’Reilly, 1995). In this paper, we advancethe Complementary Learning Systems model by usingit to provide a comprehensive treatment of recognitionmemory performance.

In the Introduction, we describe two questions thathave proved challenging for math modeling and cog-nitive neuroscience approaches to recognition, respec-tively: In the math modeling literature, there has beenconsiderable controversy regarding how to characterizethe contribution of recall (vs. familiarity) to recogni-tion memory; in the cognitive neuroscience literature, re-searchers have debated how the hippocampus (vs. medialtemporal lobe cortex) contributes to recognition. Then,we show how our modeling approach, which is jointlyconstrained by behavioral and neuroscientific data, canhelp resolve these controversies

Dual-Process Controversies

Recognition memory refers to the process of iden-tifying stimuli or situations as having been experiencedbefore, for example when you recognize a person youknow in a crowd of strangers. Recognition can be com-pared with various forms of recall memory where spe-cific content information is retrieved from memory andproduced as a response; recognition does not require re-call of specific details (e.g., one can recognize a personas being familiar without being able to recall who ex-actly they are or where you know them from). Never-theless, recognition can certainly benefit from recall ofspecific information — if you can recall that the personyou recognized at the supermarket is in fact your veteri-narian, that reinforces your feeling that you actually doknow this person. Theories that posit that recognition issupported by specific recall as well as nonspecific feel-ings of familiarity are called dual-process theories (seeYonelinas, 2002 for a thorough review of these theories).

While it is obvious that recall can (in principle) con-tribute to recognition judgments, the notion that recallroutinely contributes to item recognition performance isquite controversial. Practically all extant math models ofrecognition consist of a unitary familiarity process thatindexes in a holistic fashion the “global match” betweenthe test probe and all of the items stored in memory (e.g.,Hintzman, 1988; Gillund & Shiffrin, 1984; Humphreys,Bain, & Pike, 1989). These familiarity-only models canexplain a very wide range of recognition findings (forreviews, see Clark & Gronlund, 1996; Raaijmakers &Shiffrin, 1992; Ratcliff & McKoon, 2000) — even find-ings, that at first glance, appear to require a recall pro-cess (see, e.g., McClelland & Chappell, 1998). Further-more, the relatively small number of findings that cannot be explained using standard global matching modelstend to come from specialized paradigms like Jacoby’sprocess dissociation procedure (Jacoby, 1991; see Rat-cliff, Van Zandt, & McKoon, 1995 for discussion ofwhen global matching models can and can not accountfor process-dissociation data). As such, it is always pos-sible to treat these findings as special cases that have littlerelevance to performance on standard item-recognitiontests.

Another issue that has hindered the acceptance ofdual-process models is the difficulty inherent in measur-ing the separate contributions of recall and familiarity.Several techniques has been devised for quantitativelyestimating how recall and familiarity are contributing torecognition performance (ROC analysis, independenceremember-know, and process dissociation; see Yoneli-nas, 2001b, 2002 for review and discussion), but all ofthese techniques rely on a core set of controversial as-sumptions; for example, they all assume that recall andfamiliarity are stochastically independent. There are rea-sons to believe that the independence assumption maynot always be valid (see, e.g., Curran & Hintzman, 1995).Furthermore, there is no way to test this assumption us-ing behavioral data alone because of chicken-and-eggproblems (i.e., one needs to measure familiarity to assessits independence from recall, but one needs to assumeindependence to measure familiarity).

These chicken-and-egg problems have led to a rift be-tween math modelers and other memory researchers. Onthe empirical side, there is now a vast body of data onrecall and familiarity, gathered using measurement tech-niques that assume (among other things) independence— these data could potentially be used to constrain dual-process models. However, on the theoretical side, mod-elers are not making use of these data because of reason-able concerns about the validity of the assumptions usedto collect these data, and because single-process modelshave been quite successful at explaining recognition data(so why bother with more complex dual-process mod-


HIPPOCAMPUS

ENTORHINAL CORTEX

PERIRHINAL CORTEX

PARAHIPPOCAMPAL CORTEX

NEOCORTICAL ASSOCIATION AREAS (frontal, temporal, & parietal lobes)

MTLC

Figure 1: Schematic box diagram of neocortex, MTLC, andhippocampus. Medial temporal lobe cortex serves as the inter-face between neocortex and the hippocampus. Medial tempo-ral lobe cortex is located at the very top of the cortical pro-cessing hierarchy — it receives highly processed outputs ofdomain-specific cortical modules, integrates these outputs, andpasses them on to the hippocampus; it also receives output fromthe hippocampus, and passes this activation back to domain-specific cortical modules via feedback connections.

els). To resolve this impasse, we need some other sourceof evidence that we can use to specify the propertiesof recall and familiarity, other than the aforementionedmeasurement techniques.

Cognitive Neuroscience Approaches to Recog-nition Memory

Just as controversies exist in the math modeling lit-erature regarding the contribution of recall to recogni-tion memory, parallel controversies exist in the cognitiveneuroscience literature regarding the contribution of thehippocampus to recognition memory.

Researchers have long known that the medial tempo-ral region of the brain is important for recognition mem-ory. Patients with medial temporal lobe lesions encom-passing both the hippocampus and the medial temporallobe cortices (perirhinal, entorhinal, and parahippocam-pal cortices, which we refer to jointly as MTLC) typicallyshow impaired recall and recognition but intact perfor-mance on other memory tests (e.g., perceptual priming,skill learning; see Squire, 1992a for a review).

The finding of impaired recall and recognition in me-dial temporal amnesics is the basis for several influen-tial taxonomies of memory. Most prominently, Squire,Eichenbaum, Cohen, and others have argued that themedial temporal lobes implement a declarative mem-

ory system, which supports recall and recognition, andthat other brain structures support procedural memory(e.g., perceptual priming, motor skill learning; Squire,1992b, 1987; Cohen & Eichenbaum, 1993; Cohen, Pol-drack, & Eichenbaum, 1997; Eichenbaum, 2000). Re-searchers have argued that the medial temporal region isimportant for declarative memory because it is locatedat the top of the cortical hierarchy and therefore is ide-ally positioned to associate aspects of the current episodethat are being processed in domain-specific cortical mod-ules (e.g., Mishkin, Suzuki, Gadian, & Vargha-Khadem,1997; Mishkin, Vargha-Khadem, & Gadian, 1998). SeeFigure 1 for a schematic diagram of how hippocampus,MTLC, and neocortex are connected.

While the basic declarative memory framework iswidely accepted, attempts to tease apart the contribu-tions of different medial temporal structures have beenmore controversial. There is widespread agreementthat the hippocampus is critical for recall — focal hip-pocampal lesions lead to severely impaired recall perfor-mance. However, the data are much less clear regard-ing effects of focal hippocampal damage on recognition.Some studies have found roughly equal impairmentsin recall and recognition (e.g., Reed & Squire, 1997;Manns & Squire, 1999; Rempel-Clower, Zola, & Ama-ral, 1996; Zola-Morgan, Squire, & Amaral, 1986; Reed,Hamann, Stefanacci, & Squire, 1997), whereas otherstudies have found relatively spared recognition after fo-cal hippocampal lesions (e.g., Vargha-Khadem, Gadian,Watkins, Connelly, Van Paesschen, & Mishkin, 1997;Holdstock, Mayes, Roberts, Cezayirli, Isaac, O’Reilly, &Norman, 2002; Mayes, Holdstock, Isaac, Hunkin, &Roberts, 2002). The monkey literature parallels the hu-man literature — some studies have found relatively in-tact recognition (indexed using the delayed nonmatch-to-sample test) following focal hippocampal damage(e.g., Murray & Mishkin, 1998) whereas others havefound impaired recognition (e.g., Zola, Squire, Teng, Se-fanacci, Buffalo, & Clark, 2000; Beason-Held, Rosene,Killiany, & Moss, 1999). Spared recognition follow-ing hippocampal lesions depends critically on MTLC —whereas recognition is sometimes spared by focal hip-pocampal lesions, it is never spared after lesions that en-compass both MTLC and the hippocampus (e.g., Aggle-ton & Shaw, 1996).

Aggleton and Brown (1999) have tried to frame thedifference between hippocampal and cortical contribu-tions in terms of dual-process models of recognition. Ac-cording to this view:

� The hippocampus supports recall.

� The MTLC can support some degree of (familiarity-based) recognition on its own.


This framework captures, at a gross level, how hip-pocampal damage affects memory, but it is too vagueto be useful in explaining the considerable variabilitythat exists across patients and tests in how hippocam-pal damage affects recognition. Just as some hippocam-pal patients show more of a recognition deficit than oth-ers, some studies have found (within individual patients)greater impairment on some recognition tests than others(e.g., Holdstock et al., 2002). In the absence of furtherspecification of the hippocampal contribution (and howit differs from the contribution of MTLC), it is not possi-ble to proactively determine whether recognition will beimpaired or spared in a particular patient/test.

Aggleton & Brown attempt to flesh out their theoryby arguing that MTLC familiarity can support recog-nition of individual items, but memory for new asso-ciations between items depends on hippocampal recall(Sutherland & Rudy, 1989; Eichenbaum, Otto, & Co-hen, 1994 make similar claims). This view implies thatitem recognition should be intact but the ability to formnew associations should be impaired after focal hip-pocampal damage. However, Andrew Mayes and col-leagues have found that hippocampally-lesioned patientYR, who shows intact performance on some item recog-nition tests (e.g., Mayes et al., 2002), shows impairedperformance on other item recognition tests (e.g., Hold-stock et al., 2002), and spared performance on some teststhat require participants to associate previously unrelatedstimuli (e.g., the words “window” and “reason”; Mayes,Isaac, Downes, Holdstock, Hunkin, Montaldi, MacDon-ald, Cezayirli, & Roberts, 2001).

In summary: It is becoming increasingly evident thatthe effects of hippocampal damage are complex. Thereappears to be some functional specialization in the me-dial temporal lobe, but the simple dichotomies that havebeen proposed to explain this specialization are either toovague (recall vs. familiarity) or they are inconsistent withrecently acquired data (item memory vs. memory for newassociations).

Summary: Combining the Approaches

What should be clear at this point is that themath modeling and cognitive neuroscience approachesto recognition memory would greatly benefit from in-creased crosstalk: Math modeling approaches need anew source of constraints before they can fully explorehow recall contributes to recognition; and cognitive neu-roscience approaches need a new, more mechanisticallysophisticated vocabulary for talking about the roles ofdifferent brain structures in order to adequately charac-terize differences in how MTLC vs. hippocampus con-tribute to recognition.

The goal of our research is to achieve a synthesis of

these two approaches, by constructing a computationalmodel of recognition memory in which there is a trans-parent mapping between different parts of the modeland different subregions of hippocampus and MTLC.This mapping makes it possible to address neuroscien-tific findings using the model. For example, to predictthe effects of a particular kind of hippocampal lesion, wecan “lesion” the corresponding region of the model. Bybringing a wide range of constraints — both purely be-havioral and neuroscientific — to bear on a common setof mechanisms, we hope to achieve a more detailed un-derstanding of how recognition memory works.

Our model falls clearly in the dual-process tradition,insofar as we posit that the hippocampus and MTLCcontribute signals with distinct properties to recognitionmemory. The key, differentiating property of our modelis that — in our model — differences in the two sig-nals are grounded in architectural differences betweenthe hippocampus and MTLC; because most of these ar-chitectural differences fall along a continuum, it followsthat differences in the two signals will be more nuancedthan the dichotomies (“item vs. associative”) discussedabove.

Precis of Modeling Work

In this section, we summarize the major claims ofthe paper, highlighted in bold font, with pointers to loca-tions in the main text where these issues are discussed ingreater detail.

Complementary Learning Systems� The hippocampus is specialized for rapidly

memorizing specific events.

� The neocortex is specialized for slowly learningabout the statistical regularities of the environ-ment.

These are the central claims of the ComplementaryLearning Systems (CLS) framework (O’Reilly &Rudy, 2001; McClelland et al., 1995). Accordingto this framework, the two goals of memorizingspecific events and learning about statistical regu-larities are in direct conflict when implemented inneural networks; thus, to avoid making a tradeoff,we have evolved specialized neural systems for per-forming these tasks (see Marr, 1971; O’Keefe &Nadel, 1978; Sherry & Schacter, 1987 for similarideas, and Carpenter & Grossberg, 1993 for a con-trasting perspective).

� The hippocampus assigns distinct (pattern sepa-rated) representations to stimuli, thereby allowingit to learn rapidly without suffering catastrophic in-terference.


� In contrast, cortex assigns similar representa-tions to similar stimuli; use of overlapping rep-resentations allows cortex to represent the sharedstructure of events, and therefore makes it possiblefor cortex to generalize to novel stimuli as a func-tion of their similarity to previously encounteredstimuli.

A Dual-Process Model of Recognition� We have developed hippocampal and neocortical

models that incorporate key aspects of the biol-ogy of these structures, and instantiate the com-plementary learning systems principles outlinedabove (O’Reilly & Rudy, 2001; McClelland et al.,1995; O’Reilly & McClelland, 1994; O’Reilly &Munakata, 2000; O’Reilly, Norman, & McClelland,1998). (see Modeling Framework, p. 10).

� The neocortical model supports familiarity judg-ments based on the sharpness of representationsin MTLC. Competitive self-organization (arisingfrom Hebbian learning and inhibitory competition)causes stimulus representations to become sharperover repeated exposures (i.e., activity is concen-trated in smaller number of units; see The CorticalModel, p. 10).

– However, the neocortex can not support re-call of details from specific events, owing toits relatively low learning rate and its inabilityto sufficiently differentiate the representationsof different events.

– Familiarity is measured by the activity ofthe top

�most active units, though other

measures are possible and we have exploredsome of them, with similar results (see alsoAlternate Dependent Measures, p. 41).

� The hippocampus supports recall of previouslyencountered stimuli. When stimuli are presentedat study, the hippocampus assigns relatively non-overlapping (pattern-separated) representations tothese items in region CA3. Active units in CA3 arelinked to one another, and to a copy of the inputpattern (via region CA1). At test, presentation of apartial version of a studied pattern leads to recon-struction (pattern completion) of the original CA3representation and, through this, reconstruction ofthe entire studied pattern on the output layer (seeThe Hippocampal Model, p. 12).

– The hippocampal model is applied to recog-nition by computing the degree of matchbetween retrieved information and the re-call cue, minus the amount of mismatch.

Recall of matching information is evidencethat the cue was studied, and recall of infor-mation that mismatches the retrieval cue is ev-idence that the cue was not studied.

Part I: Basic Network Properties� The hippocampal and neocortical signals have

distinct operating characteristics. This differenceis largely due to differences in the two networks’ability to carry out pattern separation (see Simula-tion 1: Pattern Separation, p. 15, and Simulation 2:Nature of the Underlying Distributions, p. 15).

– MTLC familiarity functions as a standardsignal detection process. The familiarity dis-tributions for studied items and lures are Gaus-sian and overlap extensively. The distributionsoverlap because the representations of studieditems and lures overlap in MTLC. Some lures,by chance, have representations that overlapvery strongly with the representations of stud-ied items in MTLC and — as a result — theselures trigger a strong familiarity signal.

– In contrast, the hippocampal recall signal ismore diagnostic: Studied items sometimestrigger strong matching recall, but mostlures do not trigger any hippocampal re-call because of the hippocampus’ tendency toassign distinct representations to stimuli (re-gardless of similarity). As such, high levels ofmatching recall strongly indicate that an itemwas studied.

� There are boundary conditions on thediagnosticity of the hippocampal recallsignal. When the average amount ofof overlap between stimuli is high, hip-pocampal pattern separation mechanismsbreak down, resulting in strong recall ofshared, prototypical features (even in re-sponse to lures) and poor recall of fea-tures that are unique to particular studieditems.

� The difference in operating characteristics ismost evident on yes-no related lure recognitiontests, where lures are similar to studied items, butstudied items are dissimilar to one another. Thehippocampus strongly outperforms cortex on thesetests (see Simulation 3: YN Related Lure Simula-tions, p. 17).

– As target-lure similarity increases, lurefamiliarity increases steadily, but (up toa point) hippocampal pattern separationworks to keep lure recall at floor.


– Very similar lures trigger recall, but whenthis happens the lure can often be rejecteddue to mismatch between retrieved infor-mation and the recall cue. For example, ifa participant studies “rats” and is tested with“rat”, they might recall having studied “rats”(and reject “rat” on this basis).

� On related lure recognition tests, thepresence of any mismatching recall ishighly diagnostic of the item being alure. As such, we posit that participantspursue a recall-to-reject strategy on suchtests, whereby items are given a confident“new” response if they trigger any mis-matching recall.

� In the Sources of Variability section (p. 18) we dis-cuss different kinds of variability (e.g., variabilitydue to random initial weight settings, encoding vari-ability) and their implications for recognition per-formance.

Part II: Applications to Behavioral Phenomena

The effects of target-lure similarity on the twomodels interact with test format (yes-no versusforced-choice) (see Simulation 4: Lure Relatedness andTest Format Interactions, p. 20).

� The hippocampal advantage for related lures(observed with “yes-no” tests, as discussedabove) is mitigated by giving participants aforced choice between studied items and corre-sponding related lures at test (i.e., test A vs. A’,B vs. B’, where A’ and B’ are lures related to Aand B, respectively; see Forced-Choice Testing andCovariance, p. 20).

– Cortical performance benefits from thehigh degree of covariance in the familiarityscores triggered by studied items and corre-sponding related lures, which makes smallfamiliarity differences highly reliable.

– In contrast, use of this test format may ac-tually harm hippocampal performance, rel-ative to other test formats. On forced-choicetests with non-corresponding lures (test A vs.B’, B vs. A’), participants have two inde-pendent chances to respond correctly: On tri-als where participants fail to recall the stud-ied item (A) , they can still respond correctlyif the lure (B’) triggers mismatching recall(and is rejected on this basis). Use of cor-responding lures deprives participants of this“second chance”, insofar as recall triggered

by studied items and corresponding lures willbe highly correlated — if A does not triggerrecall, A’ will not trigger recall-to-reject (seeHippocampal FC Results, p. 21).

� The predicted interaction between target-luresimilarity and test format was obtained in exper-iments comparing a focal hippocampal amnesicwith control participants. The model predicts thathippocampally-lesioned patients, who are relyingexclusively on MTLC familiarity, should performpoorly relative to controls on standard yes-no recog-nition tests with related lures, but they should per-form relatively well on forced-choice tests with cor-responding related lures, and they should performwell relative to controls on both yes-no and forced-choice tests that use unrelated lures (insofar as bothnetworks discriminate well between studied itemsand unrelated lures). Holdstock et al. (2002; Mayeset al., 2002) found exactly this pattern of results inhippocampally-lesioned patient YR (see Tests of theModel’s Predictions, p. 22).

Associative recognition tests (i.e., study A-B, C-D;test with studied pairs and recombined pairs like A-D), can be viewed as a special case of the related lureparadigm discussed above (see Simulation 5: Associa-tive Recognition and Sensitivity to Conjunctions, p. 23).

� The hippocampal model outperforms the corti-cal model on yes-no associative recognition tests.As in the related lure simulations, the hippocampaladvantage is due to hippocampal pattern separation,and the hippocampus’ ability to carry out recall-to-reject.

– Even though the cortical model is worse atassociative recognition than the hippocam-pus, the cortical model still performs wellabove chance. This finding shows that ourcortical model has some (albeit reduced) sen-sitivity to whether items occurred together atstudy. This contrasts with other models (e.g.,Rudy & Sutherland, 1989) that posit that cor-tex supports memory for individual featuresbut not novel feature conjunctions.

� Giving participants a forced choice betweenstudied pairs and overlapping re-paired lures(test A-B vs. A-D) mitigates the hippocampal ad-vantage for associative recognition. This mirrorsthe finding that “forced choice corresponding” test-ing mitigates the hippocampal advantage for relatedlures.


� We present data from previously published studiesthat support the model’s predictions regarding howfocal hippocampal damage should affect associativerecognition performance, as a function of test for-mat (see Tests of the Model’s Predictions, p. 25).

Hippocampally- and cortically-driven recognitioncan be differentially affected by interference manipu-lations: Increasing list strength impairs discrimina-tion of studied items and lures in the hippocampalmodel, but does not impair discrimination based onMTLC familiarity. The list strength paradigm mea-sures how repeated study of a set of interference itemsaffects participants’ ability to discriminate between non-strengthened (but studied) target items and lures (seeSimulation 6: Interference and List Strength, p. 26).

� In both models, the overall effect of interfer-ence is to decrease weights to discriminative fea-tures of studied items and lures, and to increaseweights to prototypical features (which are sharedby a high proportion of items in the item set).

� The hippocampal model predicts a list strengtheffect because increasing list strength reduces re-call of discriminative features of studied items,and lure recall is already at floor. Increasing liststrength therefore has the effect of pushing the stud-ied and lure recall distributions together (reducingdiscriminability).

� The neocortical model predicts a null liststrength effect because the familiarity signaltriggered by lures has room to decrease as a func-tion of interference. Increasing list strength re-duces responding to (the discriminative featuresof) both studied items and lures, but the averagedifference in studied and lure familiarity does notdecrease. As such, discriminability does not suffer.

– In the neocortical model, lure familiarityinitially decreases more than studied-itemfamiliarity as a function of list strength,so the studied-lure gap in familiarity actuallywidens slightly with increasing interference.

– The widening of the studied-lure gap canbe explained in terms of differentiation —studying an item makes its representationoverlap less with the representations ofother, interfering items (Shiffrin, Ratcliff, &Clark, 1990). As such, studied items sufferless interference than lures.

– There are limits on the null list strength ef-fect for MTLC familiarity. With high lev-els of interference item strengthening and/or

high input overlap, the neocortical model’ssensitivity to discriminative features of stud-ied items and lures approaches floor, and thestudied and lure familiarity distributions startto converge (resulting in decreased discrim-inability).

� Data from two recent list strength experimentsprovide support for the model’s prediction thatlist strength should affect recall-based discrimina-tion but not familiarity-based discrimination (Nor-man, in press) (see Testing the Model’s Predictions,p. 29).

� We also discuss ways of modifying the learningrule to accommodate the finding that increasing listlength (i.e., adding new items to the study list)hurts recognition sensitivity more than increasinglist strength (Murnane & Shiffrin, 1991a; see ListLength Effects, p. 31).

The extent to which the neocortical and hip-pocampal recognition signals are correlated varies indifferent conditions. These issues are explored usinga more realistic “combined model” where the corticalnetwork responsible for computing familiarity serves asthe input to the hippocampal network (see Simulation 7:The Combined Model and the Independence Assumption,p. 32).

