Associative structures in animal learning: Dissociating elemental and configural processes

Neurobiology of Learning and Memory 108 (2014) 96–103

Contents lists available at SciVerse ScienceDirect

Neurobiology of Learning and Memory

journal homepage: www.elsevier .com/ locate /ynlme

Review

Associative structures in animal learning: Dissociating elementaland configural processes

1074-7427/$ - see front matter � 2013 Elsevier Inc. All rights reserved.http://dx.doi.org/10.1016/j.nlm.2013.06.002

⇑ Corresponding author. Address: School of Psychology, Cardiff University, TowerBuilding, Park Place, Cardiff CF10 3AT, United Kingdom.

E-mail address: [email protected] (R.C. Honey).

Robert C. Honey a,⇑, Mihaela D. Iordanova b, Mark Good a

a Cardiff University, United Kingdomb University of Maryland, Baltimore, USA

a r t i c l e i n f o a b s t r a c t

Article history:Available online 13 June 2013

Keywords:ElementalConfiguralHippocampus

The central concern of associative learning theory is to provide an account of behavioral adaptation that isparsimonious in addressing three key questions: (1) under what conditions does learning occur, (2) whatare the associative structures involved, and (3) how do these affect behavior? The principle focus here ison the second question, concerning associative structures, but we will have cause to touch on the othersin passing. This question is one that has exercised theorists since Pavlov’s descriptions of the conditioningprocess, where he identifies the shared significance of the study of conditioned reflexes for psychologistsand neuroscientists alike.

� 2013 Elsevier Inc. All rights reserved.

1. Historical orientation

‘‘Hence, the temporary nervous connexion is a universal physiolog-ical phenomenon both in the animal world and our own. And at thesame time it is a psychic phenomenon, which psychologists call anassociation, no matter whether it is a combination of variousactions or impressions, or that of letters, words, and thoughts.What reason might there be for drawing any distinction betweenwhat is known to a physiologist as a temporary connexion and toa psychologist as an association? Here we have a perfect coales-cence, a complete absorption of one by the other, a complete iden-tification. Psychologists seem to have likewise acknowledged this,for they (or at any rate some of them) have made statements thatexperiments with conditioned reflexes have provided associativepsychology. . .. . .with a firm basis.’’

Pavlov (1935; The Conditioned Reflex; taken from: Lectures onConditioned Reflexes (Volume 2): Conditioned Reflexes and Psy-chiatry, p. 167, 1941).

For Pavlov then, conditioned reflexes were a means of measur-ing, noninvasively but remotely, the formation of associative linksin the brain; a sentiment echoed by Konorski (1967) who statedthat ‘‘the conditioned response [is] playing the role of a ‘tracer’ allow-ing the association to be detected.’’ However, these authors alsoraised the prospect of a rapprochement between an associativeanalysis of learning, on the one hand, and its neural instantiation,

on the other hand. The field of animal learning theory, whilelargely eschewing this prospect, has aimed to answer three funda-mental questions about the nature of the conceptual nervous sys-tem through the lens provided by conditioned behavior: underwhat conditions does learning occur, what is the nature of theassociative structures that underlie learnt changes in behavior,and how is learning translated into performance? However, aswe hope to show, the analytic tools developed in the service of thisaim, together with the resulting insights that their use provides,can shape our understanding of the brain mechanisms that under-pin learning and memory; and the reverse is also true: neuroscien-tific analysis can permit a resolutions to theoretical issues thathave proven, if not intractable, then certainly elusive.

2. Associative structures: conditioning and sensorypreconditioning

For an extended period, during the first part of the 20th century,learning theory provided a theoretical framework that was bothparsimonious and dominant. The essential idea was that learningin animals could be explained by the formation of associative linksbetween the processes that are concurrently activated by a stimu-lus (S) and a response motor program (R) when both are followedby a reinforcer (e.g., Hull, 1943; Thorndike, 1911). This idea is cer-tainly parsimonious: even in bare outline, it provides simple an-swers to all three of the questions posed above: The conditionsrequired for learning are that the processes activated by the criticalevents (the S and R) occur close together in time and are ‘‘stampedin’’ by contingent reinforcement; the content of learning takes theform of an S ? R link; and learnt behavior is manifest in perfor-mance to the extent that the presentation of the stimulus is able

R.C. Honey et al. / Neurobiology of Learning and Memory 108 (2014) 96–103 97

to activate the response motor program via the link that hasformed between their corresponding processes. Such prosaic theo-rizing is admirable, but in each case the answers provided by S ? Rtheory turned out to be incomplete.