� Variability in how well items are learned at study(encoding variability) bolsters the correlation be-tween recall and familiarity signals, as was pos-tulated by Curran and Hintzman (1995) and others.

� Increasing interference reduces the recall-familiarity correlation for studied items. Thisoccurs because interference pushes raw recall andfamiliarity scores in different directions (increasinginterference reduces recall, but asymptoticallyit boosts familiarity by boosting the model’sresponding to shared, prototypical item features).

� In summary, recall and familiarity can be in-dependent when there is enough interference tocounteract the effects of encoding variability.

Partial hippocampal lesions can lead to worseoverall recognition performance than complete le-sions (see Simulation 8: Lesion Effects in the CombinedModel, p. 34).

� In the hippocampal model, partial lesions causepattern separation failure, which sharply re-duces the diagnosticity of the recall signal.Recognition performance suffers because the


noisy recall signal drowns out useful informationthat is present in the familiarity signal. Movingfrom a partial hippocampal lesion to a complete le-sion improves performance by removing this sourceof noise.

� In contrast, increasing MTLC lesion size in themodel leads to a monotonic decrease in perfor-mance. This occurs because MTLC lesions directlyimpair familiarity-based discrimination, and indi-rectly impair recall-based discrimination (becauseMTLC serves as the input to the hippocampus).

� These results are consistent with a recent meta-analysis of the lesion data, showing a negative cor-relation between recognition impairment and hip-pocampal lesion size, and a positive correlation be-tween recognition impairment and MTLC lesionsize (Baxter & Murray, 2001b).

A Note on Terminology

We use the terms “recall” and “familiarity” to de-scribe the respective contributions of the hippocampusand MTLC to recognition memory because these termsare heuristically useful. The hippocampal contribution torecognition is “recall” insofar as it involves retrieval ofspecific studied details. We use “familiarity” to describethe MTLC signal because it adheres to the definitionof familiarity set forth by Hintzman (1988; Gillund &Shiffrin, 1984) and others, i.e., it is a scalar that tracksthe global match or similarity of the test probe to studieditems.

However, we realize that the terms “recall” and “fa-miliarity” come with a substantial amount of theoreticalbaggage. Over the years, researchers have made a verylarge number of claims regarding properties of recall andfamiliarity; for example, Yonelinas has argued that re-call is a high-threshold process (e.g., Yonelinas, 2001a),and Mandler (1980) and others (e.g., Aggleton & Brown,1999) have argued that familiarity reflects memory forindividual items, apart from their associations with otheritems and contexts. By linking the hippocampal contri-bution with recall and the MTLC contribution with fa-miliarity, we do not mean to say that all of the various(and sometimes contradictory) properties that have beenascribed to recall and familiarity over the years apply tothe hippocampal and MTLC contributions, respectively.Just the opposite: We hope to “re-define” the propertiesof recall and familiarity using neurobiological data on theproperties of the hippocampus and MTLC. In the paper,we systematically delineate how the CLS model’s claimsabout hippocampal recall and MTLC familiarity deviatefrom claims made by existing dual-process theories.

Modeling Framework

Both the hippocampal and neocortical networks uti-lize the Hebbian component of O’Reilly’s Leabra al-gorithm (O’Reilly & Munakata, 2000; O’Reilly, 1998,1996; the full version of Leabra also incorporates error-driven learning, but error-driven learning was turned offin the simulations reported here). The algorithm we usedincorporates several widely-accepted characteristics ofneural computation, including Hebbian LTP/LTD (long-term potentiation/depression) and inhibitory competitionbetween neurons, that were first brought together byGrossberg (1976) (for more information on these mech-anisms see also Kohonen, 1977; Oja, 1982; Rumelhart &Zipser, 1986; Kanerva, 1988; Minai & Levy, 1994). Inour model, LTP is implemented by strengthening theconnection (weight) between two units when both thesending and receiving units are active together; LTD isimplemented by weakening the connection between twounits when the receiving unit is active, but the sendingunit is not (heterosynaptic LTD). Inhibitory competitionis implemented using a k-winners-take-all (kWTA) algo-rithm, which sets the amount of inhibition for a givenlayer such that at most k units are strongly active. Al-though the kWTA rule sets a firm limit on the number ofunits that show strong (> .25) activity, there is still con-siderable flexibility in the overall distribution of activityacross units in a layer. This is important for our discus-sion of “sharpening” in the cortical model, below. Keyaspects of the algorithm are summarized in Appendix A.

The Cortical Model

The cortical network is composed of two layers, in-put (which corresponds to cortical areas that feed intoMTLC), and hidden (corresponding to MTLC) (Fig-ure 2). The basic function of the model is for the hiddenlayer to encode regularities that are present in the inputlayer; this is achieved through the Hebbian learning rule.To capture the idea that the input layer represents manydifferent cortical areas, it consists of 24 10-unit slots,with 1 unit out of 10 active in each slot. A useful way tothink of slots is that different slots correspond to differ-ent feature dimensions (e.g., color or shape), and differ-ent units within a slot correspond to different, mutuallyexclusive features along that dimension (e.g., shapes: cir-cle, square, triangle). The hidden (MTLC) layer consistsof 1920 units, with 10% activity (i.e., 192 of these unitsare active on average for a given input). The input layeris connected to the MTLC layer via a partial feedforwardprojection where each MTLC unit receives connectionsfrom 25% of the input units. When items are presented atstudy, these connections are modified via Hebbian learn-ing.

Input patterns were constructed from prototypes by


Hidden (MTLC)

Input ( �� er-level ne �� ex)

Figure 2: Diagram of the cortical network. The corticalnetwork consists of two layers, an input layer (correspondingto “lower” cortical regions that feed into MTLC) and a hid-den layer (corresponding to MTLC). Units in the hidden layercompete to encode (via Hebbian learning) regularities that arepresent in the input layer.

randomly selecting a feature value (possibly identical tothe prototype feature value) for a random subset of slots.The number of slots that were flipped (i.e., given a ran-dom value) when generating items from the prototypeis a model parameter — increasing the number of slotsthat are flipped decreases the average overlap betweenitems. When all 24 slots are flipped, the resulting itempatterns have 10% overlap with one another on aver-age (i.e., exactly as expected by chance in a layer witha 10% activation level). Thus, with input patterns onecan make a distinction between prototypical features ofthose patterns, which have a relatively high likelihood ofbeing shared across input patterns, and non-prototypical,item-specific features of those patterns (generated by ran-domly flipping slots) which are relatively less likely to beshared across input patterns. Prototype features can bethought of as representing both high-frequency item fea-tures (e.g., if you study pictures of people from Norway,most people have blond hair) as well as contextual fea-tures that are shared across multiple items in an experi-ment (e.g., the fact that all of the pictures were viewed ona particular monitor in a particular room on a particularday). Some simulations involve more complex stimulusconstruction, as described where applicable.

Familiarity as Sharpening

To apply the cortical model to recognition, we exploitthe fact that — as items are presented repeatedly — theirrepresentations in the MTLC layer become sharper (Fig-ure 3). That is, novel stimuli weakly activate a large num-ber of MTLC units, whereas familiar (previously pre-sented) stimuli strongly activate a relatively small num-

a) b)

Figure 3: Illustration of the sharpening of hidden (MTLC)layer activation patterns in a miniature version of our corticalmodel. (a) shows the network prior to sharpening; MTLC ac-tivations (more active = lighter color) are relatively undifferen-tiated. (b) shows the network after Hebbian learning and in-hibitory competition produce sharpening; a subset of the unitsare strongly active, while the remainder are inhibited. In thisexample, we would read out familiarity by measuring the aver-age activity of the k = 5 most active units.

ber of units. Sharpening occurs because Hebbian learn-ing specifically tunes some MTLC units to representthe stimulus. When a stimulus is first presented, someMTLC units, by chance, respond more strongly to thestimulus than other units; these units get tuned by Heb-bian learning to respond even more strongly to the itemthe next time it is presented, and these strongly activeunits start to inhibit units that are less strongly active(for additional discussion of the idea that familiarizationcauses some units to drop out of the stimulus represen-tation, see Li, Miller, & Desimone, 1993). We shouldnote that sharpening is not a novel property of our model— rather, it is a general property of competitive learn-ing networks with graded unit activation values in whichthere is some kind of “contrast enhancement” within alayer (see, e.g., Grossberg, 1986, Section 23, and Gross-berg & Stone, 1986, Section 16).

The sharpening dynamic in our model is consistentwith neural data on the effects of repeated presentationof stimuli in cortex. Electrophysiological studies showthat some neurons that initially respond to a stimulus ex-hibit a lasting decrease in firing, while other neurons thatinitially respond to the stimulus do not exhibit decreasedfiring (e.g., Brown & Xiang, 1998; Li et al., 1993; Xi-ang & Brown, 1998; Miller, Li, & Desimone, 1991;Rolls, Baylis, Hasselmo, & Nalwa, 1989; Riches, Wil-son, & Brown, 1991). According to our model, this lat-ter population consists of neurons that were selected (byHebbian learning) to represent the stimulus, and the for-mer population consists of neurons that are being forcedout of the representation via inhibitory competition.


To index representational sharpness in our model —and through this, stimulus familiarity — we measure theaverage activity of the MTLC units that win the compe-tition to represent the stimulus. That is, we take the av-erage activation of the top

�(192 or 10% of the MTLC)

units computed according to the kWTA inhibitory com-petition function. This “activation of winners” (act win)measure increases monotonically as a function of howmany times a stimulus was presented at study. In con-trast, the simpler alternative measure of using the aver-age activity of all units in the layer is not guaranteed toincrease as a function of stimulus repetition — as a stim-ulus becomes more familiar, the winning units becomemore active, but losing units become less active (due toinhibition from the winning units); the net effect is there-fore a function of the exact balance between these in-creases and decreases (for an example of another modelthat bases recognition decisions on an activity readoutfrom a neural network doing competitive learning, seeGrossberg & Stone, 1986).

Although we use ��

in the simulations re-ported below, we do not want to make a strong claimthat ��

� ��is the way that familiarity is read out

from MTLC. It is the most convenient and analyticallytractable way to do this in our model, but it is far from theonly way of operationalizing familiarity, and it is unclearhow other brain structures might “read out” ��

� ��

from MTLC. We briefly describe another, more neurallyplausible familiarity measure (the time it takes for ac-tivation to spread through the network) in the GeneralDiscussion section.

Finally, we should point out that the idea (espousedabove) that the same network is involved in feature ex-traction and familiarity discrimination is controversial; inparticular, Malcolm Brown, Rafal Bogacz, and their col-leagues (e.g., Brown & Xiang, 1998; Bogacz & Brown,in press) have argued that specialized populations of neu-rons in MTLC are involved in feature extraction versusfamiliarity discrimination. At this point, it suffices to saythat our focus in this paper is on the familiarity discrimi-nation capabilities of the cortical network, rather that itsability to extract features. We address Brown and Bo-gacz’s claims in more detail in the General Discussion.

The Hippocampal Model

We have developed a “standard model” of the hip-pocampus (O’Reilly et al., 1998; O’Reilly & Munakata,2000; O’Reilly & Rudy, 2001; Rudy & O’Reilly, 2001)that implements widely-accepted ideas of hippocampalfunction (Hebb, 1949; Marr, 1971; McNaughton & Mor-ris, 1987; Rolls, 1989; O’Reilly & McClelland, 1994;McClelland et al., 1995; Hasselmo, 1995). Our goal inthis section is to describe the model in just enough de-

Input

EC_in

DG

CA3 CA1

EC_out

Figure 4: Diagram of the hippocampal network. The hip-pocampal network links input patterns in entorhinal cortex(EC) to relatively non-overlapping (pattern-separated) sets ofunits in region CA3; recurrent connections in CA3 bind to-gether all of the units involved in representing a particular ECpattern; the CA3 representation is linked back to EC via regionCA1. Learning in the CA3 recurrent connections, and in pro-jections linking EC to CA3 and CA3 to CA1, makes it possibleto recall entire stored EC patterns based on partial cues. Thedentate gyrus (DG) serves to facilitate pattern separation in re-gion CA3.

tail to motivate the model’s predictions about recognitionmemory. Additional details regarding the architecture ofthe hippocampal model (e.g., the percentage activity ineach layer of the model) are provided in Appendix B.

In the brain, entorhinal cortex (EC) is the interfacebetween hippocampus and neocortex; superficial lay-ers of entorhinal cortex send input to the hippocampus,and deep layers of entorhinal cortex receive output fromthe hippocampus (see Figure 1). Correspondingly, ourmodel subdivides EC into an EC in layer that sends in-put to the hippocampus and an EC out layer that receivesoutput from the hippocampus. Like the input layer ofthe cortical model, both EC in and EC out have a slottedstructure (24 10-unit slots, with 1 unit per slot active).

Figure 4 shows the structure of the model. The job ofthe hippocampal model, stated succinctly, is to store pat-terns of EC in activity, in a manner that supports subse-quent recall of these patterns on EC out. The hippocam-pal model achieves this goal in the following stages:

Input patterns are presented to the model by clamp-ing those patterns onto the input layer, which serves to


impose the pattern on EC in via fixed, 1-to-1 connec-tions. From EC in, activation spreads both directly andindirectly (via the dentate gyrus, DG) to region CA3.The resulting pattern of activity in CA3 is stored by Heb-bian weight changes in the feedforward pathway, and bystrengthening recurrent connections in CA3 between ac-tive units; these weight changes serve to bind the dis-parate elements of the input pattern by linking them to ashared episodic representation in CA3.

An important property of DG and CA3 is that repre-sentations in these structures are very sparse — relativelyfew units are active for a given stimulus. The hippocam-pus’ use of sparse representations gives rise to patternseparation. If only a few units are active per input pat-tern, then overlap between the hippocampal representa-tions of different items tends to be minimal (Marr, 1971;see O’Reilly & McClelland, 1994 for mathematical anal-ysis of how sparseness results in pattern separation, andthe role of the DG in facilitating pattern separation).

Next, to complete the loop, the CA3 representationneeds to be linked back to the original input pattern. Thisis accomplished by linking the CA3 representation to ac-tive units in region CA1. Like CA3, region CA1 containsa re-representation of the input pattern. However, unlikethe CA3 representation, the CA1 representation is invert-ible — if an item’s representation is activated in CA1,well-established connections between CA1 and EC outallow activity to spread back to the item’s representa-tion in EC out. Thus, CA1 serves to “translate” betweensparse representations in CA3 and more overlapping rep-resentations in EC (for more discussion of this issue, seeMcClelland & Goddard, 1996; O’Reilly et al., 1998).

At test, when a previously studied EC in pattern (ora subset thereof) is presented to the hippocampal model,the model is capable of reactivating the entire CA3 pat-tern corresponding to that item via strengthened weightsin the EC-to-CA3 pathway, and strengthened CA3 recur-rents. Activation then spreads from the item’s CA3 rep-resentation to the item’s CA1 representation via strength-ened weights, and (from there) to the item’s EC out rep-resentation. In this manner, the hippocampus manages toretrieve a complete version of the studied EC pattern inresponse to a partial cue.

The Hippocampal Recall Measure

To apply the hippocampal model to recognition, weexploit the fact that studied items tend to trigger recall(of the item itself), more so than lure items. Thus, a highlevel of match between the test probe (presented on theEC input layer) and recalled information (activated overthe EC output layer) constitutes evidence that an itemwas studied. Also, we exploit the fact that lures some-times trigger recall of information that mismatches therecall cue. Thus, mismatch between recalled informa-

tion and the test probe tends to indicate that an item wasnot studied.

For each test item, we generate a recall score usingthe formula:

��

�� (1)

where match is the number of recalled features (onEC out) that match the cue (on EC in), and mismatch islikewise the number that mismatch (a feature is countedas “recalled” if the unit corresponding to that feature inEC out shows activity > .9); numslots is a constant thatreflects the total number of feature slots in EC (24, inthese simulations).

One should appreciate that equation 1 is not the onlyway to apply the hippocampal model to recognition. Forexample, instead of looking at recall of the test cue it-self, we could attach contextual tags to items at study,leave these tags out at test, and measure the extent towhich items elicit recall of contextual tags. Also thisequation does not incorporate the fact that recall of item-specific features (i.e., features unique to particular itemsin the item set) is more diagnostic of study status thanrecall of prototypical features — if all items in the exper-iment are fish, recall of prototypical fish features (e.g.,“I studied fish”) in conjunction with a test item does notmean that you studied this particular item. We selectedthe

��

�� rule because it is a simple way

to reduce the vector output of the hippocampal modelto a scalar that correlates with the study status of testitems. Assessing the optimality of this rule, relative toother rules, and exploring ways in which different rulesmight be implemented neurally, are topics for future re-search.

The only place where we deviate from using��

�� in this paper is in our related lure

simulations, where distractors are related to particularstudied items, but studied items are reasonably distinctfrom one another. In this situation, we use a recall-to-reject rule that places a stronger weight on mismatchingrecall. The YN Related Lure Simulations section of thepaper contains a detailed account of our reasons for us-ing recall-to-reject, and the implications of using this rulein place of

��

�� .

Simulation Methodology

Our initial simulations involve a side-by-side com-parison of the cortical and hippocampal networks re-ceiving the exact same input patterns. This method al-lows us to analytically characterize differences in howthese networks respond to stimuli. A shortcoming ofthis “side-by-side” approach is that we can not exploredirect interactions between the two systems. To rem-edy this shortcoming, we also present simulations using


a combined model where the cortical and hippocampalnetworks are connected in serial (such that the corticalregions involved in computing stimulus familiarity serveas the input to the hippocampal network) — this arrange-ment more accurately reflects how cortex and hippocam-pus are arranged in the brain (see Figure 1).

Basic Methods

In our recognition simulations, for each simulated“participant” we re-randomized the connectivity patternsfor partial projections (e.g., if the specified amount ofconnectivity between layer X and layer Y was 25%, theneach unit in layer Y was linked at random to 25% of theunits in layer X) and we initialized the network weightsto random values (weight values in the cortical modelwere sampled from a uniform distribution with mean =.5 and range = .25; see Appendix B for weight initial-ization parameters for different parts of the hippocam-pal model). The purpose of this randomization was toensure that units in MTLC and the hippocampus woulddevelop specialized receptive fields (i.e., they would bemore strongly activated by some input features than oth-ers) — this kind of specialization is necessary for com-petitive learning.

After randomization was complete, the cortical andhippocampal models were (separately) given a list ofitems to learn, followed by a recognition test in whichthe models had to discriminate between 10 studied tar-get items and 10 nonstudied lure items. No learningoccurred at test. Unless otherwise specified, all of ourrecognition simulations used the same set of parameters(hereafter referred to as the basic parameters; these pa-rameters are described in detail in Appendix C). In ourbasic-parameter simulations, we used a 20-item study list(10 target items, plus 10 interference items that were nottested), and the average amount of overlap between itemswas 20% — 20% overlap between items was achieved bystarting with a 24-slot prototype pattern, and then gener-ating items by randomly selecting a feature value for 16randomly selected slots (note that each item overlaps atleast 33% with the prototype, but on average items over-lap 20% with each other).

To facilitate comparison between the models, weused hippocampal and cortical parameters that yieldedroughly matched performance across the two modelsfor both single-probe (yes-no, or YN) and forced-choice(FC) recognition tests. We matched performance in thisway to alleviate concerns that differential effects of ma-nipulations on hippocampal recall and MTLC familiarityare attributable simply to different overall levels of per-formance in the two networks. However, this matchingdoes not constitute a strong claim that hippocampal andcortical performance are — in reality — matched whenoverlap equals 20% and study list length equals 20.

Simulating YN and FC Testing

To simulate YN recognition performance, items werepresented one at a time at test, and we recorded the fa-miliarity score (in the cortical model) and recall score(in the hippocampal model) triggered by each item. Forthe cortical model, we set an unbiased criterion for eachsimulated participant by computing the average familiar-ity scores associated with studied and lure items (respec-tively) and then placing the familiarity criterion exactlybetween the studied and lure means. All items triggeringfamiliarity scores above this criterion were called “old”.

For the hippocampal model, we took a different ap-proach to criterion setting; as discussed in Simulation 2below, it is possible to set a high recall criterion that issometimes crossed by studied items, but never crossed bylures. We assume that participants are aware of this fact(i.e., that high amounts of recall are especially diagnosticof having studied an item) and set a recall criterion thatis high enough to avoid false recognition. Accordingly,in our basic-parameter simulations, we used a fixed, rel-atively high criterion for calling items “old” (recall >=.40). This value was chosen because — assuming otherparameters are set to their “basic” values — it is some-times exceeded by studied items, but never by lures (un-less lures are constructed to be similar to specific studieditems; see Simulation 3 for more details).

For both models, we used �� (computed on individ-ual participants’ hit and false alarm rates) to index YNrecognition sensitivity.1

Our method for simulating FC testing was straight-forward: We presented the two test alternatives, one at atime. For the cortical model, we recorded the ��

� ��

score associated with each test alternative, and selectedthe item with the higher score. For the hippocampalmodel, we recorded the

��

�� score as-

sociated with each test alternative, and selected the itemwith the higher score. For both models, if there was a tie,one of the two test alternatives was selected at random.

All of the simulation results reported in the text ofthe paper are significant at �� .001. In graphs of simu-lation results (starting with Figure 9), error bars indicatethe standard error of the mean, computed across simu-lated participants. When error bars are not visible, this isbecause they are too small relative to the size of the sym-bols on the graph (and thus are covered by the symbols).

1 �� , where � is the inverse of the normal distri-bution function, � is the hit rate, and � is the false alarm rate. Toavoid problems with � � being undefined when hit or false alarm ratesequal 0 or 1, we adjusted hit and false alarm rates using the correc-tion suggested by Snodgrass and Corwin (1988) prior to computing � � :� � �� !"�$#% '&(� )*� , where � = the number of “old” responses, &= the total number of items, and

�= the corrected percent “old” value.


Pattern Separation in MTLC and the Hippocampus

Input Overlap

0.0 0.2 0.4 0.6 0.8 1.0

Rep

rese

ntat

ion

Ove

rlap

0.0

0.2

0.4

0.6

0.8

1.0

MTLCHippo

Figure 5: Results of simulations exploring pattern separa-tion in the hippocampal and cortical models. In these simu-lations, we created pairs of items and manipulated the amountof overlap between paired items. The graph plots the amountof input-layer overlap for paired items versus: 1) CA3 overlapin the hippocampal model, and 2) MTLC overlap in the corti-cal model. In this graph, all points below the diagonal (dashedline) indicate pattern separation (i.e., representational overlap< input overlap). The hippocampal model showed a strongtendency towards pattern separation (CA3 overlap < < inputoverlap); the cortical model showed a smaller tendency towardspattern separation (MTLC overlap was slightly less than inputoverlap).