Take the example of sensory preconditioning (SPC), which willhave continuing relevance throughout this article: After two neu-tral stimuli have been paired (e.g., a light with a tone; e.g., Brogden,1939) directly establishing a conditioned response to the tone (bypairing it with shock) also results in the light eliciting responding.In this case, light-tone pairings result in learning without obviousreinforcement and which is behaviorally silent (for further discus-sion, see Honey, Good, & Manser, 1998a; Honey, Watt, & Good,1998b). Demonstrations of SPC undermine a simple S ? R analysisof the nature of associations that underpin animal behavior (p. 85–87, Mackintosh, 1974), and suggest a need to consider alternativeassociative structures.

Two candidate associative structures have often been advancedin the context of discussions of SPC, and standard forms of condi-tioning alike: elemental and configural. As we shall see, both ofthese types of structure can underpin the process of pattern com-pletion: re-creation of a training episode from the presentation ofone of its components; and pattern separation: enabling trainingepisodes with overlapping components to be represented as sepa-rate memories. The elemental analysis holds that the central pro-cesses or memories activated by events (e.g., the light and tone)become directly linked to one another by an elemental association(see left side of Fig. 1). In contrast, the configural analysis holdsthat these processes become linked to some third, independentconfigural representation that then codes for their co-occurrence(see right side of Fig. 1). The elemental and configural accountsprovide a ready account for simple demonstrations of SPC. Forexample, the two elementary associations resulting from thelight ? tone and tone ? shock pairings can form an associativechain that allows the light to provoke a memory of shock andthereby a conditioned response at test (but see also, Lin & Honey,2011; Lin, Dumigan, Dywer, Good & Honey, 2013; Ward-Robinson& Hall, 1996). According to a configural analysis, the light and tonebecome linked to a separate configural unit, which is later (i) linkedto shock during tone ? shock pairings, and (ii) mediates respond-ing to the light at test. This configural analysis might appear to becontrived, especially when applied to what is learnt during thesimple pairing of two stimuli. However, there is evidence fromboth studies of conditioning and parallel studies of SPC showingthat a configural analysis should not be dismissed (e.g., Iordanova,Good, & Honey, 2008). In fact, there is now compelling neuroscien-tific evidence, that we shall come to later, suggesting that both ele-mental and configural associative structures are acquired duringexposure to patterns of stimulation. But, for now, we need to con-sider briefly evidence showing that animals can acquire configuralassociations, and how this evidence has been addressed by theoriesof associative learning.

a b

Fig. 1. (a) Elemental and (b) configural associative structures that could provide thebasis for demonstrations that rats learn that two stimuli co-occur (e.g., during SPCprocedures).

3. Standard configural discriminations

One impetus for the idea that animal behavior can be based onthe formation of configural associations is straightforward: Theycan solve discriminations that should prove impossible if theywere reliant on purely elementary associations. For example, in aconfigural discrimination, rats might be placed in two visual con-texts (A and B; e.g., chambers with spotted or checked wallpapers)and receive separate presentations of two auditory stimuli (X andY; e.g., a tone and clicker). In context A, presentations of X arepaired with food while those of Y are not, and in B presentationsof Y are paired with food and those of X are not. The fact that bothof the contexts, like both of the auditory stimuli, are equally oftenpaired with food (and no food), means that animals only capable offorming elemental associations might come to show conditionedresponding (approaching the site of food delivery) when placedin either context and presented with both auditory stimuli; butthey should not show more conditioned responding during thereinforced configurations (AX and BY) than during the nonrein-forced configurations (AY and BX). The fact that they do means thata purely elemental analysis is unsustainable. However, more com-plex elemental analyses have been developed that are capable ofexplaining how configural discriminations are learnt. Accordingto these analyses, the memorial elements that are activated by acompound stimulus (AX) are not a simple product of those thatare activated by separate presentations of A and X. For example,it has been proposed that each of the four context-auditory stimu-lus combinations (i.e., AX, BX, AY, and BY) gives rise unique ele-ments (i.e., ax, ay, bx, and by; Wagner & Rescorla, 1972); or thateach stimulus (e.g., X) might activate a set of elements (cf. Atkinson& Estes, 1963) the composition of which is affected by whether it ispresented in one context (A) or another context (e.g., B; Wagner,2003). Even without further elaboration, but noting the combina-torial explosion with increases in the number of stimuli in a com-pound, it is clear that these modifications allow an elementalanalysis to be developed for the acquisition of configural discrim-inations: The elements that are uniquely activated by a given com-bination of stimuli become linked to the outcome of the trial, andthereby provide a basis for conditioned responding to be more evi-dent during the reinforced compounds (AX and BY) than the non-reinforced compounds (AY and BX).

There is evidence (for a review, see Honey, Close & Lin, 2011)that is already troublesome for both modified elemental theory(e.g., Wagner, 2003) and purely configural theories (e.g., Pearce,1994). Leaving aside this evidence, however, there is one straight-forward prediction that unites both analyses: It should not be pos-sible to observe a clear-cut dissociation between discriminationlearning problems according to whether they are (operationally)elemental or configural. This prediction follows from the assump-tions that all discriminations involve a single system that instanti-ates the same type of associative structure: either elemental orconfigural. It is just such a dissociation that has been recently ob-served using variants of a SPC procedure.

4. Dissociating elemental and configural structures in sensorypreconditioning

We have recently developed a novel set of behavioral assaysthat can be defined operationally as elemental or configural. Bothtypes of assay involved rats encountering different auditory stimuli(X and Y) in different contexts (A or B) and at different times of day(morning and afternoon). The choice of these stimuli was moti-vated, at least in part, by claims that animals can form memoriesthat integrate the components of episodic memory: whathappened (X or Y), where (context A or B) and when (morning or

98 R.C. Honey et al. / Neurobiology of Learning and Memory 108 (2014) 96–103

afternoon; see, for example, Clayton & Dickinson, 1998; Eacott,Easton, & Zinkivskay, 2005; Palmer & Good, 2011). For one set ofrats that was given elemental training, X was presented wheneverthey were placed in A, and Y was presented whenever they werein B (see Fig. 2a); and for the other set, X was presented in themorning and Y in the afternoon wherever they were placed (A orB; see Fig. 2b). These latter discriminations can be defined as ele-mental: The pairs of stimuli from one dimension (A and B; ormorning and afternoon) are consistently paired with X and Yrespectively, while those from the accompanying dimension(morning and afternoon; or A and B) are not. A conditioned re-sponse (fear, as measured by inactivity or freezing) was then estab-lished to one of the auditory stimuli (X) but not the other (Y), when

Fig. 2. Experimental designs: Contrasting elemental training involving (a) contexts or (sessions, one in a spotted and one in a checked context (A and B), in both the morning (AMY) were presented (see text for details of elemental and configural training). The revaluatand receiving pairings of one auditory stimulus (X) with shock and nonreinforced presenin the two contexts in the morning and afternoon.

they were placed in a novel context (C) at midday (for the revalu-ation stage). The question of primary interest was the level offreezing elicited by contexts A and B in the morning and afternoon.Rats given elemental training, where placement in context A wasconsistently paired with X, should be more fearful in A than B, irre-spective of the time of day. Similarly, rats given training where themorning was consistently paired with X should be more fearful inthe morning than the afternoon, irrespective of the context (A or B)in which they were placed. These predictions were assessed bymeasuring the amount of freezing in the two contexts at bothtimes of day, and dividing the amount of freezing in the context(or time of day) where X was presented, by the total amount offreezing in both contexts (or at both times of day). Using this

b) time with (c) configural training. During preexposure, rats received two training) and afternoon (PM); during these sessions auditory stimuli (tone and clicks; X and

ion sessions consisted of the rats being placed in an undecorated chamber at middaytations of the other auditory stimulus (Y). In the final test sessions, rats were placed

R.C. Honey et al. / Neurobiology of Learning and Memory 108 (2014) 96–103 99

measure, scores above .50 indicate that rats are showing the appro-priate pattern of freezing. As we shall soon see, this was the patternof results observed in control rats.