Part I: Basic Network Properties

Simulations in this section address basic propertiesof the cortical and hippocampal networks, includingdifferences in their ability to assign distinct (“pattern-separated”) representations to input patterns, and differ-ences in their operating characteristics. We also discusssources of variability in the two networks.

Simulation 1: Pattern Separation

Many of the differences between hippocampally- andcortically-driven recognition in our model arise from thefact that the hippocampal network exhibits more patternseparation than the cortical network. To document thetwo networks’ pattern separation abilities, we ran simu-lations where we manipulated the amount of overlap be-tween paired input patterns. The first item in each pairwas presented at study and the second item was presentedat test. For each pair, we measured the resulting amount

of overlap in region CA3 of the hippocampal model, andin the hidden (MTLC) layer of the cortical model. Pat-tern separation is evident when overlap between the in-ternal representations of paired items (in CA3 or MTLC)is less than the amount of input overlap.

The results of these simulations (Figure 5) confirmthat, while both networks show some pattern separation,the amount of pattern separation is larger in the hip-pocampal model. They also show that the hippocampus’ability to assign distinct representations to stimuli is lim-ited — as overlap between input patterns increases, hip-pocampal overlap eventually increases above floor levels(although it always lags behind input pattern overlap).

Simulation 2: Nature of the UnderlyingDistributions

One way to characterize how cortical and hippocam-pal contributions to recognition differ is to plot the dis-tributions of these signals for studied items and lures.Given 20% average overlap between input patterns, theMTLC familiarity distributions for studied items andlures are Gaussian and overlap strongly (Figure 6a) —this is consistent with the idea, expressed by Yonelinas,Dobbins, Szymanski, and Dhaliwal (1996) and manyothers (e.g., Hintzman, 1988), that familiarity is well-described by standard (Gaussian) signal-detection the-ory. In contrast, the hippocampal recall distributions(Figure 6b) do not adhere to a simple Gaussian model.The bulk of the lure recall distribution is located at thezero recall point, although some lures trigger above-zerorecall. The studied recall distribution is bimodal and,crucially, it extends further to the right than the lure re-call distribution, so there are some (high) recall scoresthat are sometimes triggered by studied items, but neverby lures. Thus, high levels of recall are highly diagnostic— for these parameters, if an item triggers a recall scoreof .2 or greater, you can be completely sure that it wasstudied.

The low overall level of lure recall in this simulationcan be attributed to hippocampal pattern separation. Be-cause of pattern separation, the CA3 representations oflures do not overlap strongly with the CA3 representa-tions of studied items. Because the CA3 units activatedby lures (for the most part) were not activated at study,these units do not possess strong links to CA1; as such,activity does not spread from CA3 to CA1, and recalldoes not occur.

The studied recall distribution is bimodal becauseof nonlinear attractor dynamics in the hippocampus. Ifa studied test cue accesses a sufficiently large numberof strengthened weights, it triggers pattern completion:Positive feedback effects (e.g., in the CA3 recurrents) re-


a) b)MTLC Familiarity Histogram, .2 Overlap

Familiarity

0.65 0.70 0.75 0.80

Fre

quen

cy

0.00

0.02

0.04

0.06

0.08

0.10

0.12Studied

Lure

Hippocampal Recall Histogram, .2 Overlap

Recall

-0.2 0.0 0.2 0.4 0.6 0.8 1.0

Fre

quen

cy

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8Studied

Lure

Figure 6: (a) histogram of the studied and lure MTLC familiarity distributions for 20% average overlap; (b) histogram of thestudied and lure hippocampal recall distributions for 20% average overlap

sult in strong re-activation of the CA3 and CA1 units thatwere activated at study, thereby boosting recall. Moststudied items benefit from these positive feedback ef-fects but, because of variability in initial weight values,some studied items do not have weights strong enoughto yield positive feedback. These items only weakly ac-tivate CA3 and are poorly recalled, thereby accountingfor the extra “peak” at recall = 0.

Limits on Diagnosticity: High Average Overlap

Increasing the average amount of overlap betweenitems reduces the diagnosticity of the hippocampal re-call signal. When the amount of overlap between inputpatterns is high (e.g., 40.5%, instead of 20%), both stud-ied items and lures trigger large amounts of recall, suchthat the studied and lure recall distributions are roughlyGaussian and overlap extensively (Figure 7).

High levels of lure recall occur in the high-overlapcondition because of pattern separation failure in thehippocampus. As documented in Simulation 1 (Fig-ure 5), the hippocampus loses its ability to assign dis-tinct representations to input patterns when overlap be-tween inputs is very high. In this situation, the same CA3units — the units that are most sensitive to frequently-occurring prototype features — are activated again andagain by studied patterns, and these units acquire verystrong weights to the representations of prototype fea-tures in CA1. When items are presented at test, theyactivate these “core” CA3 units to some extent (regard-less of whether or not the test item was studied), and ac-tivation spreads very quickly to CA1, leading to possi-bly erroneous recall of prototype features in response to

Hippocampal Recall Histogram, .405 Overlap

Recall

-0.2 0.0 0.2 0.4 0.6 0.8 1.0

Fre

quen

cy

0.00

0.05

0.10

0.15

0.20

0.25StudiedLure

Figure 7: Histogram of the studied and lure hippocampal re-call distributions for 40.5% average overlap

both studied items and lures. Figure 8 shows that in-creasing overlap increases the probability that prototyp-ical features of studied items and lures will be recalled,and reduces the probability that item-specific features ofstudied items will be recalled (in part, because the hip-pocampus intrudes prototypical features in place of theseitem-specific features).

In summary, The results presented here show thathippocampal recall has two modes of operation: Wheninput patterns have low-to-moderate average overlap,high levels of matching recall are highly diagnostic —studied items sometimes trigger strong recall (of item-specific and prototype features) but lures trigger virtu-


Recall of Prototypical and Item-SpecificFeatures as a Function of Overlap

Average Input Overlap

0.2 0.3 0.4 0.5

P(R

ecal

l)

0.0

0.2

0.4

0.6

0.8

1.0

Studied PrototypicalStudied Item-SpecificLure PrototypicalLure Item-Specific

Figure 8: Plot of the probability that item-specific and prototypical features of studied items and lures will be recalled, as afunction of overlap. As overlap increases, the amount of prototype recall triggered by studied items and lures increases, the amountof item-specific recall triggered by studied items decreases.

ally no recall. In contrast, when input patterns have highaverage overlap, recall functions as a standard signal-detection process — both studied items and lures triggervarying degrees of prototype recall.

Simulation 3: YN Related Lure Simulations

The strengths of the hippocampal model are most ev-ident on yes-no related lure recognition tests, where luresare similar to studied items, but studied items are dissim-ilar to one another. In this section, we show how thehippocampal model outperforms the cortical model onthese tests, because of its superior pattern separation andpattern completion abilities.

Methods

To simulate the related lure paradigm, we first cre-ated studied item patterns with 20% average overlap be-tween items. Then, for each studied (target) item, wecreated a related lure item by taking the studied item andflipping a pre-specified number of slots; to vary target-lure similarity, we varied the number of slots that weflipped to generate lures (less flipping results in moreoverlap). For comparison, we also ran simulations with“unrelated” lures that were sampled from the same itempool as studied items.

In the related lure simulations presented here (andlater in the paper), we used a recall-to-reject hippocam-pal decision rule instead of our standard

��

�� rule. According to this rule, items that trig-

ger any mismatch are given a “new” response, otherwisethe decision is based on match. We used this rule for two

reasons: First, there is extensive empirical evidence that,when lures are similar to studied items, but studied itemsare unrelated to one another, participants use recall-to-reject (see, e.g., Rotello, Macmillan, & Van Tassel,2000; Rotello & Heit, 2000; Yonelinas, 1997; but seeRotello & Heit, 1999 for a contrasting view). Second,it is computationally sensible to use recall-to-reject —we show that the presence of mismatching recall in thisparadigm is highly diagnostic of an item being nonstud-ied.

Results

Figure 9 shows the results of our related lure simu-lations: Recognition performance based on MTLC fa-miliarity gets steadily worse as lures become increas-ingly similar to studied items; in contrast, recognitionbased on hippocampal recall is relatively robust to thelure similarity manipulation. Figure 9 also shows thathippocampal performance is somewhat better for relatedlures when the recall-to-reject rule is used (as opposed to��

�� ).

The cortical model results can be explained in termsof the fact that the cortical model assigns similar repre-sentations to similar stimuli — because the representa-tions of similar lures (vs. dissimilar lures) overlap morewith the representations of studied items, similar luresbenefit more from learning that occurs at study. Thus,lure familiarity smoothly tracks target-lure similarity;increasing similarity monotonically lowers the target-lure familiarity difference, leading to decreased discrim-inability.

In contrast, the hippocampal recall signal triggeredby lures is stuck at floor until target-lure similarity >


Effect of Target-Lure Similarity on YN d'

Target-Lure Similarity

unrelated 0.5 0.7 0.9

YN

d'

0.0

0.5

1.0

1.5

2.0

2.5

Hippo with recall-to-rejectHippo (match - mismatch)MTLC

Figure 9: YN recognition sensitivity ( � � ) in the two models,as a function of target-lure similarity. Target-lure similaritywas operationalized as the proportion of input features sharedby targets and corresponding lures; note that the average levelof overlap between studied (target) items was held constant at20%. These simulations show that the hippocampal model ismore robust to increasing target-lure similarity than the corticalmodel. The figure also shows that hippocampal performancefor related lures is better when a recall-to-reject rule is usedinstead of match - mismatch

.

60%, and lures do not start to trigger above-criterion (i.e.,> .40) recall until target-lure similarity > 80%. This oc-curs because of hippocampal pattern separation – lureshave to be very similar to studied items before they ac-cess enough strengthened weights to trigger recall.

The hippocampus also benefits from the fact thatlures sometimes trigger pattern completion of the corre-sponding studied item, and can subsequently be rejectedbased on mismatch between recalled information and thetest cue. Figure 10 illustrates the point (mentioned ear-lier) that when lures resemble studied items, but studieditems are not related to one another, mismatching recallis highly diagnostic — studied items virtually never trig-ger mismatching recall, but lures sometimes do. As such,it makes sense to use a rule (like recall-to-reject) that as-signs a very high weight to mismatching recall.

Figure 10 also shows that mismatching recall trig-gered by lures increases substantially with increasingtarget-lure similarity. This increase in mismatching re-call helps offset, to some degree, increased matching re-call triggered by related lures. With recall-to-reject, theonly way that lures can trigger an old response is if theytrigger a large amount of matching recall but no mis-matching recall. The odds of this happening are verylow.

In summary: These simulations demonstrate that

Mismatching Recall as a Function of Lure Similarity



P(M

ism

atch

)

0.00

0.05

0.10

0.15

0.20

0.25

0.30Studied itemsLures

Figure 10: Plot of the probability that lures and studied itemswill trigger mismatching recall, as a function of target-lure sim-ilarity. This probability is close to floor for studied items; theprobability that lures will trigger mismatching recall increaseswith increasing target-lure similarity.

both networks can support good performance on YNrecognition tests with lures that are unrelated to studieditems. When the networks are challenged by boostingtarget-lure similarity, performance in both networks suf-fers; however, the hippocampus is more robust to thismanipulation than cortex. As such, the model predictsthat recognition discrimination based on hippocampal re-call should be better than discrimination based on MTLCfamiliarity on YN tests with related lures. This predictionis consistent with the view, expressed in several empiricalpapers, that recall is especially important for discrimi-nating between studied items and very similar distractors(e.g., Hintzman, Curran, & Oppy, 1992; Rotello et al.,2000).

Sources of Variability

The final issue that we need to address in the BasicNetwork Properties section is variability. Recognitionperformance involves detecting the presence of a vari-able memory signal against a background of noise. Manydistinct forms of variability can affect recognition per-formance; we need to carefully delineate which of thesesources of variability are present in our models, because— as we show later — different forms of variability havedifferent implications for recognition performance.

The primary source of variability in our models issampling variability: variation in how well, on average,neural units are connected to (sampled by) other neu-ral units in the network. Note that our use of the term“sampling variability” differs from how other modelershave used this term. In our model, sampling variability


is a function of variability in the initial values assignedto weights in the network. Other models use samplingvariability to refer to variability in which item featuresare presented to the model at study and test (Atkinson &Estes, 1963), or variability in which memory trace is re-trieved at test (Gillund & Shiffrin, 1984).

Sampling variability arises because, at the beginningof each simulation, weight strengths and connectivitypatterns are set randomly. As discussed earlier, this ran-domness helps units in MTLC and the hippocampus formspecialized representations of the input space. It alsohas the consequence that, by chance, some input featuresare sampled better (by units further “downstream”) thanother input features.

An important property of sampling variability is thatit decreases as network size increases. Intuitively, as thenumber of units and connections increases, the odds thatany one input pattern will be “undersampled” relativeto another decreases. We conducted simulations to ex-plore this issue, and the results were very clear: As weincreased network size, variability in MTLC familiarityand hippocampal recall scores steadily decreased, and � �scores steadily increased.

In a network scaled to the approximate size of thehuman brain, sampling variability would likely be negli-gible. Therefore, we conclude that other forms of vari-ability must be at play in the human brain; we brieflydescribe some other sources of variability below.

Other Sources of Variability

A potentially important source of variability in recalland familiarity scores is variability in how well stimuliare encoded at study. This kind of encoding variabilitycan arise, for example, if participants’ attention fluctu-ates over the course of an experiment — some items willbe encoded more strongly than others, leading to higherrecall and familiarity scores at test.

An important property of encoding variability, whichis not true of sampling variability, is that it affects stud-ied items and related lures in tandem. That is, encodingfluctuations that boost the memory signal triggered by astudied item also boost the memory signal triggered bylures that are similar to that studied item (e.g., if “cat” isencoded so as to be especially familiar, the related lure“cats” will also be highly familiar). In contrast, sam-pling variability operates independently on each inputfeature; in small networks where sampling variability isthe dominant source of variance, noise associated withsampling of non-shared (discriminative) features of over-lapping stimuli counteracts much of the shared variabil-ity in memory scores triggered by these items. We revisitthis issue later, when we explore how lure relatedness in-teracts with test format (Simulation 4).

Another source of variability in recall and famil-iarity scores is variability in pre-experimental exposureto stimuli: Some stimuli have been encountered exten-sively prior to the experiment, in many different contexts;other stimuli are relatively novel; for evidence that pre-experimental presentation frequency affects recognitionmemory, see Dennis and Humphreys (2001). Variabilityin pre-experimental exposure (like encoding variability,but unlike sampling variability) affects studied items andrelated lures in tandem.

Finally, in addition to variability in how much testitems overlap with pre-experimental memory traces,there is also variability in how much items overlap withother items presented in the experiment; this kind ofvariability also affects studied items and related lures intandem. Overlap-related variability is already presentin the model, but its effect on performance is typicallydwarfed by sampling variability. Consequently, variabil-ity in overlap should play a much larger role, proportion-ally, in larger networks with minimal sampling variabil-ity.

Sources of Variability: Summary

In summary, the basic model (as described above)is strongly influenced by sampling variability, and lacksother, plausible sources of variability such as encodingvariability. Given that sampling variability is not likely tobe a factor in human recognition memory performance,one might conclude that we should eliminate this sourceof variability and incorporate other sources. Unfortu-nately, this is not practical at present — models that arelarge enough to eliminate sampling variability cannot befeasibly run on available computational hardware. Fur-thermore, adding more variability on top of samplingvariability in our small networks leads to poor perfor-mance, unless other steps are taken to compensate forincreased variability (e.g., increasing the learning rate).

Because of these limitations, we refrain from makingstrong predictions about how manipulations affect vari-ance in this paper. Nevertheless, we can still use the basicmodel to explain many phenomena that do not dependon the exact source of variability. Also, it is relativelystraightforward to supplement the basic model with otherforms of variability on an “as needed” basis, and we dothis to make some important points in Simulation 4.

Part II: Applications To BehavioralPhenomena

The simulations in this part of the paper build on thebasic results described earlier, by applying the modelsto a wide range of empirical recognition memory phe-


nomena (e.g., how does interference affect recognitionperformance in the two models). Whenever possible, wepresent data that speak to the model’s predictions.

Simulation 4: Lure Relatedness and TestFormat Interactions

As we saw in Figure 9, the model predicts that thehippocampus should outperform cortex on standard YNtests where participants have to discriminate betweenstudied items and related lures. In this section, we showhow giving participants a forced choice between stud-ied items and corresponding related lures benefits per-formance in the cortical model but not the hippocampalmodel, thereby mitigating the hippocampal advantage.

Forced-Choice Testing and Covariance

In an forced-choice (FC) test, participants are simul-taneously presented with a studied item and a lure, andare asked to select the studied item. The specific ver-sion of this test that boosts cortical performance involvespairing studied items with corresponding related lures(i.e., lures related to the paired studied item; for exam-ple, study “RAT”, test “RAT” vs. “RATS”).

The central insight as to why this format improvescortical performance with related lures is that, eventhough related lures trigger strong feelings of familiar-ity (because they overlap with the studied items), corre-sponding studied items are reliably more familiar. Be-cause performance in an FC test is based on the differ-ence in familiarity between paired items, even small dif-ferences can drive good performance, as long as they arereliable.

The reliability of the familiarity difference dependson where variability comes from in the model. Asdiscussed in the previous section, some kinds of vari-ability (e.g., differences in encoding strength and pre-experimental exposure) necessarily affect studied and re-lated lure familiarity in tandem, whereas other kinds ofvariability (e.g., sampling variability) do not. When theformer kind of variability predominates, the familiarityvalues of studied items and corresponding lures will behighly correlated, and therefore their difference will bereliable. When sampling variability predominates, thestudied-lure familiarity difference will be somewhat lessreliable.

More formally, the beneficial effect of using an FCtest depends on covariance in the familiarity scores trig-gered by studied items and corresponding related lures(Hintzman, 1988, 2001). The variance of the studied-lurefamiliarity difference is given by the following equation:��

�� (2)

FC Accuracy in MTLC



Pro

port

ion

Cor

rect

0.5

0.6

0.7

0.8

0.9

1.0

Corresponding luresNon-corresponding lures

Figure 11: FC accuracy in the MTLC model as a func-tion of target-lure similarity, using corresponding and non-corresponding FC testing. For high levels of target-lure sim-ilarity, FC performance is slightly better with correspondinglures vs. with non-corresponding lures.

where�

represents familiarity of studied items and �that of lures,

�� is variance and � �� is covariance be-

tween�

and � . Equation 2 shows that increasing co-variance reduces the variance of the

� �� familiaritydifference, which in turn boosts FC performance.

FC Simulations Using the Basic Cortical Model

Our first concern was to assess how much FC testingwith corresponding related lures benefits performancein our basic cortical model. Towards this end, we ransimulations using a paradigm introduced by Hintzman(1988); in these simulations, we compared FC perfor-mance with corresponding related lures (i.e., study A, B;test A vs. A’, B vs. B’, where A’ and B’ are lures re-lated to A and B, respectively) to FC performance withnon-corresponding lures (e.g., study A, B; test A vs. B’;B vs. A’). To the extent that there is covariance betweenstudied items and corresponding lures, this will benefitperformance in the corresponding lure condition relativeto the non-corresponding lures.

As shown in Figure 11, FC performance ishigher with corresponding related lures than with non-corresponding lures — this replicates the empirical re-sults obtained by Hintzman (1988) and shows that thereis some covariance present in the basic cortical model.To quantify the level of covariance underlying these re-sults, we can compute the following ratio:

� � � ��

�� (3)

When�

= 1, covariance will completely offset studiedand lure variance, and the studied-lure familiarity differ-


ence will be completely reliable (i.e., variance = 0);�

= 0means that there is no covariance. For target-lure similar-ity = .92, the covariance ratio

�= .27 in the correspond-

ing condition and�

= -.01 in the non-corresponding con-dition. Thus, the model exhibits roughly one third themaximal level of covariance possible.

In summary: Although the basic model qualitativelyexhibits an FC advantage with corresponding relatedlures, this advantage is not quantitatively very large. Thisis because the dominant source of variability in the basiccortical model is sampling variability, which — as dis-cussed above — does not reliably affect studied itemsand corresponding lures in tandem.

Cortical and Hippocampal Simulations With EncodingVariability

Next, we wanted to explore a more realistic scenarioin which the contribution of sampling variability to over-all variability is small, relative to other forms of variabil-ity (like encoding variability) that affect studied itemsand corresponding lures in tandem. Increasing the rela-tive contribution of encoding variability should increasecovariance, and thereby increase the extent to which thecortical model benefits from use of an FC-correspondingtest. We were also interested in how test format affectsthe hippocampal model’s ability to discriminate betweenstudied items and related lures (when encoding variabil-ity is high). To address these issues, we ran simula-tions with added encoding variability in both the corti-cal and hippocampal models, where we manipulated testformat (FC-corresponding vs. FC-non-corresponding vs.YN) and lure relatedness.

Methods

We added encoding variability using the followingsimple manipulation: For each item at study, the learn-ing rate was scaled by a random number from the 0-to-1uniform distribution. However, this manipulation by it-self does not achieve the desired result; the influence ofencoding variability is still too small relative to samplingvariability, and overall performance levels with addedencoding variability are unacceptably low. To boost therelative impact of encoding variability (and overall per-formance), we also increased the learning rate in bothmodels to three times its usual value. Under this regime,random scaling of the learning rate at study has a muchlarger effect on studied-item (and related-lure) familiar-ity than random differences in how well features are sam-pled. We should note that using a large learning rate hassome undesirable side-effects (e.g., increased interfer-ence), but these side-effects are orthogonal to the ques-tions we are asking here. As with the hippocampal re-lated lure simulations presented earlier, the hippocampalsimulations presented here used a recall-to-reject deci-

sion rule. We applied this rule to FC testing in the follow-ing manner: If one item triggered mismatching recall, butthe other item did not, the second item was selected; oth-erwise, the item triggering a higher

��

��

recall score was selected.

Cortical FC Results

As expected, the corresponding vs. non-corresponding difference for the cortical modelwas much larger in our simulations with encodingvariability (Figure 12a) than in our simulations withoutencoding variability (Figure 11). Computing the averagecovariance/variance ratio for the .92 target-lure overlapcondition shows that

�= .62 for corresponding lures

vs.�

= -.05 for non-corresponding lures. This is morethan double the covariance in the basic model (.62 vs..27), and confirms our intuition that decreasing thecontribution of sampling variability relative to encod-ing variability should increase covariance and boostperformance in the FC-corresponding condition.

Hippocampal FC Results

In contrast to the cortical model results, FC-corresponding and FC-non-corresponding performancewere almost identical in the hippocampal model (Fig-ure 12a). It seems clear that the same argumentsabout covariance benefiting FC-corresponding perfor-mance should hold for hippocampus as well as for cortex.Why then does the hippocampus behave differently thanthe cortex in this situation? This can be explained bylooking at what happens on trials where the studied itemis not recalled — on these trials, participants can stillrespond correctly if the lure triggers mismatching recall(and is rejected on this basis). The key insight is thatstudied recall and lure misrecall are independent whennon-corresponding lures are used (in effect, participantsget two independent chances to make a correct response),but they are highly correlated when corresponding luresare used — if the studied item does not trigger any recall,the corresponding lure probably will not trigger any re-call either. Thus, using corresponding lures can actuallyhurt performance in the hippocampal model by deprivingparticipants of an extra, independent chance to respondcorrectly (via recall-to-reject) on trials where studied re-call fails. This harmful effect of covariance cancels outthe beneficial effects of covariance described earlier.

Because the cortical model benefits from FC-corresponding (vs. non-corresponding) testing but thehippocampal model does not, the performance of the cor-tical model relative to the hippocampal model is better inthis condition.

YN Results

The results of our YN simulations with encodingvariability (Figure 12b) were identical to the results of


FC Testing


unrelated 0.71 0.92

Pro

port

ion

Cor

rect

0.5

0.6

0.7

0.8

0.9

1.0

Hippo CHippo NMTLC CMTLC N

YN Testing


unrelated 0.71 0.92

YN

d'

0.0

0.5

1.0

1.5

2.0

2.5

HippoMTLC

a)

b)

Figure 12: Cortical and hippocampal related-lure simulationsincorporating strong encoding variability. (a) Results of FCsimulations. When encoding variability is present, the corti-cal model benefits very strongly from use of corresponding vs.non-corresponding lures (more so than in our simulations with-out encoding variability). In contrast, the hippocampal model(using recall-to-reject) performs equally well with correspond-ing vs. non-corresponding lures. (b) Results of YN simulationsusing the same parameters. As in our previous related-lure sim-ulations (Figure 9), the cortical model is severely impaired rel-ative to the hippocampal model on these tests.

our earlier YN-related-lure simulations. As before, wefound that the hippocampal model was much better thanthe cortical model at discriminating between studieditems and related lures on YN tests.

Tests of the Model’s Predictions

One way to test the model’s predictions is to lookat recognition in patients with focal, relatively completehippocampal damage. Presumably, these patients are

Figure 13: Sample stimuli from the Holdstock et al. (2002)related lure experiment. Participants studied pictures of objects(e.g., the horse shown in the upper left). At test, participantshad to discriminate studied pictures from three highly relatedlures (e.g., the horses shown in the upper right, lower left, andlower right).

relying exclusively on MTLC familiarity when makingrecognition judgments (in contrast to controls, who haveaccess to both hippocampal recall and MTLC familiar-ity). As such, patients should perform poorly relative tocontrols on tests where hippocampus outperforms cortex,and they should perform relatively well on tests wherehippocampus and cortex are evenly matched. Apply-ing this logic to the results shown in Figure 12, patientsshould be impaired on YN recognition tests with relatedlures, but they should perform relatively well on FC-corresponding tests with related lures, and on tests withunrelated lures (regardless of test format).

To test this prediction, we collaborated with AndrewMayes and Juliet Holdstock to test patient YR, who suf-fered focal hippocampal damage, sparing surroundingMTLC regions (for details of the etiology and extent ofYR’s lesion, see Holdstock et al., 2002). YR is severelyimpaired at recalling specific details — thus, YR has torely almost exclusively on MTLC familiarity when mak-ing recognition judgments. Holdstock et al. (2002) de-veloped YN and FC tests with highly related lures thatwere closely matched for difficulty, and administeredthese tests to patient YR and her controls. Figure 13shows sample stimuli from this experiment. Results fromthis experiment can be compared to results from 15 otherYN item recognition tests and 25 other FC item recogni-tion tests that used lures that were less strongly related tostudied items (Mayes et al., 2002); we refer to these testsas the YN-low-relatedness and FC-low-relatedness tests,respectively.

Figure 14 shows that, exactly as we predicted, YR


Performance of Patient YR Relative to Controls

YN Hi FC Hi YN Low FC Low

SD

sA

bove

/Bel

owC

ontr

olM

ean

-4

-3

-2

-1

0

1

2

Figure 14: Performance of YR relative to controls on matchedYN and FC-corresponding tests with highly related (Hi) lures;the graph also plots YR’s average performance on 15 YN testsand 25 FC tests with less strongly related (Low) lures. YR’sscores are plotted in terms of number of standard deviationsabove or below the control mean. For the YN and FC low-relatedness tests, error bars indicate the maximum and mini-mum z-scores achieved by YR (across the 15 YN tests and the25 FC tests, respectively). YR was significantly impaired rel-ative to controls on the YN test with highly related lures (i.e.,her score was �� SDs below the control mean) but YR per-formed slightly better than controls on the FC test with highlyrelated lures. YR was not significantly impaired, on average,on the tests that used less strongly related lures.

was significantly impaired on a YN recognition test thatused highly related lures, but showed relatively sparedperformance on an FC version of the same test (YR ac-tually performed slightly better than the control mean onthis test). This pattern can not be explained in terms ofdifficulty confounds (i.e., YR performing worse, relativeto controls, on the more difficult test) — controls foundthe YN test with highly related lures to be slightly eas-ier than the FC test. Figure 14 also shows that YR was,on average, unimpaired on YN-low-relatedness and FC-low-relatedness tests. YR performed worse on the YNtest with highly related lures than on any of the 15 YN-low-relatedness tests; this difference can not be attributedto the YN-low-relatedness tests being easier than the YNtest with highly related lures: YR showed unimpairedperformance on the 8 YN low-relatedness tests that con-trols found to be more difficult than the YN test withhighly related lures; for these eight tests her mean z-scorewas 0.04 (

��= 0.49; minimum = -0.54; maximum =

0.65; J. Holdstock, personal communication). We haveyet to test the model’s prediction regarding use of FC-corresponding vs. FC-non-corresponding tests with re-lated lures; based on the results shown in Figure 12, themodel predicts that YR will be more strongly impaired

on FC tests with non-corresponding (vs. corresponding)related lures.

Simulation 5: Associative Recognition andSensitivity to Conjunctions

In this section, we explore the two networks’ perfor-mance on associative recognition tests. On these tests,participants have to discriminate between studied pairsof stimuli (A-B, C-D) and associative lures generated byrecombining studied pairs (A-D, B-C). To show above-chance associative recognition performance, a networkmust be sensitive to whether features occurred together atstudy; sensitivity to individual features does not help dis-criminate between studied pairs and recombined lures.The hippocampus’ ability to rapidly encode and storefeature conjunctions is not in dispute — this is a centralfeature of practically all theories of hippocampal func-tioning, including ours (e.g., Rudy & Sutherland, 1995;Squire, 1992b; Rolls, 1989; Teyler & Discenna, 1986).In contrast, many theorists have argued that neocortex isnot capable of rapidly forming new conjunctive represen-tations (i.e., representations that support differential re-sponding to conjunctions vs. their constituent elements)on its own; see O’Reilly and Rudy (2001) for a review.

Associative recognition can be viewed as a specialcase of the related lure paradigm described earlier. Assuch, we would expect the hippocampus to outperformcortex on YN associative recognition tests, because ofits superior pattern separation abilities, and its ability toreject similar lures based on mismatching recall.

Methods

In our associative recognition simulations, 20 itempairs were presented at study — each “pair” consistedof a 12-slot pattern concatenated with another 12-slotpattern; at test, studied pairs were presented, along withthree types of lures: associative (re-paired) lures, featurelures (generated by pairing a studied item with a non-studied item), and novel lures (generated by pairing twononstudied items). Our initial simulations used a YN testformat.

YN Results

As expected, the hippocampal model outperformedthe cortical model in our YN associative recognition sim-ulations (Figure 15). The other important result from theYN associative recognition simulations was that corti-cal performance was well above chance. This indicatesthat cortex is sensitive (to some degree) to feature co-occurrence in addition to individual feature occurrence.


Associative Recognition in the Two Models

Lure Type

Novel Feature Re-paired

YN

d'

1.5

1.6

1.7

1.8

1.9

2.0

2.1

HippoMTLC

Figure 15: Results of YN associative recognition simulationsin the cortical and hippocampal models. Using parameters thatyield matched performance for unrelated (novel) lures, cortexis impaired relative to the hippocampus at associative recog-nition; nonetheless, cortex performs well above chance on theassociative recognition tests.

The ability of the cortical model to encode stimulusconjunctions can be explained in terms of the fact thatcortex, like the hippocampus, uses sparse representations(as enforced by the k-winners-take-all algorithm). ThekWTA algorithm forces units to compete to represent in-put patterns, and units that are sensitive to multiple fea-tures of a given input pattern (i.e., feature conjunctions)are more likely to win the competition than units that areonly sensitive to single input features. Representationsare more conjunctive in the hippocampus than in cor-tex because representations are more sparse (i.e., thereis stronger inhibitory competition) in the hippocampusthan in the cortex. For additional computational supportfor the notion that cortex should encode low-order con-junctive representations, see O’Reilly and Busby (2002).

Effects of Test Format

In Simulation 4 we showed how giving participantsa forced choice between studied items and correspond-ing related lures mitigates the hippocampal advantage forrelated lures. Analogously, giving participants a forcedchoice between overlapping studied pairs and lures (FC-OLAP testing: study A-B, C-D test A-B vs. A-D) miti-gates the hippocampal advantage for associative recogni-tion. In both cases, performance suffers because partic-ipants do not get an “extra chance” to respond correctly(via recall-to-reject) when studied recall fails.

Typically, FC-OLAP tests are structured in a way

that emphasizes the shared item: Participants are asked,“which of these items was paired with A: B or D?”. Thisencourages participants use a strategy where they cuewith the shared item (A), and select the choice (B orD) that best matches retrieved information. The prob-lem with this algorithm is that success or failure dependsentirely on whether the shared cue (A) triggers recall; ifA fails to trigger recall of B, participants are forced toguess. In contrast, on FC associative recognition testswith non-overlapping choices (FC-NOLAP tests: studyA-B, C-D, E-F; test A-B vs. C-F), participants have mul-tiple, independent chances to respond correctly; even ifA does not trigger recall of B, participant can still re-spond correctly if they recall that C was paired with D(not F).

To demonstrate how recall-to-reject differentiallybenefits FC-NOLAP performance (relative to FC-OLAPperformance), we ran FC-NOLAP and FC-OLAP simu-lations in the hippocampal model using a recall-to-rejectrule. For comparison purposes, we ran another set ofsimulations where decisions were based purely on theamount of matching recall.

Methods

These simulations used a cued recall algorithmwhere, for each test pair (e.g., A-B), we cued with thefirst pair item and measured how well recalled informa-tion matched the second pair item. On FC-OLAP tests,we cued with the shared pair item (A) for both test al-ternatives (A-B vs. A-D). On FC-NOLAP tests we cuedwith the first item of each test alternative (A from A-B,and C from C-D). We had to adjust some model param-eters to get the model to work well using partial cues;specifically, we used a higher-than usual learning rate(.03 vs. .01) to help foster pattern completion of informa-tion not in the cue, and we increased the activation crite-rion for counting a feature as recalled (from .90 to .95)to compensate for the fact that the output of the model isless clean with partial cues.

Results

As expected, FC-NOLAP performance was higher inthe recall-to-reject condition (vs. the “match-only” con-dition) but FC-OLAP performance did not benefit at all(Figure 16). Because of this differential benefit, FC-NOLAP performance was better overall than FC-OLAPperformance in the recall-to-reject condition. Consis-tent with this prediction, Clark, Hori, and Callan (1993)found better performance on an FC-NOLAP associativerecognition test than on an FC-OLAP associative recog-nition test. They explained this finding in a manner thatis consistent with our account — they argued that partic-ipants were using recall of studied pairs to reject lures,


Hippocampal Associative Recognition Performance

Test Format

FC-OLAP FC-NOLAP

Pro

port

ion

Cor

rect

0.5

0.6

0.7

0.8

0.9

1.0

Match onlyRecall-to-reject

Figure 16: Associative recognition performance in the hip-pocampal model, as a function of recall decision rule (matchonly vs. recall-to-reject) and test format (FC-OLAP vs. FC-NOLAP). FC-NOLAP performance benefits from use of recall-to-reject, but FC-OLAP performance does not benefit at all.When recall-to-reject is used, FC-NOLAP performance is bet-ter than FC-OLAP performance.

and that participants have more unique (independent)chances to recall useful information in the FC-NOLAPcondition.


The implications of the above simulation results forpatient performance are clear: Patients with focal hip-pocampal lesions should be impaired, relative to con-trols, on YN associative recognition tests, but they shouldbe relatively less impaired on FC-OLAP associativerecognition tests.

No one has yet conducted a direct comparison of howwell patients with hippocampal damage perform relativeto controls, as a function of test format. However, thereare several relevant data points in the literature. Krollet al. (1996) studied associative recognition performancein a patient with bilateral hippocampal damage (causedby anoxia), as well as other patients with less focal le-sions. In Experiment 1 of the Kroll et al. (1996) pa-per, participants studied 2-syllable words (e.g., “barter”,“valley”) and had to discriminate between studied wordsand words created by re-combining studied words (e.g.,“barley”). Results from the patient with bilateral hip-pocampal damage, as well as control data, are plotted inFigure 17. In keeping with the model’s predictions, thepatient showed impaired YN associative recognition per-formance, but YN discrimination with novel lures (whereneither part of the stimulus was studied) was intact. Fur-thermore, even though the patient was impaired at as-

Effect of Hippocampal Damage (Kroll et al., 1996)

Lure Type

Novel Feature Re-paired

YN

d'

0.0

0.5

1.0

1.5

2.0

2.5

3.0 ControlsPatient

Figure 17: Associative recognition in a patient with bilateralhippocampal damage (taken from Kroll et al. (1996), Experi-ment 1); � � scores for the patient and controls were computedbased on average hit and false alarm rates published in Table 3of that paper. The patient performed better than adult controlsat discriminating studied items from novel lures but was worsethan controls at discriminating studied items from feature lures(where one part of the stimulus was old and one part was new)and was much worse that controls when lures were generatedby re-combining studied stimuli. The pattern reported here isqualitatively consistent with the model’s predictions as shownin Figure 15.

sociative recognition, the patient’s performance in thisexperiment was above chance. This is consistent withthe idea that cortex is sensitive (to some degree) to fea-ture conjunctions. However, this study does not speakto whether cortex can form novel associations betweenpreviously unrelated stimuli — because stimuli (includ-ing lures) were familiar words, participants do not nec-essarily have to form a new conjunctive representation tosolve this task.

Two studies (Vargha-Khadem et al., 1997; Mayeset al., 2001) have examined how well patients with focalhippocampal damage perform on FC-OLAP tests whereparticipants are cued with one pair item and have to saywhich of two items was paired with that item at study.The Vargha-Khadem et al. (1997) study used unrelatedword pairs, nonword pairs, familiar face pairs, and unfa-miliar face pairs as stimuli, and the Mayes et al. (2001)study used unrelated word pairs as stimuli. In all ofthese studies, the hippocampally-lesioned patients wereunimpaired. This is consistent with the model’s predic-tion that patients should perform relatively well, com-pared to controls, on FC-OLAP tests. Furthermore, thepatients’ excellent performance on these tests despitehaving hippocampal lesions, coupled with the fact thatthe tests used novel pairings, provides clear evidencethat cortex is capable of forming new conjunctive rep-


resentations (that are strong enough to support recogni-tion, if not recall) after a single study exposure. Onecaveat is that, while MTLC appears capable of support-ing good associative recognition performance when theto-be-associated stimuli are the same kind (e.g., words),it performs less well when the to-be-associated stimuliare of different kinds (e.g., objects and locations; Hold-stock et al., 2002; Vargha-Khadem et al., 1997). Hold-stock et al. (2002) argue that MTLC familiarity can notsupport object-location associative recognition becauseobject and location information do not converge fully inMTLC.

As a final note: While the studies discussed abovefound patterns of spared and impaired recognition per-formance after focal hippocampal damage that are con-sistent with the model’s predictions, it is important tokeep in mind that many studies have found an “across-the-board” impairment in declarative memory tasks fol-lowing hippocampal damage. For example, Stark andSquire (2002) found that patients with hippocampal dam-age were impaired to a roughly equal extent on tests withnovel vs. re-paired lures. We address the question of whysome hippocampal patients show “across-the-board” vs.selective deficits in Simulation 8, below.

Simulation 6: Interference and List Strength

We now turn to the fundamental issue of interference:How does studying an item affect recognition of other,previously studied items? In this section, we first re-view general principles of interference in networks likeours that incorporate Hebbian LTP and LTD. We thenshow how a list strength interference manipulation (de-scribed in detail below) differentially affects discrimina-tion based on hippocampal recall vs. MTLC familiarity.

General Principles of Interference in our Mod-els

At the most general level, interference occurs in ourmodels whenever different input patterns have overlap-ping internal representations. In this situation, studyinga new pattern tunes the overlapping units so they aremore sensitive to the new pattern, and less sensitive tothe unique (discriminative) features of other patterns.

Figure 18 is a simple illustration of this tuning pro-cess. It shows a network with a single hidden unit thatreceives input from five input units. Initially, the hiddenunit is activated to a roughly equal extent by two differ-ent input patterns, A and B. Studying pattern A has twoeffects: Hebbian LTP increases weights to active inputfeatures, and Hebbian LTD decrements weights to inac-tive input features. These changes bolster the extent to

�� B

BEFORE studying �� sA and B

AFTER studying ��

��LTD

Figure 18: Illustration of how Hebbian learning causes inter-ference in our models. Initially (top two squares) the hiddenunit responds equally to patterns A and B. The effects of study-ing pattern A are shown in the bottom two squares. Studyingpattern A boosts weights to features that are part of pattern A(including features that are shared with pattern B), and reducesweights to features that are not part of pattern A. These changesresult in a net increase in responding to pattern A, and a net de-crease in responding to pattern B.

which pattern A activates the hidden unit. The effects oflearning on responding to pattern B are more complex:LTP boosts weights to features that are shared by pat-terns A and B, but LTD reduces weights to features thatare unique to pattern B.

If you train a network of this type on a large numberof overlapping patterns (e.g., several pictures of fish), thenetwork will become more and more sensitive to featuresthat are shared across the entire item set (e.g., the factthat all studied stimuli have fins), and it will become lessand less responsive to discriminative features of individ-ual stimuli (e.g., the fact that one fish has a large greenstriped dorsal fin). In the long run, this latter effect isharmful to recognition performance — if the network’ssensitivity to the unique, discriminative features of stud-ied items and lures hits floor, then the network will not beable to respond differentially to studied items and luresat test. However, in the absence of floor effects, the ex-tent to which recognition is harmed depends on the extentto which interference differentially affects responding tostudied items and lures. In the next section, we explorethis issue in the context of our two models.

List Strength Simulations

We begin our exploration of interference by simulat-ing how list strength affects recognition in the two mod-els; specifically, how does strengthening some list itemsaffect recognition of other (non-strengthened) list items?(Ratcliff, Clark, & Shiffrin, 1990)


Effect of List Strength onRecognition in MTLC

YN

d'

0.0

0.5

1.0

1.5

2.0

2.5

Effect of List Strength onRecognition in the Hippocampus


YN

d'

0.0

0.5

1.0

1.5

2.0

2.5

Weak interf.Strong interf.

Size of the LSE in MTLCand the Hippocampus


LSE

:W

eak

Int.

d' -

Str

ong

Int.

d'

-0.20.00.20.40.60.81.01.2

MTLCHippo

a) b) c)

.10 .20 .26 .33 .50.41 .10 .20 .26 .33 .50.41 .10 .20 .26 .33 .50.41


Weak interf.Strong interf.

Figure 19: Results of list strength simulations in the two models. (a) shows MTLC results; (b) shows hippocampal results; (c) re-plots data from (a) and (b) as list strength difference scores (weak interference � � - strong interference � � ) to facilitate comparisonacross models. For low-to-moderate levels of overlap (up to .26), there was a significant list strength effect in the hippocampalmodel but not in the cortical model; for higher levels of overlap there was a list strength effect in both models.

List Type Target Items Interference Itemsweak interf.: bike robot apple cat tree towelstrong interf.: bike robot apple cat tree towel

cat tree towelcat tree towel

Table 1: List strength compares weak interference lists withstrong interference lists. In both conditions, participants areasked to discriminate between targets (e.g., bike) and nonstud-ied lures (e.g., coin) at test.

Methods

In our list strength simulations, we compared twoconditions: a weak interference condition and a stronginterference condition. In the weak interference condi-tion, participants study target items once and interfer-ence items once. The strong interference condition is thesame except interference items are strengthened at study(e.g., by presenting them multiple times, or presentingthem for a longer duration). In both conditions, par-ticipants have to discriminate between target items andnonstudied lures at test. If strengthening of interferenceitems (in the strong interference condition) impairs targetvs. lure discrimination, relative to the weak interferencecondition, this is a list strength effect. A simple diagramof the procedure is provided in Table 1.

The study list was comprised of 10 target items,followed by 10 interference items. Interference itemstrength was manipulated by increasing the learning ratefor these items (from .01 to .02). In our models, strength-ening by repetition and strengthening by increasing thelearning rate have qualitatively similar effects; however,quantitatively, repetition has a larger effect on weights

(e.g., doubling the number of presentations leads to moreweight change than doubling the learning rate), becausethe initial study presentation alters how active units willbe on the next presentation, and greater activity leadsto greater learning (according to the Hebb rule). Wealso manipulated average between-item overlap (rangingfrom 10% to 50%) to see how this factor interacts withlist strength — intuitively, increasing overlap should in-crease interference.

Results

In our cortical network simulations (Figure 19a),there was no effect of list strength on recognition wheninput pattern overlap was relatively low (up to .26), butthe list strength effect (LSE) was significant for higherlevels of input overlap. In contrast, the hippocampal net-work showed a significant LSE for all levels of inputoverlap (Figure 19b); the size of the hippocampal LSEincreased with increasing overlap (except in the .5 over-lap condition, where the LSE was compressed by flooreffects). Figure 19c directly compares the size of theLSE in the two models.