For rats given configural training, the situation was different: Inthe morning, X was presented in A and Y was presented in B,whereas in the afternoon Y was presented in A and X was pre-sented in B (see Fig. 2c). This training can be defined as configural:The stimuli from each of the dimensions (e.g., X or Y) are equallyoften paired with those stimuli from both of the other dimensions(e.g., A or B, morning or afternoon). According to a simple elemen-tal analysis, this will mean that X, for example, will be capable ofprovoking activity in the memories of each of the elements fromthe other dimensions (i.e., A, B, morning and afternoon). Underthese conditions, this analysis predicts that pairing X with shockshould result in conditioned fear in (i) contexts A and B, and (ii)the morning and afternoon; and the rats should show equal fearin test compounds AM + A, AM + B, PM + A and PM + B. In contrast,if the rats given configural training represented each of the fourpatterns as configurations (AM + A + X, AM + B + Y, PM + A + Y andPM + B + X), then establishing fear to X should result in the config-urations involving X becoming linked to shock, and thus AM + Aand PM + B should elicit greater fear than AM + B and PM + A. Inshort, the rats should be more fearful in context A than in contextB in the morning, and more fearful in context B than A in theafternoon.

We assessed the accuracy of these predictions by measuring theamount of freezing in the two contexts at both times of day. To re-duce individual variability we computed separate freezing ratioscores for the AM tests and the PM tests. The ratios for both teststook the same form: The amount of freezing in context A dividedby the amount of freezing in both contexts A and B. With this mea-sure, if the rats are showing more freezing in context A than in con-text B in the AM tests, then their scores should be above .50;whereas if they are more fearful in context B than in context A inthe PM tests, then their scores should be below .50. As we will soonsee, control rats given configural training responded in the wayanticipated: with the ratio scores being above .50 in the morningand below .50 in the afternoon (cf. Iordanova et al., 2008).

As we have already indicated, the fact that our behavioral as-says can be defined as elemental or configural does not require thatthe observed pattern of test performance is underpinned by differ-ent types of associative structure. The results observed in our con-trol rats (and those with sham lesions) after both elemental andconfigural training could be explained in terms of either modifiedelemental associative structures or configural structures; providedthat these structures can be updated (in some way) as a result ofthe revaluation stage when X and Y are presented alone. Two gen-eral observations are consistent with such attempts to build a par-simonious analysis. The similarity of the four, three-elementtraining patterns to one another (elemental: AM + A + X,AM + B + Y, PM + A + X, PM + B + Y; or AM + A + X, AM + B + X,PM + A + Y, PM + B + Y; configural: AM + A + X, PM + A + Y,AM + B + Y, PM + B + X) can be shown, at least on average, to beequivalent in the two training conditions (see Iordanova, Burnett,Good, & Honey, 2011). And, as will become clear, both forms oftraining resulted in similar levels of performance: in terms of thedivergence of the scores from .50 in rats given elemental and con-figural training. However, such a parsimonious analysis, whetherbased upon elemental or configural principles, is not sustainablegiven the central findings from our laboratory. Namely, a clearand reliable dissociation between test performance that is basedupon elemental and configural training resulting from a varietyof manipulations of the hippocampus (Iordanova, Burnett,Aggleton, Good, & Honey, 2009; Iordanova, Good, & Honey, 2011;Iordanova et al., 2011; see also Good, Barnes, Staal, McGregor, &Honey, 2007).

Iordanova et al. (2011; see also, Iordanova et al., 2009) demon-strated that rats with lesions of the hippocampus (made beforetraining) exhibited a clear dissociation between the test perfor-mance based on elemental and configural training. Rats withsham lesions (like controls) and those with lesions of the hippo-campus (group HPC) were equally likely to have scores that weresignificantly above .50, and did so after both types of elementaltraining (context relevant and time-of-day relevant; see Fig. 3a).There was some indication of superior performance in rats givenelemental training involving times of day than those given ele-mental training involving contexts; however, under somewhatdifferent training conditions the opposite tendency was observed(see Fig. 3c). Importantly, the effects in controls were always mir-rored in rats that had received manipulations involving the hippo-campus (see Fig. 3a–c). Turning now to rats that were givenconfigural training, those with sham lesions were more fearfulin context A than in B in the morning (with scores above .50),and more fearful in context B than A in the afternoon (with scoresbelow .50; see Fig. 3d; cf. Iordanova et al., 2008). However, rats ingroup HPC were equally likely to show fear in contexts A and B inthe morning and afternoon (i.e. their scores diverged little from.50; see also Iordanova et al., 2009; see Fig. 3d). Two things areworth noting about this deficit: It did not reflect the fact thatthe rats with lesions were unable to freeze: their overall levelsof freezing across the four types of test trials were similar to thosein sham-operated rats; and it did not reflect differences in the ex-tent to which X and Y supported differential conditioned fear –which was found to be equivalent in a final (extinction) test inwhich X and Y were presented in context C at midday. This disso-ciation between procedures that are operationally elemental andconfigural, therefore, provides compelling support for the generalidea that both elemental and configural associative structures areacquired during SPC procedures that make use of the same stimuliand training procedures.

It could be deemed premature to attach too much significanceto the dissociation described in the previous paragraphs, were itnot for the fact that it is highly reliable and has enabled the roleof the hippocampus to be explored in some detail. Thus, essentiallythe same elemental-configural dissociation can be observed whensynaptic transmission in the hippocampus is temporarily disrupted(by infusions of muscimol, MUS) either during the revaluationstage (compare the combined elemental results in Fig. 3b, withthe configural results in 3f; Iordanova et al., 2011b) or during thetest (compare Fig. 3c with h; Iordanova et al., 2009). The same dis-sociation is observed when NMDA receptor-dependent synapticplasticity in the hippocampus has been blocked (by AP5) duringthe revaluation stage (compare the elemental results in Fig. 3bwith the configural results 3e). However, unlike muscimol, whichblocks synaptic transmission, AP5 had no effect when adminis-tered immediately before the test (see Fig. 3g; Iordanova et al.,2011b). These observations demonstrate that the hippocampus isinvolved in the updating of configural memories with newinformation (during the revaluation stage) and the retrieval ofthese updated memories (at test); but they leave open the role ofthe hippocampus (and NMDA receptor-dependent plasticity) inthe initial formation of the configural representations. Notwith-standing this observation, the robust nature of the behavioraldissociations suggests that we need to give serious considerationto the idea that both elemental and configural associativestructures underpin associative learning, with the important ca-veat that this dissociation is probably restricted to some classesof stimuli and training procedures (cf. Coutureau, Killcross, Good,Marshall, Ward-Robinson & Honey, 2002; Good & Honey, 1991;Honey & Good, 1993). This suggestion raises many issues. Forexample, are the two types of associative structure entirelyseparate or do they interact in the way assumed by Rudy and

a b c

d e f

g h

Fig. 3. Dissociating elemental and configural learning. Panels a–c; mean freezing ratios during the test in rats given elemental training: (a) hippocampal (HPC) or shamlesions before elemental training (separately presented for context and time), (b) artificial cerebrospinal fluid (aCSF), muscimol (Mus) or AP5 immediately prior to revaluation(combined results for rats given context or time elemental training), and (c) aCSF or Mus immediately prior to the test (separately presented for context and time). Panels d-h;mean freezing ratios during the test in rats given configural training: (d) HPC or sham lesions prior to configural training; (e) aCSF or AP5 immediately prior to revaluation; (f)aCSF or Mus immediately prior to revaluation; (g) aCSF or AP5 immediately prior to the test, and (h) aCSF or Mus immediately prior to the test.