We also measured the direct effect of strengthen-ing interference items on memory for those items; bothmodels exhibited a robust item strength effect wherebymemory for interference items was better in the stronginterference condition (e.g., for 20% input overlap,interference-item � � increased from 2.13 to 3.22 in thehippocampal model; in the cortical model, � � increasedfrom 2.08 to 3.12), thereby confirming that our strength-ening manipulation was effective.

The data are puzzling: For moderate amounts ofoverlap, the hippocampus shows an interference effectdespite its ability to carry out pattern separation, and cor-


Effect of List Strength on Studied Recall

Recall

0.0 0.2 0.4 0.6 0.8 1.0

Fre

quen

cy

0.00

0.04

0.08

0.12

0.16

0.20

Weak interferenceStrong interference

Figure 20: Studied recall histograms for the strong and weakinterference conditions, 20% overlap condition. Increasing liststrength pushes the studied recall distribution to the left (to-wards zero).

tex — which has higher baseline levels of pattern overlap— does not show an interference effect. We address thehippocampal results first.

Interference in the Hippocampal Model

Understanding the hippocampal list strength effect isquite straightforward. Even though there is less overlapbetween representations in the hippocampus than in cor-tex, there is still some overlap (primarily in CA1, butalso in CA3). These overlapping units cause interfer-ence — specifically, recall of discriminative features ofstudied items is impaired through Hebbian LTD occur-ring in the CA3-CA1 projection and (to a lesser extent)in projections coming in to CA3. Importantly, for low-to-moderate levels of input pattern overlap, the amountof recall triggered by lures is at floor, and therefore itcan not decrease as a function of interference. Puttingthese two points together, the net effect of interference isto move the studied distribution downwards towards theat-floor lure distribution, which increases the overlap be-tween distributions and therefore impairs discriminabil-ity (Figure 20).

Interference in the Cortical Model

Next, we need to explain why a LSE was not obtainedin the cortical model (for low-to-moderate levels of inputoverlap). The critical difference between the cortical andhippocampal models is that lure familiarity is not at floorin the cortical network, thereby opening up the possibil-ity that lure familiarity (as well as studied familiarity)

Effect of List Strength on Raw Familiarity, 20% Overlap

List Strength (Interference Lrate / Target Lrate)

0 5 10 15 20 25 30

Fam

iliar

ity

0.71

0.72

0.73

0.74

0.75

0.76Studied itemsLures

Effect of List Strength on the Studied - Lure Familiarity Gap


0 5 10 15 20 25 30

Fam

iliar

ityD

iffer

ence

0.00

0.01

0.02

0.03

0.04

a)

b)

Figure 21: (a) Plot of how target and lure familiarity are af-fected by list strength (with 20% input overlap). Initially, targetand lure familiarity decrease; however, with enough interfer-ence, target and lure familiarity start to increase. (b) Plot ofthe difference in target and lure familiarity, as a function of liststrength; initially, the difference increases slightly, but then itdecreases.

might decrease as a function of interference. Discrim-inability is a function of the difference in studied andlure familiarity (as well as the variance of these distribu-tions); therefore, if lure familiarity decreases as much as(or more) than studied familiarity as a function of inter-ference, overall discriminability may be unaffected. Thisis in fact what occurs in the cortical model.

Figure 21a shows how raw familiarity scores trig-gered by targets and lures change as a function of in-terference. Initially, both studied and lure familiarity de-crease; this occurs because interference reduces weightsto discriminative features of studied items and lures.There is also an interaction whereby lure representa-tions degrade more quickly than studied representations,so the studied-lure gap in familiarity increases slightly(Figure 21b; note that while the increase is numericallysmall, it is highly significant).

However, with more interference, the studied-lure


gap in familiarity starts to decrease. This occurs becauseweights to discriminative features of lures begin to ap-proach floor (and — as such — can not show any ad-ditional decrease as a function of interference). Also,raw familiarity scores start to increase; this occurs be-cause, in addition to reducing weights to discriminativefeatures, interference also boosts weights to shared itemfeatures. At first, the former effect outweighs the latter,and there is a net decrease in familiarity. However, whenweights to discriminative features approach floor, theseweights no longer decrease enough to offset the increasein weights to shared features, and there is a net increasein familiarity.

In this simulation, list strength did not substantiallyaffect the variance of the familiarity signal; variance in-creased numerically with increasing list strength but theincrease was extremely small (e.g., 10x strengthening ofthe 10 interference items led to a 3% increase in vari-ance). However, we can not rule out the possibility thatadding other forms of variability to the model (and elimi-nating sampling variability; see the Sources of Variabilitysection) might alter the model’s predictions about howlist strength affects variance.

Differentiation

The finding that lure representations initially degradefaster than studied representations can be explained interms of the principle of differentiation, which was firstarticulated by Shiffrin et al. (1990); see also McClellandand Chappell (1998). Shiffrin et al. argued that studyingan item makes its memory representation more selective,such that the representation is less likely to be activatedby other items.

In our model, differentiation is a simple consequenceof Hebbian learning (as it is in McClelland & Chap-pell, 1998). As discussed above, Hebbian learning tunesMTLC units so that they are more sensitive to the stud-ied input pattern (because of LTP), and less sensitive toother, dissimilar input patterns (because of LTD). Be-cause of this LTD effect, studied-item representations areless likely to be activated by interference items than lure-item representations; as such, studied items suffer lessinterference than lures. As an example of how studyingan item pulls its representation away from other items,in the 20% input overlap simulations the average amountof MTLC overlap between studied target items and inter-ference items (expressed in terms of vector dot product)was .150, whereas the average overlap between lures andinterference items was .154; this difference was highlysignificant.

Boundary Conditions on the Null LSE

It should be clear from the above explanation thatwe do not always expect a null list strength effect for

Effect of List Strength on Raw Familiarity, 40.5% Overlap


0 5 10 15 20 25 30

Fam

iliar

ity

0.71

0.72

0.73

0.74

0.75

0.76

0.77

Studied ItemsLures

Figure 22: Plot of how target and lure familiarity are affectedby list strength with 40.5% input overlap. When overlap ishigh, target and lure familiarity increase right from the start,and the target-lure familiarity gap monotonically decreases.

MTLC familiarity. With enough interference, the corti-cal model’s overall sensitivity to discriminative featuresalways approaches floor and the studied and lure famil-iarity distributions converge. The amount of overlap be-tween items determines how quickly the network arrivesat this degenerate state — more overlap yields faster de-generation. When overlap is high, raw familiarity scoresincrease (and the familiarity gap decreases) right fromthe start; this is illustrated in Figure 22, which plots tar-get and lure familiarity as a function of list strength, for40.5% input overlap.


The main novel prediction from our models is that(modulo the boundary conditions outlined above) recog-nition based on hippocampal recall should exhibit a LSE,whereas recognition based on MTLC familiarity shouldnot.

Consistent with the hippocampal model’s prediction,some studies have found a LSE for cued recall (e.g.,Kahana & Rizzuto, submitted; Ratcliff et al., 1990) al-though not all studies that have looked for a cued re-call LSE have found one (e.g., Bauml, 1997). How-ever, practically all published studies that have lookedfor a LSE for recognition have failed to find one (Ratcliffet al., 1990; Murnane & Shiffrin, 1991a, 1991b; Ratcliff,Sheu, & Gronlund, 1992; Yonelinas, Hockley, & Mur-dock, 1992; Shiffrin, Huber, & Marinelli, 1995). Al-though this null LSE for recognition is consistent withour cortical model’s predictions (viewed in isolation), itis nevertheless somewhat surprising that overall recogni-tion scores do not reflect the hippocampal model’s ten-dency to produce a recognition LSE.

One way to reconcile the null LSE for recognition


List Strength Effect for Recall,Familiarity, and Old/New Recognition

d'(R) d'(F) d'(Old)

LSE

(Wea

kIn

t.d'

-S

tron

gIn

t.d'

)

-0.4

-0.2

0.0

0.2

0.4

0.6

Figure 23: Plot of the size of the list strength effect forthree dependent measures: � �� , recall-based discrimination;

� � �� , familiarity-based discrimination; and � � �� , discrim-ination computed based on “old/new” responses. Error barsindicate 95% confidence intervals. The LSE was significant for� � �� , but not for � � �� or � � �� .

with the model’s predictions is to argue that hippocampalrecall was not making enough of a contribution, relativeto MTLC familiarity, on existing tests. This explanationleads to a clear prediction: List strength effects shouldbe obtained for recognition tests and measures that loadmore heavily on the recall process. Norman (in press)tested this prediction in two distinct ways.

Self-Report Measures

In one experiment, Norman (in press) collected self-report measures of recall and familiarity — whenevera participant called an item “old”, there were askedwhether they specifically “remembered” details fromwhen the item was studied, whether the item just seemed“familiar”. To estimate recall-based discrimination, weplugged the probability of saying “remember” to studieditems and lures into the formula for � � . We computedfamiliarity-based discrimination using the IndependenceRemember-Know technique described in Jacoby, Yoneli-nas, and Jennings (1997).

Norman (in press) found that the effect of list strengthon old/new recognition sensitivity was nonsignificant inthis experiment, replicating the null LSE obtained byRatcliff et al. (1990). However, if you break recogni-tion into its component processes, it is clear that liststrength does affect performance (Figure 23). As pre-dicted, there was a significant LSE for discriminationbased on recall; in contrast, list strength had no effectwhatsoever on familiarity-based discrimination. 2 We

2Norman (in press) did not report how list strength affects

should emphasize that the technique we used to estimatefamiliarity-based discrimination assumes that recall andfamiliarity are independent — the independence assump-tion is discussed in more detail in Simulation 7. Also,as discussed by Norman (in press), we can not conclu-sively rule out (in this case) the possibility that the ob-served LSE for recall-based discrimination was causedby shifting response bias, with no real change in sensitiv-ity. Nonetheless, the overall pattern of results is highlyconsistent with the predictions of our model.

Lure Relatedness

Norman (in press) also looked at how list strengthaffects discrimination in the plurals paradigm (Hintz-man et al., 1992), where participants have to discrimi-nate between studied words, related switched-plurality(SP) lures (e.g., study “scorpion”, test “scorpions”), andunrelated lures. The model predicts that the ability todiscriminate between studied words and related SP luresshould depend on recall (see Simulation 3. Thus, weshould find a significant list strength effect for studiedvs. SP discrimination, but not necessarily for studied vs.unrelated discrimination, which can also be supported byfamiliarity.

Furthermore, we can also look at SP vs. unrelatedpseudodiscrimination, i.e., how much more likely areparticipants to say old to related vs. unrelated lures.Familiarity boosts pseudodiscrimination (insofar as SPlures will be more familiar than unrelated lures), but re-call of plurality information lowers pseudodiscrimina-tion (by allowing participants to confidently reject SPlures). Hence, if increasing list strength reduces recall ofplurality information (but has no effect familiarity-baseddiscrimination) the net effect should be an increase inpseudodiscrimination (a negative LSE).

As predicted, the LSE for studied vs. SP lure discrim-ination was significant; the LSE for studied vs. unrelatedlure discrimination was nonsignificant; and there was asignificant negative LSE for SP lure vs. unrelated lurepseudodiscrimination (Figure 24).

In summary: We obtained data consistent with themodel’s prediction of a LSE for recall-based discrimi-nation, and a null LSE for familiarity-based discrimina-tion, using two very different methods of isolating thecontributions of these processes (collecting self-reportdata, and using related lures). The model’s predictionsregarding the boundary conditions of the null LSE forfamiliarity-based discrimination remain to be tested. Forexample, our claim that discrimination asymptoticallygoes to zero with increasing list strength implies that itshould be possible to obtain a LSE (in paradigms that

familiarity-based discrimination — these results are being presentedfor the first time here.


Effect of List Strength on Discrimination ofStudied Items (S), Related Lures (SP), and

Unrelated Lures (U)

S vs. SP S vs. U SP vs. ULSE

(Wea

kIn

t.A

z-

Str

ong

Int.

Az)

-0.12

-0.08

-0.04

0.00

0.04

0.08

Figure 24: Results from the plurals LSE experiment. In thisexperiment, recognition sensitivity was measured using �� (anindex of the area under the ROC curve; Macmillan & Creel-man, 1991). The graph plots the size of the list strength effectfor three different kinds of discrimination: Studied vs. relatedswitched-plurality lures (S vs. SP); studied vs. unrelated lures(S vs. U); and related vs. unrelated lure pseudodiscrimination(SP vs. U). Error bars indicate 95% confidence intervals. Therewas a significant LSE for studied vs. related lure discrimina-tion, and there was a significant negative LSE for related vs.unrelated lure pseudodiscrimination.

have previously yielded a null LSE) by increasing thenumber of training trials for interference items.

List Length Effects

Thus far, our interference simulations have focusedon list strength. Here, we briefly discuss how list lengthaffects performance in the two models. In contrast tothe list strength paradigm, which measures how strength-ening items already on the list interferes with memoryfor other items, the list length paradigm measures howadding new (previously nonstudied) items to the list in-terferes with memory for other items.

The basic finding is that our model, as currently con-figured, makes the same predictions regarding list lengthand strength effects — adding new items and strengthen-ing already-presented items induce a comparable amountof weight change and therefore result in comparable lev-els of interference. As with list strength, the model pre-dicts a dissociation whereby (so long as overlap betweenitems is not too high) list length affects discriminationbased on hippocampal recall but not MTLC familiar-ity. Yonelinas (1994) obtained evidence consistent withthis prediction using the process dissociation procedure(which assumes independence).

However, some extant evidence appears to contradict

the model’s prediction of parallel effects of list lengthand list strength. In particular, Murnane and Shiffrin(1991a) and others have obtained a dissociation wherebylist length impairs recognition sensitivity but a closelymatched list strength manipulation does not. This findingimplies that adding new items to the list is more disrup-tive to existing weight patterns than repeating already-studied items. We are currently exploring different waysof implementing this dynamic in our model.

Fast Weights

One particularly promising approach is to add a large,transient fast weight component to the model that reachesits maximum value in one study trial and then decays ex-ponentially; subsequent presentations of the same itemsimply reset fast weights to their maximum value, anddecay begins again (see Hinton & Plaut, 1987 for anearly implementation of a similar mechanism). This dy-namic is consistent with neurobiological evidence show-ing a large, transient component to long-term poten-tiation (LTP) (e.g., Malenka & Nicoll, 1993; Bliss &Collingridge, 1993).

The assumptions outlined above imply that the mag-nitude of fast weights at test (for a particular item) willbe a function of the time elapsed from the most recentpresentation of the item; the number of times that theitem was repeated before this final presentation does notmatter. As such, presenting interference items for thefirst time (increasing list length) should have a large ef-fect on fast weights, but repeating already-studied in-terference items (increasing list strength) should haveless of an effect on the configuration of fast weights attest. Preliminary cortical-model simulations incorporat-ing fast weights (in addition to standard, non-decayingweights) have shown a list length/strength dissociation,but more work is needed to explore the relative meritsof different implementations of fast weights (e.g., shouldthe fast weight component incorporate LTD as well asLTP), and how the presence of fast weights interacts withthe other predictions outlined in this paper.

The idea that list length effects are attributable toquickly decaying weights implies that interposing a de-lay between study and test (thereby allowing the fastweights to decay) should greatly diminish the list lengtheffect. In keeping with this prediction, a recent studywith a 5 minute delay between study and test did notshow a list length effect (Dennis & Humphreys, 2001).The next step in testing this hypothesis will be to run ex-periments that parametrically manipulate study-test lagwhile measuring list length effects.


Simulation 7: The Combined Model and theIndependence Assumption

Up to this point, we have explored the properties ofhippocampal recall and MTLC familiarity by presentinginput patterns to separate hippocampal and neocorticalnetworks — this approach is useful for analytically map-ping out how the two networks respond to different in-puts, but it does not allow us to explore interactions be-tween the two networks. One important question thatcannot be addressed using separate networks is the sta-tistical relationship between recall and familiarity. Asmentioned several times throughout this paper, all ex-tant techniques for measuring the distinct contributionsof recall and familiarity to recognition performance as-sume that they are stochastically independent (e.g., Ja-coby et al., 1997). This assumption can not be directlytested using behavioral data because of the chicken-and-egg problems described in the Introduction.

To assess the validity of the independence assump-tion, we implemented a combined model in which thecortical system serves as the input to the hippocampus— this arrangement more accurately reflects how the twosystems are connected in the brain. Using the combinedmodel, we show that there is no simple answer regardingwhether or not hippocampal recall and MTLC familiar-ity are independent. Rather, the extent to which theseprocesses are correlated is a function of different (situa-tionally varying) factors, some of which boost the corre-lation and some of which reduce the correlation. In thissection, we briefly describe the architecture of the com-bined model, and then we use the model to explore twomanipulations that push the recall-familiarity correlationin opposite directions: encoding variability and interfer-ence (list length).

Combined Model Architecture

The combined model is structurally identical to thehippocampal model except the projection from Input toEC in has modifiable connections (and 25% random con-nectivity) instead of fixed 1-to-1 connectivity. Thus, theInput-to-EC in part of the combined model has the samebasic architecture and connectivity as the separate corti-cal model. This makes it possible to read out our ��

� ��

familiarity measure from the EC in layer of the com-bined model (i.e., the EC in layer of the combined modelserves the same functional role as the MTLC layer of theseparate cortical network).

There are, however, a few small differences betweenthe cortical part of the combined model, and the separatecortical network. First, the EC in layer of the combinedmodel is constrained to learn “slotted” representations,where only one unit in each 10-unit slot is strongly ac-

Recall-Familiarity Correlation for Studied Items

P(Encoding Failure)

0.0 0.1 0.2 0.3 0.4 0.5

Cor

rela

tion

0.0

0.1

0.2

0.3

0.4

0.5

Figure 25: Simulations exploring the effect of encoding vari-ability on the recall-familiarity correlation for studied items.As encoding variability (operationalized as the probability thatparticipants will “blink” and fail to encode half of an item’sfeatures at study) increases, the recall-familiarity correlationincreases.

tive; limiting the range of possible EC representationsmakes it easier for the hippocampus to learn a stablemapping between CA1 representations and EC represen-tations. Second, the EC in layer for the combined modelonly has 240 units, compared to 1920 units in the MTLClayer of the separate cortical network. This reduced sizederives from computational necessity — use of a largerEC in would require a larger CA1, which together wouldmake the simulations run too slowly on current hard-ware. This smaller hidden layer in the combined modelmakes the familiarity signal more subject to samplingvariability, and thus recognition �� is somewhat worse,but otherwise it functions just as before. We used thesame basic cortical and hippocampal parameters as in ourseparate-network simulations, except we used input pat-terns with 32.5% overlap — this level of input overlapyields approximately 24% overlap between EC in pat-terns at study.

In the combined model, the absolute size of therecall-familiarity correlation is likely to be inflated, rela-tive to the brain, because of the high level of samplingvariability present in our model. Sampling variabilityleads to random fluctuations in the sharpness of corticalrepresentations, which induce a correlation (insofar assharper representations trigger larger familiarity scores,and they also propagate better into the hippocampus, bol-stering recall). Because of this issue, our simulationsbelow focus on identifying manipulations that affect thesize of the correlation, rather than characterizing the ab-solute size of the correlation.


Recall-Familiarity Correlation for Studied Items

List Length

10 40 70 100 130 160

Cor

rela

tion

0.1

0.2

0.3

Raw Familiarity and Recall Scores for Studied Items

List Length

10 40 70 100 130 160

Fam

iliar

ity

0.66

0.68

0.70

0.72

Rec

all

0.0

0.2

0.4

0.6

0.8

1.0

Studied familiarityStudied recall

a) b)

Figure 26: Simulations exploring how interference affects the recall-familiarity correlation. (a) Increasing list length reducesthe recall-familiarity correlation for studied items. (b) Increasing list length leads to a decrease in studied-item recall scores, butstudied-item familiarity scores stay relatively constant.

Effects of Encoding Variability

Curran and Hintzman (1995) pointed out that encod-ing variability can boost the recall-familiarity correla-tion; if items vary substantially in how well they areencoded, poorly-encoded items will be unfamiliar andwill not be recalled; well-encoded items will be morefamiliar, and more likely to trigger recall. We ran sim-ulations in the combined model where we manipulatedencoding variability by varying the probability of par-tial encoding failure (i.e., encoding only half of an item’sfeatures) from 0 to .50. For each simulated participant,we read out the MTLC familiarity signal ( ��

� ��, read

out from the EC in layer) and the hippocampal recall sig-nal (

��

�� between EC out and EC in) on

each trial, and measured the correlation between thesesignals (across trials). The results of these simulationsare presented in Figure 25; as expected, increasing en-coding variability increased the recall-familiarity corre-lation in the model.

Interference Induced Decorrelation

In the next set of simulations, we show how the pres-ence of interference between memory traces can reducethe recall-familiarity correlation. In the Interference andList Strength section of the paper, we discussed how thetwo systems are differentially affected by interference:Hippocampal recall scores for studied items tend to de-crease with interference; familiarity scores decrease less,and sometimes increase, because increased sensitivity toshared prototype features compensates for lost sensitivity

to discriminative item-specific features. Insofar as itemsvary in how much interference they are subject to (dueto random differences in between-item overlap), and in-terference pushes recall and familiarity in different di-rections, it should be possible to use interference as a“wedge” to decorrelate recall and familiarity.

We ran simulations measuring how the recall-familiarity correlation changed as a function of interfer-ence (operationalized using a list length manipulation).There were 10 target items followed by 0 to 150 (non-tested) interference items. As expected, increasing listlength lowers the recall-familiarity correlation for stud-ied items (Figure 26a) in the model.

We can get further insight into these results by look-ing at how interference affects raw familiarity and re-call scores for studied items in these simulations (Fig-ure 26b). With increasing interference, recall decreasessharply but familiarity stays relatively constant; this dif-ferential effect of interference works to de-correlate thetwo signals. Once recall approaches floor, recall and fa-miliarity are affected in a basically similar manner (i.e.,not much); this lack of a differential effect explains whythe recall-familiarity correlation does not continue to de-crease all the way to zero.

In summary: The combined model gives us a princi-pled means of predicting how different factors will affectthe statistical relationship between recall and familiarity.Given the tight coupling of the cortical and hippocampalnetworks in the combined model, one might think thata positive correlation is inevitable. However, the resultspresented here show that — to a first approximation —


independence can be achieved, so long as factors that re-duce the correlation (e.g., interference) exert more of aninfluence than factors that bolster the correlation (e.g.,encoding variability). Future research will focus on iden-tifying additional factors that affect the size of the corre-lation.

Simulation 8: Lesion Effects in the CombinedModel

In this section, we show how the combined modelcan provide a more sophisticated understanding of theeffects of different kinds of medial temporal lesions.Specifically, we show that (in the combined model)partial hippocampal lesions can sometimes result inworse overall recognition performance than completehippocampal lesions. In contrast, increasing MTLC le-sion size monotonically reduces overall recognition per-formance.