100 R.C. Honey et al. / Neurobiology of Learning and Memory 108 (2014) 96–103

Sutherland (1989) and Kehoe (1988; Iordanova et al., 2011b;McLaren, Forrest, & McLaren, 2012).

5. Hybrid associative systems: Parallel or interactive

The view that a complete understanding of associative learningshould include qualitatively distinct elemental and configural asso-ciations might be contentious for those whose primary focus is onunderstanding the conceptual nervous system (but see McLarenet al., 2012); however, it is less so in the field of behavioral neuro-science (e.g., Rudy & Sutherland, 1989). 1 Taking this view seriouslyprompts a number of subsidiary questions. For example, do the twosystems ordinarily operate in parallel or do they interact, and if they

1 Indeed, it remains logically possible there are also distinct configural learningprocesses (see Coutureau et al., 2002; see also, Cowell, Bussey, & Saksida, 2010); orparallel (more-or-less complex) elemental systems (cf. Wagner, 2003), that aredifferently reliant on the hippocampus (and other neural structures). Notwithstand-ing these additional complexities, the dissociations that we have observed stillprovide a firm basis for appealing to separable (possibly interacting) associativesystems.

do interact then how does this interaction play out. In fact, the resultsof other studies suggest that what is learnt during training proce-dures that are operationally elemental interacts with what is learntduring procedures that are operationally configural (e.g., Allman &Honey, 2006; Alvarado & Rudy, 1992; Honey & Ward-Robinson,2002; Lin & Honey, 2010; Schreurs & Kehoe, 1987). These results be-gin to suggest that if there are elemental and configural systems thenthey are likely to interact. However, we begin by considering the sim-pler suggestion that the two systems operate entirely separately, inparallel, and examine the application of this idea to the results thathave been the focus of interest thus far. Is there internal evidence,within the dissociations described by Iordanova et al. (2009 and2011a,b), that is helpful in determining whether or not the elementaland configural systems operate in parallel?

5.1. Parallel systems

The simplest analysis that is licensed by the dissociations de-scribed by Iordanova et al. (2009 and 2011a,b), assumes that thereare separate elemental and configural systems that run in parallel,

a

c

b

Fig. 4. Hybrid associative structures with (a and b) parallel independent elementaland configural systems, or (c) a ‘default’ elemental system that recruits a configuralsystem. AM refers to morning, A to a context, and X to an auditory stimulus. Dottedlines indicate acquired links, and solid lines indicate structural links that carryactivation in proportion to the error at each of the nodes that are activated by thethree components (AM, A and X); additional input to the configural system arisesfrom error resulting from the associative activation of components that are absent(and not shown).

R.C. Honey et al. / Neurobiology of Learning and Memory 108 (2014) 96–103 101

with one involving the hippocampus and the other not. Accordingto this analysis, a given pattern of stimulation activates two sepa-rate populations of elements, with one set being connected to oneanother by direct links (see Fig. 4a), and the other set being boundto one another by becoming separately linked to a configural rep-resentation (see Fig. 4b). In Fig. 4a and b, the elements that provideinput to both systems are separate, but otherwise the same;although this need not be the case. Both systems would be capableof separately representing patterns during training that can beoperationally defined as elemental, but only the configural systemcould do so when the training is operationally configural. The con-sequences of exposing this hybrid parallel system to elemental andconfigural training are asymmetric.

First consider the consequences of exposing both systems toconfigural training: AM + A + X, PM + A + Y, AM + B + Y andPM + B + X. This exposure will result in the formation of a net ofpair-wise associations within the elemental system: AM-A, AM-B,PM-A, PM-B, AM-X, AM-Y, PM-X, PM-Y, A-X, A-Y, B-X, and B-Y;and it will result in the formation of four configural representa-tions: AM + A + X, PM + A + Y, AM + B + Y and PM + B + X. After Xhas been paired with shock, all of the test compounds (AM + A,PM + A, AM + B and PM + B) will be capable of eliciting fear throughthe elemental links that their components have with X (and there-by shock; see also Lin, Dumigan, Dwyer, Good, & Honey, 2013;Ward-Robinson & Hall, 1996). And, the test compounds AM + Aand PM + B will also strongly activate configural units linked toshock (i.e., those involving X: AM + A + X and PM + B + X). Conse-quently, disrupting the operation of the configural system willmean that while the test compounds will not elicit different levelsof fear, they will nevertheless elicit fear. So far, so good: The con-texts did provoke appreciable fear after the various manipulationsof the hippocampus. Now consider the consequences of exposingboth systems to elemental training (e.g., AM + A + X, AM + B + Y,PM + A + X, PM + B + Y). This training will result in strong elemen-tary associations: A-X and B-Y; and other weaker associations. Itwill also result in configural representations involving each ofthe compounds. Now, after conditioning trials with X, there willbe two bases for the test compounds AM + A and PM + A to elicitmore fear than AM + B and PM + B: (1) because A is strongly linkedto X, and (2) because AM + A and PM + A are similar to configuralrepresentations involving X. So, disrupting the operation of theconfigural system (e.g., through lesions, inactivation during reval-uation or test, or blocking synaptic plasticity during revaluation)in rats given elemental training should also mean that the criticaltest compounds would elicit less fear. However, as we have alreadyseen, there was no evidence that disrupting the function of the hip-pocampus reduced test performance in rats given elemental train-ing (see Iordanova et al. (2009 and 2011a,b). These observationsmight reflect some (consistent) lack of sensitivity in our test proce-dures to the effects of a variety of neural manipulations; but thisseems implausible given the fact that there appeared to be no ceil-ing or floor effects. Without explicit performance rules for how ele-mental and configural knowledge is combined to influencebehavior (in this case fear), then predictions about test perfor-mance that is underpinned by two systems rather than one aremoot. However, the clear set of dissociations that we have ob-served suggest that we should at least consider the merits of a hy-brid associative system in which elemental and configuralprocesses interact with one another. We consider one version ofsuch an interactive system below.

5.2. Interactive systems

One potential way in which elemental and configural systemmight interact is depicted in Fig. 4c. This analysis is based partly

upon the general intuition that the default associative learningprocess is elemental, and that configural learning processes be-come increasingly likely to be recruited when elemental structureshave proven error-prone. This error could occur in the context offorming separate representations of patterns of stimulation involv-ing the components of episodic memory (which auditory stimuluswas presented where and when; Iordanova et al., 2008) or predict-ing the outcome of a trial (e.g., food or no food) in standard confi-gural discriminations. This loosely specified proposal can be madeless so by imagining how a purely elemental system might behavewhen it is confronted with a problem that is operationally configu-ral: AM + A + X, PM + A + Y, AM + B + Y and PM + B + X.

To begin with, direct associations will form between the com-ponents of each trial (e.g., AM, A and X); but the efficacy of theseassociations will be weakened on the other three trial types (e.g.,PM + A + Y), and these components (i.e., AM, A and X) will enterinto association with those presented on the other three trialstypes (e.g., PM, B and Y). The result of these training trials will bethat when the original trial type is encountered once again(AM + A + X) the pattern of activation generated by the input pat-tern will not match that generated by the current pattern of asso-ciations: the associations among the components of the trial (A, Xand food) will be weak, and these components will also activateother components that are not present. In short, during a configu-ral training procedure there will be a persistent difference betweenthe components that are present on a given trial and those that areassociatively provoked (i.e., there will be a mismatch or error). Un-der these conditions, a configural system might receive activationfrom the elemental system – allowing the elements that populatethis system to be represented configurally (see Fig. 4c). In Fig. 4c,the internal error for each component (the difference betweenthe activity generated by the presentation of the input patternand that generated by the links among the components) is shownas directly influencing the configural unit (AM A X) via structural(i.e., non-plastic) links; and the error generated by activated com-ponents that are absent can be imagined to do likewise. In such anetwork, elemental training will be less likely to result in both