Partial Lesion Simulation Methods

We ran one set of simulations exploring the effects offocal hippocampal lesions, and another set of simulationsexploring the effects of focal MTLC lesions. In all of thelesion simulations, the size of the lesion (in terms of %of units removed) was varied from 0% to 95% in 5%increments. In the hippocampal lesion simulations, welesioned all of the hippocampal subregions (DG, CA1,CA3) equally by percentage; in the MTLC lesion sim-ulations, we lesioned EC in. To establish comparablebaseline (pre-lesion) recognition performance betweenthe hippocampal and cortical networks, we boosted thecortical learning rate to .012 instead of .004; this increasecompensates for the high amount of sampling variabilitypresent in the (240 unit) EC in layer of the combinedmodel. In these simulations, we used forced-choice test-ing to maximize comparability with animal lesion studiesthat used this format (e.g., Baxter & Murray, 2001b).

To simulate overall recognition performance whenboth processes are contributing, we used a decision rulewhereby if one item triggered a larger positive recallscore (

��

�� ) than the other, then that item

was selected; otherwise, if��

�� ,

the decision falls back on familiarity. This rule incorpo-rates the assumption (shared by other dual-process mod-els, e.g., Jacoby et al., 1997) that recall takes precedenceover familiarity. This differential weighting of recallcan be justified in terms of our finding that, in normalcircumstances, hippocampal recall in the CLS model ismore diagnostic than MTLC familiarity. Furthermore,it is supported by data showing that recall is associatedwith higher average recognition confidence ratings than

Effect of Hippocampal Damage on FC Recognition

Lesion Size (Percent)

0 20 40 60 80 100

Pro

port

ion

Cor

rect

0.5

0.6

0.7

0.8

0.9

1.0

CombinedHippocampal recallMTLC familiarity

Figure 27: Effect of hippocampal damage on forced-choice(FC) recognition performance. This graph plots FC accuracybased on MTLC familiarity, hippocampal recall, and a combi-nation of the two signals, as a function of hippocampal lesionsize. FC accuracy based on MTLC familiarity is unaffected byhippocampal lesion size. FC accuracy based on hippocampalrecall declines steadily as lesion size increases. FC accuracybased on a combination of recall and familiarity is affected ina nonmonotonic fashion by lesion size: initially, combined FCaccuracy declines; however, going from a 75% to 100% hip-pocampal lesion leads to an increase in combined FC perfor-mance.

familiarity (e.g., Yonelinas, 2001b).

Hippocampal Lesions

Figure 27 shows how hippocampal FC performance,cortical FC performance, and combined FC performance(using the decision rule just described) vary as a functionof hippocampal lesion size. As one might expect, hip-pocampal FC performance decreases steadily as a func-tion of hippocampal lesion size, while cortical perfor-mance is unaffected by hippocampal damage (insofar asfamiliarity is computed before activity feeds into the hip-pocampus). The most interesting result is that hippocam-pal lesion size has a nonmonotonic effect on combinedrecognition performance. At first, combined FC accu-racy decreases with increasing hippocampal lesion size;however, going from a 75% to 100% hippocampal lesionactually improves combined FC performance.

Why is recognition performance worse for partial(75%) vs. complete lesions? The key to understandingthis finding is that partial hippocampal damage impairsthe hippocampus’ ability to carry out pattern separation.We assume that lesioning a layer lowers the total numberof units, but does not decrease the number of active units;accordingly, representations become less sparse and theaverage amount of overlap between patterns increases.There is neurobiological support for this assumption: In


Effect of Hippocampal Damage on Studied and Lure Recall

Lesion Size (Percent)

0 20 40 60 80 100

P(R

ecal

l>0)

0.0

0.2

0.4

0.6

0.8

1.0

StudiedLure

Figure 28: Plot of the probability that studied items and lureswill trigger a positive �� score, as a as a func-tion of hippocampal lesion size. For studied items, the proba-bility declines monotonically as a function of lesion size. Forlures, the probability first increases, then decreases, as a func-tion of lesion size.

the brain, the activity of excitatory pyramidal neuronsis regulated primarily by inhibitory interneurons (Dou-glas & Martin, 1990); assuming that both excitatory andinhibitory neurons are damaged by lesions, this loss ofinhibition is likely to result in a proportional increase inactivity for the remaining excitatory neurons.

Pattern separation failure (induced by partial dam-age) results in a sharp increase in the amount of recalltriggered by lures. Combined recognition performancein these participants suffers because the noisy recall sig-nal drowns out useful information that is present in thefamiliarity signal. Moving from a partial hippocampallesion to a complete lesion improves performance by re-moving this source of noise.

Figure 28 provides direct support for the “noisy hip-pocampus” theory; it shows how the probability thatlures will trigger a positive

��

�� score in-

creases steadily with increasing hippocampal lesion sizeuntil it reaches a peak of .41 (for 55% hippocampal dam-age). However, as lesion size approaches 60%, the prob-ability that lures and studied items will trigger a positive��

�� score starts to decrease. This oc-

curs for two reasons: First, because of pattern separationfailure in CA3, all items start to trigger recall of the sameprototypical features (which mismatch item-specific fea-tures of the test probe); second, CA1 damage reducesthe hippocampus’ ability to translate CA3 activity intothe target EC pattern. As the number of items triggeringpositive

��

�� scores decreases, control of

the recognition decision reverts to familiarity. This ben-efits recognition performance insofar as familiarity doesa better job of discriminating between studied items andlures than the recall signal generated by the lesioned hip-

Effect of MTLC Damage on FC Recognition

MTLC Lesion Size (Percent)

0 20 40 60 80 100

Pro

port

ion

Cor

rect

0.5

0.6

0.7

0.8

0.9

1.0

CombinedHippo recallMTLC familiarity

Figure 29: Effect of MTLC (specifically, EC in) damage onforced-choice (FC) recognition performance. This graph plotsFC accuracy based on MTLC familiarity, hippocampal recall,and a combination of the two signals, as a function of EC inlesion size. All three accuracy scores (recall-alone, familiarity-alone, and combined) decline steadily with increasing lesionsize.

pocampus.

MTLC Lesions

Turning to the effects of MTLC lesions, we findthat lesioning EC in hurts both cortical and hippocampalrecognition performance (Figure 29). Therefore, com-bined recognition performance decreases steadily as afunction of MTLC lesion size. The observed deficits re-sult from the fact that overlap between EC in patterns in-creases with lesion size — this has a direct adverse effecton cortically-based discrimination; furthermore, becauseEC in serves as the input layer for the hippocampus, in-creased EC in overlap leads to increased hippocampaloverlap, which hurts recall.

Relevant Data and Implications

The simulation results reported above are consistentwith the results of a recent meta-analysis conducted byBaxter and Murray (2001b). This meta-analysis incor-porated results from several studies that have looked athippocampal and perirhinal (MTLC) lesion effects onrecognition in monkeys using a delayed-nonmatching-to-sample (DNMS) paradigm. Baxter and Murray (2001b)found, in keeping with our results, that partial hippocam-pal lesions can lead to larger recognition deficits thanmore complete lesions — in the meta-analysis, hip-pocampal lesion size and recognition impairment werenegatively correlated. The Baxter and Murray meta-analysis also found that perirhinal lesion size and recog-


nition impairment were positively correlated, just as wefound that MTLC lesion size and impairment were cor-related in our simulations.

Baxter and Murray’s results are highly controversial;Zola and Squire (2001) re-analyzed the data in the Bax-ter and Murray meta-analysis, using a different set ofstatistical techniques that control, e.g., for differences inmean lesion size across studies, and found that the nega-tive correlation between hippocampal lesion size and im-pairment reported by Baxter and Murray (2001b) was nolonger significant (although there was still a nonsignifi-cant trend in this direction; see Baxter & Murray, 2001afor further discussion of this issue). Our results con-tribute to this debate by providing a novel and principledaccount of how a negative correlation might come intobeing, in terms of pattern separation failure resulting ina noisy recall signal that participants nevertheless relyupon when making recognition judgments. Of course,our account is not the only way of explaining why par-tial hippocampal lesions might impair recognition morethan complete lesions. Another possibility, suggested byMumby, Wood, Duva, Kornecook, Pinel, and Phillips(1996), is that a damaged hippocampus might disruptneocortical processing via epileptiform activity.

The results reported here speak to the debate overwhy hippocampally lesioned patients sometimes showrelatively spared recognition performance on standardrecognition tests with unrelated lures, and sometimes donot. The model predicts that patients with relativelycomplete hippocampal lesions (that nonetheless spareMTLC) should show relatively spared performance, andpatients with smaller hippocampal lesions (that reducethe diagnosticity of recall without eliminating the signaloutright) should show an “across-the-board” declarativememory deficit without any particular sparing of recog-nition. This view implies that the range of etiologies thatproduce selective sparing should be quite narrow: If thelesion is too small, you end with a partially-lesioned hip-pocampus that injects noise into the recognition process;if the lesion is too large, then you hit perirhinal cortex (inaddition to the hippocampus) and this leads to deficits infamiliarity-based recognition.

General Discussion

In this section of the paper, we review how our mod-eling work speaks to extant debates over how to charac-terize the respective contributions of recall (vs. familiar-ity) and hippocampus (vs. MTLC) to recognition. Next,we compare our model to other neurally-inspired and ab-stract computational models of recognition. We concludeby discussing limitations of the models and future direc-tions for research.

Implications for Theories of How Recall Con-tributes to Recognition

As discussed in the Introduction, the dual-process ap-proach to recognition has become increasingly prevalentin recent years, but this enterprise is based on a weakfoundation. Yonelinas, Jacoby, and their colleagues havedeveloped several techniques for measuring the contribu-tions of recall and familiarity based on behavioral data,but these techniques rely on a core set of assumptionsthat have not been tested; furthermore, some of these as-sumptions (e.g., independence) can not be tested basedon behavioral data alone, because of chicken-and-eggproblems.

More specifically, measurement techniques such asprocess dissociation and ROC analysis are built arounddual-process signal detection theory (Jacoby et al., 1997;Yonelinas, 2001b). In addition to the independence as-sumption, this theory assumes that familiarity is a Gaus-sian signal-detection process, but recall is a dual high-threshold process (i.e., recall is all-or-none; studied itemsare sometimes recalled as being “old”, but lures neverare; lures are sometimes recalled as being “new”, butstudied items never are). The CLS model does not as-sume any of the above claims to be true — rather, itscore assumptions are based on the functional proper-ties of hippocampus vs. MTLC. As such, we can usethe CLS model to evaluate the validity of dual-processsignal-detection theory.

Results from our cortical model are generally consis-tent with the idea that familiarity is a Gaussian signal-detection process. In contrast, results from our hip-pocampal model are not consistent with the idea that re-call is a high-threshold process — recall in our modelis not all-or-none, and lures sometimes trigger above-zero recall. Our claim that lures can trigger some match-ing recall, and that the number of recalled details variesfrom item to item, is consistent with the Source Moni-toring Framework set forth by Marcia Johnson and hercolleagues (e.g., Mitchell & Johnson, 2000).

Although the recall process in our model is notstrictly high-threshold, it is not Gaussian either. If over-lap between stimuli is not too high, our recall process isapproximately consistent with dual-high-threshold the-ory, in the sense that (for a given experiment) there isa level of matching recall that is sometimes exceeded bystudied items, but not by lures, and there is a level of mis-matching recall that is sometimes exceeded by lures, butnot by studied items. Furthermore, we assume that par-ticipants routinely set their recall criterion high enough toavoid false recognition of unrelated lures, and that partic-ipants set their criterion for recall-to-reject high enoughto avoid incorrect rejection of studied items. As such, themodel provides some support (conditional on overlap not


being too high) for Yonelinas’ claim that lures will not becalled “old” based on recall, and that studied items willnot be called “new” based on recall.

Regarding the independence assumption: As men-tioned earlier, Curran and Hintzman (1995) and othershave criticized this assumption on the grounds that somedegree of encoding variability is likely to be present inany experiment, and encoding variability results in a pos-itive recall-familiarity correlation. Our results confirmthis latter conclusion, but they also show that interfer-ence reduces the recall-familiarity correlation. As such,it is possible to achieve independence (assuming there isenough interference) even when encoding variability ispresent.

Overall, although the details of the CLS model’s pre-dictions are not strictly consistent with the assumptionsbehind dual-process signal detection theory, it is safe tosay that the CLS model’s predictions are broadly con-sistent with these assumptions; accordingly, we wouldexpect measurement techniques that rely on these as-sumptions to yield meaningful results most of the time.The main contribution of the CLS model is to establishboundary conditions on the validity of these assumptions— e.g., the assumption that lures will not be called “old”based on recall does not hold true when there is exten-sive overlap between stimuli, and the independence as-sumption is more likely to hold true in situations whereinterference is high and encoding variability is low thanwhen the opposite is true.

Implications for Theories of How the Hip-pocampus (vs. MTLC) Contributes to Recog-nition

To a first approximation, the CLS model resemblesthe neuropsychological theory of episodic memory setforth by Aggleton and Brown (1999) (A&B) — boththeories posit that the hippocampus is essential for re-call and that MTLC can support judgments of stimulusfamiliarity on its own. However, this resemblance isonly superficial. As discussed in the Introduction, simplylinking “recall” and “familiarity” to hippocampus andMTLC, without further specifying how these processeswork, does not allow us to predict when recognition willbe impaired or spared after hippocampal damage. Ev-erything depends on how you unpack the terms “recall”and “familiarity”, and the CLS and A&B theories unpackthese terms in completely different ways.

The A&B theory unpacks recall and familiarity interms of a simple, verbally stated dichotomy wherebyrecall is necessary for forming new associations but fa-miliarity is sufficient for item recognition. In contrast tothe A&B theory, the CLS model grounds its conception

of recall and familiarity in terms of very specific ideasabout the neurocomputational properties of hippocam-pus and MTLC. Moreover, we have implemented theseideas in a working computational model that can be usedto test their sufficiency and generate novel predictions.

According to the CLS model, practically all dif-ferences between cortical and hippocampal contribu-tions to recognition can be traced back to graded differ-ences in how information is represented in these struc-tures. For example, the fact that hippocampal repre-sentations are more sparse than cortical representationsin our model implies that hippocampal recall will bemore sensitive to conjunctions than MTLC familiarity,but crucially both signals should show some sensitivityto conjunctions (see Simulation 5 for more discussionof this point). This graded approach makes it possiblefor the CLS model to explain dissociations within sim-ple categories like “memory for new associations” —for example, the finding from Mayes et al. (2001) thathippocampally-lesioned patient YR sometimes shows in-tact FC word-word associative recognition (presumably,based on MTLC familiarity) despite being impaired atcued recall of newly learned paired associates (Mayeset al., 2002; see Vargha-Khadem et al., 1997 for a similarpattern of results). As discussed earlier, these dissocia-tions are problematic for A&B’s “item” vs. “associative”dichotomy.

Lastly, we should mention an important commonal-ity between the CLS model and the A&B theory: Boththeories predict sparing of item recognition (with unre-lated lures) relative to recall in patients with completehippocampal lesions. While some studies have foundthis pattern of results in patients with focal hippocam-pal damage (e.g., Mayes et al., 2002), it is potentiallyquite problematic for both models that other studies havefound roughly equivalent deficits in recognition (withunrelated lures) and recall following focal hippocampaldamage (e.g., Manns & Squire, 1999). We do not claimto fully understand why some hippocampally-lesionedpatients show relatively spared item recognition but oth-ers do not. However, it may be possible to account forsome of this variability in terms of the idea, set forth byBaxter and Murray (2001b) and explored in Simulation8, that partial hippocampal lesions are especially harmfulto recognition. This idea is still highly controversial andneeds to be tested directly, using experiments that carryout careful, parametric manipulations of lesion size (con-trolling for other factors like lesion technique and task).

Comparison with Abstract ComputationalModels of Recognition

Whereas our model incorporates explicit claimsabout how different brain structures (hippocampus and


MTLC) support recognition memory, most computa-tional models of recognition memory are abstract inthe sense that they do not make specific claims abouthow recognition is implemented in the brain. The REMmodel presented by Shiffrin and Steyvers (1997) repre-sents the state of the art in abstract modeling of recog-nition memory (see McClelland & Chappell, 1998 for avery similar model). REM carries out a Bayesian com-putation of the likelihood that an ideal observer shouldsay “old” to an item, based on the extent to which thatitem matches (and mismatches) stored memory traces.Our cortical familiarity model resembles REM and otherabstract models in several respects: Like the abstractmodels, our cortical model computes a scalar that tracksthe “global match” between the test probe and storedmemory traces; furthermore, both our cortical model andmodels like REM posit that differentiation (Shiffrin et al.,1990) contributes to the null LSE. Thus, our modelingwork relies critically on insights that were developed inthe context of abstract models like REM.

We can also draw a number of contrasts betweenthe CLS model and abstract Bayesian models. Onemajor difference is that our model posits that two pro-cesses (with distinct operating characteristics) contributeto recognition whereas abstract models attempt to ex-plain recognition data in terms of a single familiarity pro-cess. Another difference is that — in our model — inter-ference occurs at study, when one item re-uses weightsthat are also used by another item, whereas REM positsthat memory traces are stored in a non-interfering fash-ion, and that interference arises at test, whenever the testitem spuriously matches memory traces corresponding toother items.

The question of whether interference occurs at study(or only at test) has a long history in memory research.Other models that posit structural interference at studyhave found large and sometimes excessive effects of in-terference on recognition sensitivity. For example, Rat-cliff (1990) presented a model consisting of a 3-layerfeedforward network that learns to reproduce input pat-terns in the output layer; the dependent measure at test ishow well the recalled (output) pattern matches the input.Like our hippocampal model, the Ratcliff model showsinterference effects on recognition sensitivity because re-call of discriminative features of lures is at floor; as such,any decrease in studied recall necessarily pushes thestudied and lure recall distributions closer together. TheRatcliff model generally shows more interference thanour hippocampal model because there is more overlapbetween “hidden” representations in the Ratcliff modelvs. in our hippocampal model.

Based on these results (and others like them), Mur-nane and Shiffrin (1991a) concluded that interference-

at-study models may be incapable of explaining the nullrecognition LSE obtained by Ratcliff et al. (1990). Animportant implication of the work presented here is thatinterference-at-study models do not always show exces-sive effects of interference on recognition sensitivity; ourcortical model predicts a null recognition LSE becauseincreasing list strength reduces lure familiarity slightlymore than studied-item familiarity, so the studied-luregap in familiarity is preserved.

In this paper, we have focused on describing qual-itative model predictions, and the boundary conditionsof these predictions. Working at this level, it is clear thatthere are some fundamental differences in the predictionsof the CLS model vs. models like REM. Because study-ing one item degrades the memory traces of other items,our model predicts — regardless of parameter settings— that the curve relating interference (e.g., list lengthor list strength) to recognition sensitivity will alwaysasymptotically go to zero with increasing interference.In contrast, in REM it is possible to completely eliminatethe deleterious effects of interference items on perfor-mance through differentiation; if interference items arepresented often enough, they could become so stronglydifferentiated that the odds of them spuriously matchinga test item are effectively zero; whether or not this actu-ally happens depends on model parameters.

Comparison with Other Neural Network Mod-els of Hippocampal and Cortical Contributionsto Recognition

Models of Hippocampus

The hippocampal component of the CLS model ispart of a long tradition of hippocampal modeling (e.g.,Marr, 1971; McNaughton & Morris, 1987; Rolls, 1989;Levy, 1989; Touretzky & Redish, 1996; Burgess &O’Keefe, 1996; Wu, Baxter, & Levy, 1996; Treves &Rolls, 1994; Moll & Miikkulainen, 1997; Hasselmo &Wyble, 1997). Although different hippocampal modelsmay differ slightly in the functions they ascribe to par-ticular hippocampal subcomponents, a remarkable con-sensus has emerged regarding how the hippocampussupports episodic memory (i.e., by assigning minimallyoverlapping CA3 representations to different episodes,with recurrent connectivity serving to bind together theconstituent features of those episodes). In the presentmodeling work, we build on this shared foundation byapplying these biologically-based computational model-ing ideas to a rich domain of human memory data (for anapplication of the same basic model to animal learningdata, see O’Reilly & Rudy, 2001).

The Hasselmo and Wyble (1997) model (hereafter,H&W) deserves special consideration because it is the


only one of the aforementioned hippocampal models thathas been used to simulate patterns of behavioral list-learning data. The architecture of this model is gener-ally similar to the architecture of the CLS hippocampalmodel, except H&W make a distinction between itemand (shared) context information, and posit that itemand context information are kept separate throughout theentire hippocampal processing pathway, except in CA3where recurrent connections allow for item-context asso-ciations; furthermore, in the H&W model recognition isbased on the extent to which item representations triggerrecall of shared contextual information associated withthe study list. The H&W model predicts that recogni-tion of studied items should be robust to factors that de-grade hippocampal processing because — insofar as allstudied items have the same “context vector” — the CA3representation of shared context information will be verystrong and thus easy to activate. However, the fact thatthe CA3 context representation is easy to activate impliesthat related lures will very frequently trigger false alarmsin the H&W model (in contrast to the CLS model, whichpredicts low hippocampal false alarms to related lures).The H&W model also predicts a null list strength effectfor hippocampally-driven recognition, and a null maineffect of item strength on hippocampally-driven recogni-tion (in contrast to our model, which predicts that bothitem strength and list strength effects should be obtainedin the hippocampus). Thus, because H&W use a differ-ent hippocampal recognition measure, and separate itemand context representations, their model generates recog-nition predictions that are very different from the CLShippocampal model’s predictions. However, we shouldemphasize that, if H&W used the same recognition mea-sure as our model (match - mismatch) and factored itemrecall as well as context recall into recognition decisions,their model and the CLS model would likely make verysimilar predictions because the two model architecturesare so similar.

Models of Neocortical Contributions to RecognitionMemory

In recent years, several models (besides ours) havebeen developed that address the role of MTLC in famil-iarity discrimination; see Bogacz and Brown (in press)for a detailed comparison of the properties of differentfamiliarity-discrimination models. Some of these mod-els, like ours, posit that familiarity discrimination in cor-tex arises from Hebbian learning that tunes a popula-tion of units to respond strongly to the stimulus (e.g.,Bogacz, Brown, & Giraud-Carrier, 2001; Sohal & Has-selmo, 2000), although the details of these models differfrom ours (for example, the Bogacz et al., 2001 and So-hal & Hasselmo, 2000 models posit that both homosy-naptic Hebbian LTD — decrease weights if the sending

unit is active but the receiving unit is not — and heterosy-naptic Hebbian LTD — decrease weights if the receivingunit is active but the sending unit is not — are impor-tant for familiarity discrimination, whereas our modelonly incorporates heterosynaptic Hebbian LTD). Othermodels incorporate radically different mechanisms, e.g.,anti-Hebbian learning that reduces connection strengthsbetween active neurons (Bogacz & Brown, in press).

One difference between the models proposed by Bo-gacz & Brown vs. our model is that — in the Bogacz& Brown models — familiarity is computed by a spe-cialized population of novelty detector units that are notdirectly involved in extracting and representing stimulusfeatures (see Bogacz & Brown, in press for a review ofempirical evidence that supports this claim). In contrast,our combined model does not posit the existence of spe-cialized novelty detector units; rather, the layer (EC in)where we read out the ��

� ��familiarity signal con-

tains the highest (most refined) cortical representation ofthe stimulus, which in turn serves as the input to the hip-pocampus.

We should note, however, that our model could easilybe transformed into a model with specialized novelty de-tectors, without altering any of the predictions outlined inthe main body of the paper. To do this, we could connectthe input layer of the combined model directly to CA3,CA1, and DG, and we could have this input layer feed inparallel to a second cortical layer (with no connectionsto the hippocampus) where we compute ��

� ��. The

units in this second cortical layer could be labeled “spe-cialized novelty detectors” insofar as they are not servingany other important function in the model. This changewould not affect the functioning of the cortical part of themodel in any way.

Having the same layer serve as the input to the “nov-elty detection” layer and the hippocampus could, in prin-ciple, affect the predictions of the combined model, butas a practical matter none of the combined model predic-tions outlined in Simulations 7 and 8 would be affectedby this change. For example, if cortical units involvedin computing novelty/familiarity were not involved inpassing features to the hippocampus, then it would bepossible in principle to disrupt familiarity for a partic-ular stimulus without disrupting hippocampal recall ofthat stimulus (by lesioning the “novelty detector” units).However, this is probably not possible in practice — ac-cording to Brown and Xiang (1998), perirhinal neuronsinvolved in familiarity discrimination and stimulus rep-resentation are topographically interspersed, so lesionslarge enough to affect one population of neurons shouldalso affect the other.

A major issue raised by Bogacz and Brown (in press)is whether networks like ours that extract features via


Hebbian learning have adequate capacity to explain ourability to discriminate between very large numbers of fa-miliar and unfamiliar stimuli (e.g., Standing, 1973 foundthat people can discriminate between studied and non-studied pictures after studying a list of thousands of pic-tures). Bogacz and Brown (in press) argue that — even ina “brain-sized” version of our cortical model — the net-work’s tendency to represent shared (prototypical) fea-tures at the expense of features that discriminate betweenitems will result in unacceptably poor performance afterstudying large numbers of stimuli. Bogacz and Brown(in press) point out that the anti-Hebbian model that theypropose does not have this problem; this model ignoresfeatures that are shared across patterns and, as such, hasa much higher capacity. A possible problem with theanti-Hebbian model is that it may show too little inter-ference. More research is needed to assess whether ournetwork architecture, suitably scaled, can explain find-ings like Standing (1973), and — if not — how it can bemodified to accommodate this result (without eliminat-ing its ability to explain interference effects on recogni-tion discrimination).

At this point in time, it is difficult to directly com-pare our model’s predictions to the predictions of othercortical familiarity models because the models have beenapplied to different data domains — we have focused onexplaining detailed patterns of behavioral data, whereasthe other models have focused on explaining single-cellrecording data in monkeys. Bringing the different mod-els to bear on the same data points is an important topicfor future research. While the CLS model can not makedetailed predictions about spiking patterns of single neu-rons, it does make predictions regarding how firing rateswill change as a function of familiarity. For example,the model predicts that, for a particular stimulus, neuronsthat show decreased (vs. asymptotically strong) firing inresponse to repeated presentation of that stimulus shouldbe neurons that initially had a less strong response to thestimulus (and therefore lost the competition to representthe stimulus).

Future Directions

Future research will address limitations of the modelthat were mentioned earlier. In the Sources of Variabil-ity section, we discussed how the model incorporatessome sources of variability that we plan to remove (sam-pling variability) and lacks some sources of variabilitythat we plan to add. Increases in computer processingspeed will make it possible to grow our networks to thepoint where sampling variability is negligible, and wewill replace lost sampling variability by adding encodingvariability and variability in pre-experimental presenta-tion frequency to the model. Including pre-experimental

variability (by presenting test items in other contexts avariable number of times prior to the start of the experi-ment) will allow us to address a range of interesting phe-nomena, including the so-called frequency mirror effect,whereby hits tend to be higher for low-frequency (LF)stimuli than for high-frequency (HF) stimuli, but falsealarms tend to be higher for HF stimuli than LF stim-uli (see, e.g., Glanzer, Adams, Iverson, & Kim, 1993);recently, several studies have obtained evidence suggest-ing that recall is responsible for the LF hit-rate advan-tage and familiarity is responsible for the HF false-alarm-rate advantage (Reder, Nhouyvanisvong, Schunn, Ayers,Angstadt, & Hiraki, 2000; Joordens & Hockley, 2000;Reder et al. also present an abstract dual-process modelof this finding).

Furthermore, we plan to directly address the questionof how participants make decisions based on recall andfamiliarity. Clearly, people are capable of employing avariety of different decision strategies that can differen-tially weight the different signals that emerge from thecortex and hippocampus. One way to address this issue isto conduct empirical Bayesian analyses to delineate howthe optimal way of making recognition decisions in ourmodel varies as a function of situational factors, and thencompare the results of these analyses with participants’actual performance. A specific idea that we plan to ex-plore in detail is that participants discount recall of pro-totype information, because prototype recall is much lessdiagnostic than item-specific recall. The frontal lobesmay play an important part in this discounting process— for example, Curran, Schacter, Norman, and Galluc-cio (1997) studied a frontal-lesioned patient (BG) whofalse alarmed excessively to nonstudied items that wereof the same general type as studied items; one way ofexplaining this finding is that BG has a selective deficitin discounting prototype recall. Thus, the literature onfrontal lesion effects may provide important constraintson how recognition decision-making works, by showinghow it breaks down.

Supplementing the model with a more principledtheory of how participants make recognition decisionswill make it possible for us to apply the model to awider range of recognition phenomena, for example sit-uations where recall and familiarity are placed in oppo-sition (e.g., Jacoby, 1991). We could also begin to ad-dress the rich literature on how different manipulationsaffect recognition ROC curves (e.g., Ratcliff et al., 1992;Yonelinas, 1994).

Another topic for future research involves improvingcrosstalk between the model and neuroimaging data. Inprinciple, we should be able to predict fMRI activationsduring episodic recognition tasks by reading out activa-tion from different subregions of the model; to achieve


this goal, we will need to build a “back end” onto themodel that relates changes in (simulated) neuronal activ-ity to changes in the hemodynamic response that is mea-sured by fMRI. Finally, Curran (2000) has isolated whatappear to be distinct ERP correlates of recall and famil-iarity; as such, we should be able to use the model topredict how these recall and familiarity waveforms willbe affected by different manipulations. Our first attemptalong these lines was successful; we found that — as pre-dicted by the model — increasing list strength did not af-fect how well the ERP familiarity correlate discriminatesbetween targets and lures, but list strength adversely af-fected how well the ERP recall correlate discriminatesbetween targets and lures (Norman, Curran, & Tepe, inpreparation).

Alternate Dependent Measures

We also plan to explore other, more biologicallyplausible ways of reading out familiarity from the cor-tical network. While ��

� ��has the virtue of being

easy to compute, it is not immediately clear how someother structure in the brain could isolate the activity ofonly the winning units (because “losing” units are stillactive to some small extent, and there are many morelosing units than winning units). One promising alter-native measure is settle time: the time it takes for activ-ity to spread through the network (more concretely, wemeasured the number of processing cycles needed foraverage activity in MTLC to reach a criterion value =.03). This measure exploits the fact that activity spreadsmore quickly for familiar vs. unfamiliar patterns. The��

measure is more biologically plausible than��

, insofar as it only requires some sensitivity tothe average activity of a layer, and some ability to assesshow much time elapses between stimulus onset and ac-tivity reaching a pre-determined criterion. Preliminarysimulation results show that

�� yields good � �

scores, and – like ��

– does not show a list strengtheffect on � � for our basic parameters (20% overlap). Fur-ther research is necessary to determine if the qualitativeproperties of ��

� ��and

�� are completely

identical or if there are manipulations that affect themdifferently.

Conclusion

We have provided a comprehensive initial treat-ment of the domain of recognition memory using ourbiologically-based neural network model of the hip-pocampus and neocortex. This work extends a similarlycomprehensive application of the same basic model to arange of animal learning phenomena (O’Reilly & Rudy,2001). Thus, we are encouraged by the breadth and depthof data that can be accounted for within our framework.Future work can build upon this foundation to address a

range of other human and animal memory phenomena.

Acknowledgments

This work was supported by ONR grant N00014-00-1-0246, NSF grant IBN-9873492, and NIH programproject MH47566. KAN was supported by NIH NRSAFellowship MH12582. We thank Rafal Bogacz, NeilBurgess, Tim Curran, David Huber, Michael Hasselmo,Caren Rotello, and Craig Stark for their very insightfulcomments on an earlier draft of this manuscript. KANis now at Princeton University, Department of Psychol-ogy, Green Hall, Princeton, NJ 08544, email: [email protected].

Appendix A: Algorithm Details

This appendix describes the computational details ofthe algorithm that was used in the simulations. The al-gorithm is identical to the Leabra algorithm described inO’Reilly and Munakata (2000; O’Reilly, 1998), exceptthe error-driven-learning component of the Leabra algo-rithm was not used here; see Grossberg (1976) for a sim-ilar model. Interested readers should refer to O’Reillyand Munakata (2000) for more details regarding the al-gorithm and its historical precedents.

Pseudocode

The pseudocode for the algorithm that we used isgiven here, showing exactly how the pieces of the algo-rithm described in more detail in the subsequent sectionsfit together.

Outer loop: Iterate over events (trials) within anepoch. For each event, settle over cycles of updating:

1. At start of settling, for all units:

(a) Initialize all state variables (activation, v_m,etc).

(b) Apply external patterns.

2. During each cycle of settling, for all non-clampedunits:

(a) Compute excitatory netinput ( �� or �� ,eq 6).

(b) Compute kWTA inhibition for each layer,based on �

(eq 9):

i. Sort units into two groups based on � :

top�

and remaining� �� to

�.

ii. Set inhib conductance � between � and

� �� (eq 8).


(c) Compute point-neuron activation combiningexcitatory input and inhibition (eq 4).

3. Update the weights (based on linear current weightvalues), for all connections:

(a) Compute Hebbian weight changes (eq 10).

(b) Increment the weights and apply contrast-enhancement (eq 12).

Point Neuron Activation Function

Leabra uses a point neuron activation function thatmodels the electrophysiological properties of real neu-rons, while simplifying their geometry to a single point.This function is nearly as simple computationally asthe standard sigmoidal activation function, but the morebiologically-based implementation makes it consider-ably easier to model inhibitory competition, as describedbelow. Further, using this function enables cognitivemodels to be more easily related to more physiologicallydetailed simulations, thereby facilitating bridge-buildingbetween biology and cognition.

The membrane potential��

is updated as a functionof ionic conductances � with reversal (driving) potentials�

as follows:

� �� (4)

with 3 channels ( � ) corresponding to:�

excitatory input;�leak current; and

�inhibitory input. Following elec-

trophysiological convention, the overall conductance isdecomposed into a time-varying component � � � � com-puted as a function of the dynamic state of the network,and a constant � � that controls the relative influence ofthe different conductances. The equilibrium potentialcan be written in a simplified form by setting the exci-tatory driving potential (

� � ) to 1 and the leak and in-hibitory driving potentials (

��and

� ) of 0:

�� (5)

which shows that the neuron is computing a balance be-tween excitation and the opposing forces of leak andinhibition. This equilibrium form of the equation canbe understood in terms of a Bayesian decision makingframework (O’Reilly & Munakata, 2000).

The excitatory net input/conductance � � � � or �� iscomputed as the proportion of open excitatory channelsas a function of sending activations times the weight val-ues:

� � � � � � � �� (6)

The inhibitory conductance is computed via the kWTAfunction described in the next section, and leak is a con-stant.

Activation communicated to other cells ( � � ) is athresholded ( � ) sigmoidal function of the membrane po-tential with gain parameter � :

� � � � � �� ! #" %$'&)( (7)

where * �,+ � is a threshold function that returns 0 if �(� �and � if -/. � . This sharply-thresholded function isconvolved with a Gaussian noise kernel ( 0 �21 � �43 ),which reflects the intrinsic processing noise of biologicalneurons. This produces a less discontinuous determinis-tic function with a softer threshold that is better suitedfor graded learning mechanisms (e.g., gradient descent).

k-Winners-Take-All Inhibition

Leabra uses a kWTA function to achieve sparsedistributed representations (c.f., Minai & Levy, 1994).Although two different versions are possible (seeO’Reilly & Munakata, 2000 for details), only the sim-pler form was used in the present simulations. A uni-form level of inhibitory current for all units in the layeris computed as follows:

� � � �� 65 � � � �

�� (8)

where � �75 � � is a parameter for setting the inhibitionbetween the upper bound of �

and the lower bound of� �� . These boundary inhibition values are computed as

a function of the level of inhibition necessary to keep aunit right at threshold:

� � �98�:� � � � � �;� � � � :� � � �<� �;� � � � (9)

where �98� is the excitatory net input without the biasweight contribution — this allows the bias weights tooverride the kWTA constraint.

In the basic version of the kWTA function used here,� and �

�� are set to the threshold inhibition value forthe

� � � and� � � � � most excited units, respectively.

Thus, the inhibition is placed exactly to allow�

unitsto be above threshold, and the remainder below thresh-old. For this version, the 5 parameter is almost always.25, allowing the

� � � unit to be sufficiently above theinhibitory threshold. We should emphasize that, whenmembrane potential is at threshold, unit activation in themodel = .25; as such, the kWTA algorithm places a firmupper bound on the number of units showing activation> .25, but it does not set an upper bound on the num-ber of weakly active units (i.e., units showing activationbetween 0 and .25).


Activation dynamics similar to those produced by thekWTA function have been shown to result from simu-lated inhibitory interneurons that project both feedfor-ward and feedback inhibition (O’Reilly & Munakata,2000). Thus, although the kWTA function is somewhatbiologically implausible in its implementation (e.g., re-quiring global information about activation states andusing sorting mechanisms), it provides a computation-ally effective approximation to biologically plausible in-hibitory dynamics.

Hebbian Learning

The simplest form of Hebbian learning adjusts theweights in proportion to the product of the sending ( � )and receiving ( �� ) unit activations: � � � � � � � . Theweight vector is dominated by the principal eigenvectorof the pairwise correlation matrix of the input, but it alsogrows without bound. Leabra uses essentially the samelearning rule used in competitive learning or mixtures-of-Gaussians (Rumelhart & Zipser, 1986; Nowlan, 1990;Grossberg, 1976), which can be seen as a variant of theOja normalization (Oja, 1982):

�� (10)

Rumelhart and Zipser (1986) and O’Reilly and Mu-nakata (2000) showed that, when activations are inter-preted as probabilities, this equation converges on theconditional probability that the sender is active given thatthe receiver is active.

To renormalize Hebbian learning for sparse input ac-tivations, equation 10 can be re-written as follows:

� � � ��* � � � � � � � � � �� #+ (11)

where an�

value of 1 gives equation 10, while a largervalue can ensure that the weight value between uncorre-lated but sparsely active units is around .5. Specifically,we set

� � � � and � � � 1 3 �75 � � 1 3 �� , where �is the sending layer’s expected activation level, and 5 �(called savg cor in the simulator) is the extent to whichthis sending layer’s average activation is fully correctedfor ( 5 � � � gives full correction, and 5 � � � yields nocorrection).

Weight Contrast Enhancement

One limitation of the Hebbian learning algorithm isthat the weights linearly reflect the strength of the condi-tional probability. This linearity can limit the network’sability to focus on only the strongest correlations, whileignoring weaker ones. To remedy this limitation, weintroduce a contrast enhancement function that magni-fies the stronger weights and shrinks the smaller ones

Area Units Activity (pct)EC 240 10.0DG 1600 1.0CA3 480 4.0CA1 640 10.0

Table 2: Sizes of different subregions and their activity levelsin the model.

Projection Mean Var Scale % ConEC to DG, CA3 .5 .25 1 25(perforant path)DG to CA3 .9 .01 25 4(mossy fiber)CA3 recurrent .5 .25 1 100CA3 to CA1 .5 .25 1 100(schaffer)

Table 3: Properties of modifiable projections in the hippocam-pal model: Mean initial weight strength, variance of the initialweight distribution, scaling of this projection relative to otherprojections, and percent connectivity.

in a parametric, continuous fashion. This contrast en-hancement is achieved by passing the linear weight val-ues computed by the learning rule through a sigmoidalnonlinearity of the following form:

� � � ��

�� " �� ( "�� (12)

where � � is the contrast-enhanced weight value, and the

sigmoidal function is parameterized by an offset�

anda gain � (standard defaults of 1.25 and 6, respectively,used here).

Appendix B: Hippocampal Model Details

This section provides a brief summary of key archi-tectural parameters of the hippocampal model. Activitylevels, layer size, and projection parameters were set tomirror the “consensus view” of the functional architec-ture of the hippocampus described, e.g., by Squire, Shi-mamura, and Amaral (1989).

Table 2 shows the sizes of different hippocampal sub-regions and their activity levels in the model. These ac-tivity levels are enforced by setting appropriate

�param-

eters in the Leabra kWTA inhibition function. As dis-cussed in the main text, activity is much more sparse inDG and CA3 than in EC.

Table 3 shows the properties of the four modifiableprojections in the hippocampal model. For each simu-lated participant, connection weights in these projectionsare set to values randomly sampled from a uniform dis-tribution with mean and variance (range) as specified in


the table. The “scale” factor listed in the table shows howinfluential this projection is, relative to other projectionscoming into the layer, and “percent connectivity” speci-fies the percentage of units in the sending layer that areconnected to each unit in the receiving layer. Relativeto the perforant path, the mossy fiber pathway is sparse(i.e., each CA3 neuron receives a much smaller numberof mossy fiber synapses than perforant path synapses)and strong (i.e., a given mossy fiber synapse has a muchlarger impact on CA3 unit activation than a given per-forant path synapse). The CA3 recurrents and the Schaf-fer collaterals projecting from CA3 to CA1 are relativelydiffuse, so that each CA3 neuron and each CA1 neuronreceive a large number of inputs sampled from the entireCA3 population.

The connections linking EC in to CA1, and fromCA1 to EC out, are not modified in the course of the sim-ulated memory experiment. Rather, we pre-train theseconnections so they form an invertible mapping, wherebythe CA1 representation resulting from a given EC inpattern is capable of re-creating that same pattern onEC out. CA1 is arranged into eight columns (consist-ing of 80 units apiece); each column receives input fromthree slots in EC in and projects back to the correspond-ing three slots in EC out. See O’Reilly and Rudy (2001)for discussion of why CA1 is structured in columns.

Lastly, our model incorporates the claim, set forthby Michael Hasselmo and his colleagues, that the hip-pocampus has two functional “modes”: an encodingmode, where CA1 activity is primarily driven by EC in,and a retrieval mode, where CA1 activity is primar-ily driven by stored memory traces in CA3 (e.g., Has-selmo & Wyble, 1997); recently, Hasselmo, Bodelon,and Wyble (2002) presented evidence that these twomodes are linked to different phases of the hippocampaltheta rhythm. While we find the “theta rhythm” hypoth-esis to be compelling, we decided to implement the twomodes in a much simpler way — specifically, we set thescaling factor for the EC in to CA1 projection to a largevalue (6) at study, and we set the scaling factor to zeroat test. This manipulation captures the essential differ-ence between the two modes without adding unnecessarycomplexity to the model.

Appendix C: Basic Parameters

20 items at study: 10 target items (which are tested)followed by 10 interference items (which are not tested)

20% overlap between input patterns (flip 16/24 slots)

Fixed high recall criterion, recall = .40

The following are other basic parameters, most ofwhich are standard default parameter values for the

Leabra algorithm:

Parameter Value Parameter Value� �0.15 � � 0.235� 0.15 � 1.0� � 1.00 � � 1.0�� 0.15 � 0.25� .02 � 600

MTLC � .004 Hippo � .01MTLC savg cor .4 Hippo savg cor 1

For those interested in exploring the model in moredetail, it can be obtained by sending email to the firstauthor.


References

Aggleton, J. P., & Brown, M. W. (1999). Episodic mem-ory, amnesia, and the hippocampal-anterior thalamicaxis. Behavioral and Brain Sciences, 22, 425–490.

Aggleton, J. P., & Shaw, C. (1996). Amnesia and recog-nition memory: A re-analysis of psychometric data.Neuropsychologia, 34, 51–62.

Atkinson, R. C., & Estes, W. K. (1963). Stimulus sam-pling theory. In R. D. Luce, R. R. Bush, & E. Galanter(Eds.), Handbook of mathematical psychology. NewYork: Wiley.

Bauml, K. (1997). The list-strength effect: Strength-dependent competition or suppression? PsychonomicBulletin and Review, 4, 260–264.

Baxter, M. G., & Murray, E. A. (2001a). Effects of hip-pocampal lesions on delayed nonmatching-to-samplein monkeys: A reply to Zola and Squire (2001). Hip-pocampus, 11, 201–203.

Baxter, M. G., & Murray, E. A. (2001b). Opposite rela-tionship of hippocampal and rhinal cortex damage todelayed nonmatching-to-sample deficits in monkeys.Hippocampus, 11, 61–71.

Beason-Held, L. L., Rosene, D. L., Killiany, R. J., &Moss, M. B. (1999). Hippocampal formation lesionsproduce memory impairments in the rhesus monkey.Hippocampus, 9, 562–574.

Bliss, T. V. P., & Collingridge, G. L. (1993). A synap-tic model of memory: long-term potentiation in thehippocampus. Nature, 361, 31–39.

Bogacz, R., & Brown, M. W. (in press). Comparison ofcomputational models of familiarity discrimination inperirhinal cortex. Hippocampus.

Bogacz, R., Brown, M. W., & Giraud-Carrier, C. (2001).Model of familiarity discrimination in the perirhinalcortex. Journal of Computational Neuroscience, 10,5–23.

Brown, M. W., & Xiang, J. Z. (1998). Recognition mem-ory: Neuronal substrates of the judgement of prior oc-currence. Progress in Neurobiology, 55, 149–189.

Burgess, N., & O’Keefe, J. (1996). Neuronal computa-tions underlying the firing of place cells and their rolein navigation. Hippocampus, 6, 749–762.

Carpenter, G. A., & Grossberg, S. (1993). Normal andamnesic learning, recognition and memory by a neu-ral model of cortico-hippocampal interactions. Trendsin Neurosciences, 16, 131–137.

Clark, S. E., & Gronlund, S. D. (1996). Global match-ing models of recognition memory: How the models

match the data. Psychonomic Bulletin and Review, 3,37–60.

Clark, S. E., Hori, A., & Callan, D. E. (1993).Forced-choice associative recognition: Implicationsfor global-memory models. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 19,871–881.

Cohen, N. J., & Eichenbaum, H. (1993). Memory, am-nesia, and the hippocampal system. Cambridge, MA:MIT Press.

Cohen, N. J., Poldrack, R. A., & Eichenbaum, H. (1997).Memory for items and memory for relations in theprocedural/declarative memory framework. In A. R.Mayes, & J. J. Downes (Eds.), Theories of organicamnesia (pp. 131–178). Hove, UK: Psychology Press.

Curran, T. (2000). Brain potentials of recollection andfamiliarity. Memory and Cognition, 28, 923.

Curran, T., & Hintzman, D. L. (1995). Violationsof the independence assumption in process dissocia-tion. Journal of Experimental Psychology: Learning,Memory, and Cognition, 21, 531–547.

Curran, T., Schacter, D. L., Norman, K., & Galluccio,L. (1997). False recognition after a right frontal lobeinfarction: Memory for general and specific informa-tion. Neuropsychologia, 35, 1035.

Dennis, S., & Humphreys, M. S. (2001). A context noisemodel of episodic word recognition. PsychologicalReview, 108, 452–477.

Douglas, R. J., & Martin, K. A. C. (1990). Neocortex.In G. M. Shepherd (Ed.), The synaptic organizationof the brain (Chap. 12, pp. 389–438). Oxford: OxfordUniversity Press.

Eichenbaum, H. (2000). Cortical-hippocampal networksfor declarative memory. Nature Reviews Neuro-science, 1, 41–50.

Eichenbaum, H., H., Otto, T., & Cohen, N. J. (1994).Two functional components of the hippocampal mem-ory system. Behavioral and Brain Sciences, 17(3),449–518.

Gillund, G., & Shiffrin, R. M. (1984). A retrieval modelfor both recognition and recall. Psychological Review,91, 1–67.

Glanzer, M., Adams, J. K., Iverson, G. J., & Kim, K.(1993). The regularities of recognition memory. Psy-chological Review, 100, 546–567.

Grossberg, S. (1976). Adaptive pattern classification anduniversal recoding I: Parallel development and codingof neural feature detectors. Biological Cybernetics,23, 121–134.


Grossberg, S. (1986). The adaptive self-organization ofserial order in behavior: Speech, language, and motorcontrol. In E. C. Scwab, & H. C. Nusbaum (Eds.),Pattern recognition in humans and machines, volumei: Speech perception. New York: Academic Press.

Grossberg, S., & Stone, G. (1986). Neural dynamicsof word recognition and recall: Attentional priming,learning, and resonance. Psychological Review, 93,46–74.

Hasselmo, M. E. (1995). Neuromodulation and corticalfunction: Modeling the physiological basis of behav-ior. Behavioural Brain Research, 67, 1–27.

Hasselmo, M. E., Bodelon, C., & Wyble, B. P. (2002).A proposed function for hippocampal theta rhythm:Separate phases of encoding and retrieval enhance re-versal of prior learning. Neural Computation, 14,793–818.

Hasselmo, M. E., & Wyble, B. (1997). Free recall andrecognition in a network model of the hippocampus:Simulating effects of scopolamine on human memoryfunction. Behavioural Brain Research, 89, 1–34.

Hebb, D. O. (1949). The organization of behavior. NewYork: Wiley.

Hinton, G. E., & Plaut, D. C. (1987). Using fast weightsto deblur old memories. Proceedings of the 9th An-nual Conference of the Cognitive Science Society (pp.177–186). Hillsdale, NJ: Erlbaum.

Hintzman, D. L. (1988). Judgments of frequencyand recognition memory in a multiple-trace memorymodel. Psychological Review, 95, 528–551.

Hintzman, D. L. (2001). Similarity, global matching, andjudgments of frequency. Memory and Cognition, 29,547–556.

Hintzman, D. L., Curran, T., & Oppy, B. (1992). Effectsof similiarity and repetition on memory: Registrationwithout learning. Journal of Experimental Psychol-ogy: Learning, Memory, and Cognition, 18, 667–680.

Holdstock, J. S., Mayes, A. R., Roberts, N., Cezayirli,E., Isaac, C. L., O’Reilly, R. C., & Norman, K. A.(2002). Under what conditions is recognition sparedrelative to recall after selective hippocampal damagein humans? Hippocampus, 12, 341–351.

Humphreys, M. S., Bain, J. D., & Pike, R. (1989). Differ-ent ways to cue a coherent memory system: A theoryfor episodic, semantic, and procedural tasks. Psycho-logical Review, 96, 208–233.

Jacoby, L. L. (1991). A process dissociation framework:Separating automatic from intentional uses of mem-ory. Journal of Memory and Language, 30, 513–541.

Jacoby, L. L., Yonelinas, A. P., & Jennings, J. M. (1997).The relation between conscious and unconscious (au-tomatic) influences: A declaration of independence.In J. D. Cohen, & J. W. Schooler (Eds.), Scientific ap-proaches to consciousness (pp. 13–47). Mahway, NJ:Lawrence Erlbaum Associates.

Joordens, S., & Hockley, W. E. (2000). Recollection andfamiliarity through the looking glass: When old doesnot mirror new. Journal of Experimental Psychology:Learning, Memory, and Cognition, 26, 1534.

Kahana, M. J., & Rizzuto, D. (submitted). An analy-sis of the recognition-recall relation in four distributedmemory models.

Kanerva, P. (1988). Sparse distributed memory. Boston:Bradford MIT.

Kohonen, T. (1977). Associative memory: A system the-oretical approach. Berlin: Springer-Verlag.

Kroll, N. E. A., Knight, R. T., Metcalfe, J., Wolf, E. S.,& Tulving, E. (1996). Cohesion failure as a source ofmemory illusions. Journal of Memory and Language,35, 176–196.

Levy, W. B. (1989). A computational approach to hip-pocampal function. In R. D. Hawkins, & G. H. Bower(Eds.), Computational models of learning in simpleneural systems (pp. 243–304). San Diego, CA: Aca-demic Press.

Li, L., Miller, E. K., & Desimone, R. (1993). Therepresentation of stimulus familiarity in anterior in-ferior temporal cortex. Journal of Neurophysiology,69, 1918–1929.

Macmillan, N. A., & Creelman, C. D. (1991). Detec-tion theory: A user’s guide. New York: CambridgeUniversity Press.

Malenka, R. C., & Nicoll, R. A. (1993). NMDA receptor-dependent synaptic plasticity: Multiple forms andmechanisms. Trends in Neurosciences, 16, 521–527.

Mandler, G. (1980). Recognizing: The judgment of pre-vious occurrence. Psychological Review, 87, 252–271.

Manns, J. R., & Squire, L. R. (1999). Impaired recogni-tion memory on the Doors and People test after dam-age limited to the hippocampal region. Hippocampus,9, 495–499.

Marr, D. (1971). Simple memory: A theory for archicor-tex. Philosophical Transactions of the Royal Society(London) B, 262, 23–81.

Mayes, A. R., Holdstock, J. S., Isaac, C. L., Hunkin,N. M., & Roberts, N. (2002). Relative sparing ofitem recognition memory in a patient with adult-onset


damage limited to the hippocampus. Hippocampus,12, 325–340.

Mayes, A. R., Isaac, C. L., Downes, J. J., Holdstock,J. S., Hunkin, N. M., Montaldi, D., MacDonald, C.,Cezayirli, E., & Roberts, J. N. (2001). Memory forsingle items, word pairs, and temporal order in a pa-tient with selective hippocampal lesions. CognitiveNeuropsychology, 18, 97–123.

McClelland, J. L., & Chappell, M. (1998). Familiar-ity breeds differentiation: a subjective-likelihood ap-proach to the effects of experience in recognitionmemory. Psychological Review, 105, 724.

McClelland, J. L., & Goddard, N. H. (1996). Consid-erations arising from a complementary learning sys-tems perspective on hippocampus and neocortex. Hip-pocampus, 6, 654–665.

McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C.(1995). Why there are complementary learning sys-tems in the hippocampus and neocortex: Insights fromthe successes and failures of connectionist models oflearning and memory. Psychological Review, 102,419–457.

McNaughton, B. L., & Morris, R. G. M. (1987).Hippocampal synaptic enhancement and informationstorage within a distributed memory system. Trendsin Neurosciences, 10(10), 408–415.

Miller, E. K., Li, L., & Desimone, R. (1991). A neuralmechanism for working and recognition memory ininferior temporal cortex. Science, 254, 1377–9.

Minai, A. A., & Levy, W. B. (1994). Setting the activitylevel in sparse random networks. Neural Computa-tion, 6, 85–99.

Mishkin, M., Suzuki, W., Gadian, D. G., & Vargha-Khadem, F. (1997). Hierarchical organization of cog-nitive memory. Philosophical Transactions of theRoyal Society, London, B, 352, 1461–1467.

Mishkin, M., Vargha-Khadem, F., & Gadian, D. G.(1998). Amnesia and the organization of the hip-pocampal system. Hippocampus, 8, 212–216.

Mitchell, K. J., & Johnson, M. K. (2000). Source mon-itoring: Attributing mental experiences. In E. Tulv-ing, & F. I. M. Craik (Eds.), The oxford handbook ofmemory (pp. 179–195). New York: Oxford UniversityPress.

Moll, M., & Miikkulainen, R. (1997). Convergence-zoneepisodic memory: Analysis and simulations. NeuralNetworks, 10, 1017–1036.

Mumby, D. G., Wood, E. R., Duva, C. A., Kornecook,T. J., Pinel, J. P. J., & Phillips, A. G. (1996). Ischemia-induced object recognition deficits in rats are attenu-

ated by hippocampal lesions before or soon after is-chemia. Behavioral Neuroscience, 110, 266–281.

Murnane, K., & Shiffrin, R. (1991a). Interference andthe representation of events in memory. Journal of Ex-perimental Psychology: Learning Memory and Cog-nition, 17, 855–874.

Murnane, K., & Shiffrin, R. M. (1991b). Word repeti-tions in sentence recognition. Memory and Cognition,19, 119–130.

Murray, E. A., & Mishkin, M. (1998). Object recognitionand location memory in monkeys with excitotoxic le-sions of the amygdala and hippocampus. Journal ofNeuroscience, 18, 6568.

Norman, K., Curran, T., & Tepe, K. (in preparation). In-terference effects on ERP correlates of recall and fa-miliarity.

Norman, K. A. (in press). Differential effects of liststrength on recollection and familiarity. Journal of Ex-perimental Psychology: Learning, Memory, and Cog-nition.

Nowlan, S. J. (1990). Maximum likelihood competitivelearning. In D. S. Touretzky (Ed.), Advances in neuralinformation processing systems, 2 (pp. 574–582). SanMateo, CA: Morgan Kaufmann.

Oja, E. (1982). A simplified neuron model as a principalcomponent analyzer. Journal of Mathematical Biol-ogy, 15, 267–273.

O’Keefe, J., & Nadel, L. (1978). The hippocampus as acognitive map. Oxford, England: Oxford UniversityPress.

O’Reilly, R. C. (1996). The Leabra model of neural in-teractions and learning in the neocortex. PhD thesis,Carnegie Mellon University, Pittsburgh, PA.

O’Reilly, R. C. (1998). Six principles for biologically-based computational models of cortical cognition.Trends in Cognitive Sciences, 2(11), 455–462.

O’Reilly, R. C., & Busby, R. S. (2002). Generalizablerelational binding from coarse-coded distributed rep-resentations. In T. G. Dietterich, S. Becker, & Z.Ghahramani (Eds.), Advances in Neural InformationProcessing Systems (NIPS) 14. Cambridge, MA: MITPress.

O’Reilly, R. C., & McClelland, J. L. (1994). Hippocam-pal conjunctive encoding, storage, and recall: Avoid-ing a tradeoff. Hippocampus, 4(6), 661–682.

O’Reilly, R. C., & Munakata, Y. (2000). Computationalexplorations in cognitive neuroscience: Understand-ing the mind by simulating the brain. Cambridge,MA: MIT Press.


O’Reilly, R. C., Norman, K. A., & McClelland, J. L.(1998). A hippocampal model of recognition memory.In M. I. Jordan, M. J. Kearns, & S. A. Solla (Eds.), Ad-vances in neural information processing systems 10(pp. 73–79). Cambridge, MA: MIT Press.

O’Reilly, R. C., & Rudy, J. W. (2001). Conjunctive rep-resentations in learning and memory: Principles ofcortical and hippocampal function. Psychological Re-view, 108, 311–345.

Raaijmakers, J. G., & Shiffrin, R. M. (1992). Models forrecall and recognition. Annual Review of Psychology,43, 205–234.

Ratcliff, R. (1990). Connectionist models of recognitionmemory: Constraints imposed by learning and forget-ting functions. Psychological Review, 97, 285–308.

Ratcliff, R., Clark, S., & Shiffrin, R. (1990). The liststrength effect: I. Data and discussion. Journal of Ex-perimental Psychology: Learning Memory and Cog-nition, 16, 163–178.

Ratcliff, R., & McKoon, G. (2000). Memory models. InE. Tulving, & F. I. M. Craik (Eds.), The Oxford hand-book of memory (pp. 571–581). New York: OxfordUniversity Press.

Ratcliff, R., Sheu, C.-F., & Gronlund, S. D. (1992). Test-ing global memory models using ROC curves. Psy-chological Review, 99, 518–535.

Ratcliff, R., Van Zandt, T., & McKoon, G. (1995). Pro-cess dissociation, single-process theories, and recog-nition memory. Journal of Experimental Psychology:General, 124, 352–374.

Reder, L. M., Nhouyvanisvong, A., Schunn, C. D., Ay-ers, M. S., Angstadt, P., & Hiraki, K. A. (2000). Amechanistic account of the mirror effect for word fre-quency: A computational model of remember-knowjudgments in a continuous recognition paradigm.Journal of Experimental Psychology: Learning,Memory, and Cognition, 26, 294–320.

Reed, J. M., Hamann, S. B., Stefanacci, L., & Squire,L. R. (1997). When amnesic patients perform well onrecognition memory tests. Behavioral Neuroscience,111, 1163–1170.

Reed, J. M., & Squire, L. R. (1997). Impaired recog-nition memory in patients with lesions limited to thehippocampal formation. Behavioral Neuroscience,111, 667–675.

Rempel-Clower, N. L., Zola, S. M., & Amaral, D. G.(1996). Three cases of enduring memory impairmentafter bilateral damage limited to the hippocampal for-mation. Journal of Neuroscience, 16, 5233.

Riches, I. P., Wilson, F. A., & Brown, M. W. (1991). Theeffects of visual stimulation and memory on neurons

of the hippocampal formation and the neighbouringparahippocampal gyrus and inferior temporal cortexof the primate. Journal of Neuroscience, 11, 1763–79.

Rolls, E. T. (1989). Functions of neuronal networks inthe hippocampus and neocortex in memory. In J. H.Byrne, & W. O. Berry (Eds.), Neural models of plas-ticity: Experimental and theoretical approaches (pp.240–265). San Diego, CA: Academic Press.

Rolls, E. T., Baylis, G. C., Hasselmo, M. E., & Nalwa,V. (1989). The effect of learning on the face selec-tive responses of neurons in the cortex in the superiortemporal sulcus of the monkey. Experimental BrainResearch, 76, 153–164.

Rotello, C. M., & Heit, E. (1999). Two-process modelsof recognition memory: Evidence for recall-to-reject.Journal of Memory and Language, 40, 432.

Rotello, C. M., & Heit, E. (2000). Associative recog-nition: a case of recall-to-reject processing. Memoryand Cognition, 28, 907.

Rotello, C. M., Macmillan, N. A., & Van Tassel, G.(2000). Recall-to-reject in recognition: Evidencefrom ROC curves. Journal of Memory and Language,43, 67–88.

Rudy, J. W., & O’Reilly, R. C. (2001). Conjunctive repre-sentations, the hippocampus, and contextual fear con-ditioning. Cognitive, Affective, and Behavioral Neu-roscience, 1, 66–82.

Rudy, J. W., & Sutherland, R. J. (1989). The hippocam-pal formation is necessary for rats to learn and remem-ber configural discriminations. Behavioural Brain Re-search, 34, 97–109.

Rudy, J. W., & Sutherland, R. W. (1995). Configural as-sociation theory and the hippocampal formation: Anappraisal and reconfiguration. Hippocampus, 5, 375–389.

Rumelhart, D. E., & Zipser, D. (1986). Feature discoveryby competitive learning. In D. E. Rumelhart, J. L. Mc-Clelland, & PDP Research Group (Eds.), Parallel dis-tributed processing. Volume 1: Foundations (Chap. 5,pp. 151–193). Cambridge, MA: MIT Press.

Sherry, D. F., & Schacter, D. L. (1987). The evolutionof multiple memory systems. Psychological Review,94(4), 439–454.

Shiffrin, R., Ratcliff, R., & Clark, S. (1990). The liststrenth effect: II. Theoretical mechanisms. Journalof Experimental Psychology: Learning Memory andCognition, 16, 179–195.

Shiffrin, R. M., Huber, D. E., & Marinelli, K. (1995).Effects of category length and strength on familiarity


in recognition. Journal of Experimental Psychology:Learning, Memory and Cognition, 21, 267–287.

Shiffrin, R. M., & Steyvers, M. (1997). a model forrecognition memory: REM – retrieving effectivelyfrom memory. Psychonomic Bulletin and Review, 4,145–166.

Snodgrass, J. G., & Corwin, J. (1988). Pragmatics ofmeasuring recognition memory: Applications to de-mentia and amnesia. Journal of Experimental Psy-chology: General, 117(1), 34–50.

Sohal, V. S., & Hasselmo, M. E. (2000). A model forexperience-dependent changes in the responses of in-ferotemporal neurons. Network : Computation inNeural Systems, 11, 169.

Squire, L. R. (1987). Memory and brain. Oxford, Eng-land: Oxford University Press.

Squire, L. R. (1992a). Declarative and nondeclarativememory: Multiple brain systems supporting learn-ing and memory. Journal of Cognitive Neuroscience,4(3), 232–243.

Squire, L. R. (1992b). Memory and the hippocampus: Asynthesis from findings with rats, monkeys, and hu-mans. Psychological Review, 99, 195–231.

Squire, L. R., Shimamura, A. P., & Amaral, D. G. (1989).Memory and the hippocampus. In J. H. Byrne, &W. O. Berry (Eds.), Neural models of plasticity: Ex-perimental and theoretical approaches (pp. 208–239).San Diego, CA: Academic Press.

Standing, L. (1973). Learning 10,000 pictures. QuarterlyJournal of Experimental Psychology, 25, 207–222.

Stark, C. E. L., & Squire, L. R. (2002, April). Memoryfor associations in patients with hippocampal damage.9th Annual Meeting of the Cognitive Neuroscience So-ciety.

Sutherland, R. J., & Rudy, J. W. (1989). Configural asso-ciation theory: The role of the hippocampal formationin learning, memory, and amnesia. Psychobiology,17(2), 129–144.

Teyler, T. J., & Discenna, P. (1986). The hippocampalmemory indexing theory. Behavioral Neuroscience,100, 147–154.

Touretzky, D. S., & Redish, A. D. (1996). A theory ofrodent navigation based on interacting representationsof space. Hippocampus, 6, 247–270.

Treves, A., & Rolls, E. T. (1994). A computational anal-ysis of the role of the hippocampus in memory. Hip-pocampus, 4, 374–392.

Vargha-Khadem, F., Gadian, D. G., Watkins, K. E., Con-nelly, A., Van Paesschen, W., & Mishkin, M. (1997).Differential effects of early hippocampal pathology on

episodic and semantic memory. Science, 277, 376–380.

Wu, X., Baxter, R. A., & Levy, W. B. (1996). Contextcodes and the effect of noisy learning on a simplifiedhippocampal CA3 model. Biological Cybernetics, 74,159–165.

Xiang, J. Z., & Brown, M. W. (1998). Differential en-coding of novelty, familiarity, and recency in regionsof the anterior temporal lobe. Neuropharmacology,37, 657–676.

Yonelinas, A. P. (1994). Receiver-operating character-istics 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 foritem and associative information: The contribution ofrecollection and familiarity. Memory and Cognition,25, 747–763.

Yonelinas, A. P. (2001a). Consciousness, control, andconfidence: The 3 cs of recognition memory. Journalof Experimental Psychology General, 130, 361–379.

Yonelinas, A. P. (2001b). Consciousness, control, andconfidence: The three Cs of recognition memory.Journal of Experimental Psychology: General, 130,361–379.

Yonelinas, A. P. (2002). The nature of recollection andfamiliarity: A review of 30 years of research. Journalof Memory and Language, 46, 441–517.

Yonelinas, A. P., Dobbins, I., Szymanski, M. D., &Dhaliwal, H. S.and King, L. (1996). Signal-detection,threshold, and dual-process models of recognitionmemory: ROCs and conscious recollection. Con-sciousness and Cognition, 5, 418–441.

Yonelinas, A. P., Hockley, W. E., & Murdock, B. B.(1992). Tests of the list-strength effect in recogni-tion memory. Journal of Experimental Psychology:Learning, Memory, and Cognition, 18, 345–355.

Zola, S. M., & Squire, L. R. (2001). Relationshipbetween magnitude of damage to the hippocampusand impaired recognition memory in monkeys. Hip-pocampus, 11, 92–98.

Zola, S. M., Squire, L. R., Teng, E., Sefanacci, L., Buf-falo, E., & Clark, R. E. (2000). Impaired recognitionmemory in monkeys after damage limited to the hip-pocampal region. Journal of Neuroscience, 20, 451–463.

Zola-Morgan, S., Squire, L. R., & Amaral, D. G. (1986).Human amnesia and the medial temporal region: En-during memory impairment following a bilateral le-sion limited to field ca1 of the hippocampus. Journalof Neuroscience, 6, 2950–2967.

Documents

Modeling Hippocampal and Neocortical Contributions to