102 R.C. Honey et al. / Neurobiology of Learning and Memory 108 (2014) 96–103

elemental and configural learning, and it will be possible to ob-serve a clear dissociation between them.2

One interesting feature of the preceding analysis is that at theoutset of training there will always be a mismatch between thepattern of input created by the presentation of a pattern of stimu-lation (e.g., AM + A + X) and the internally generated activity (with-in the population of elements, AM, A and X) – because the strengthof any links will not match the input to the system. Provided it isthe case that this mismatch is sufficiently large then a given pat-tern will always be represented in the configural system and theelemental system early in training. This analysis thereby allowsthe possibility that the hippocampus might be recruited duringboth elemental and configural training, at least initially (cf. Rudy& O’Reilly, 2001; see also, Nadel & Willner, 1980) – a possibilitythat is consistent with findings suggesting that rapid learningabout contexts can be dependent on the hippocampus (e.g., Kim& Fanselow, 1992; Rudy, 2009; but see, Good & Honey, 1997;McNish, Gewirtz, & Davis, 1997). Notwithstanding this possibility,it is only with more extensive training that a persistent differencewill emerge between the error produced in the elemental systemby elemental and configural procedures; and it is this differencethat will result in the ongoing or increased recruitment of the con-figural system.

According to the informal analysis developed in the precedingparagraphs, there are two potential theoretical loci where the hip-pocampus might play a significant role: In aspects of the process ofmismatch or error detection, that determines the degree to whichthe configural system will be recruited (see Sokolov, 1963; Vinog-radova, 1970; see also, Honey, Watt & Good, 1998; Honey & Good,2000a, 2000b) or in instantiating (at least on a temporary basis) as-pects of the configural structure itself. The former possibility ismost consistent with the finding that blocking NMDA receptor-dependent plasticity in the hippocampus during the (elemental)revaluation stage selectively disrupts new learning involving con-figural memories (Iordanova et al., 2011b); while the latter re-ceives most obvious support from the fact that blocking synaptictransmission during the test disrupts performance based on confi-gural training (Iordanova et al., 2009). In either case, the fact thatthe elemental system is the default means that performance afterelemental training will be relatively unaffected by disruption tothe hippocampus (with the possible exception of during initialtraining trials), whereas there will be a clear disruption to perfor-mance that is reliant on configural processes.

6. Concluding comments and caveats

We have described a set of studies that have yielded a clear dis-sociation between SPC procedures that can be operationally de-fined as either elemental or configural. The fact that thehippocampus was required for configural but not elemental vari-ants of SPC has been foreshadowed: if not firmly established by awealth of direct data, then certainly in computational analyses(e.g., Rudy & Sutherland, 1989). Thus, evidence concerning the roleof the hippocampus in standard configural discriminations turnsout to be quite inconsistent (see Coutureau et al., 2002), and sub-ject to the criticisms that the elemental and configural tasks are of-

2 It might also seem possible to develop a yet more integrated elemental–configural learning system wherein the inputs (or at least some of them) are part ofthe same hybrid system (see McLaren et al., 2012), and the error is propagateddirectly to the configural units through the input-hidden layer associative links. It isworth noting, however, that default nature of the elemental system and theseparation of the two systems depicted in Fig. 4 allows one to avoid the predictionthat manipulating the hippocampus will disrupt a SPC effect that follows elementaltraining. As we have seen, this prediction is a consequence of allowing parallelelemental and configural structures to be acquired during elemental training – aprediction that is also difficult for a fully interactive system to avoid.

ten poorly matched (e.g., Good & Honey, 1991). However, the set ofresults reported by Iordanova et al. (2009 and 2011a,b) begin toprovide more compelling grounds for appealing to distinct elemen-tal and configural systems – an appeal that is anathema to promi-nent theoretical treatments of associative learning (e.g., Pearce,1994; Wagner, 2003). The potential theoretical value of this setof results highlights the need to establish their generality. Arethe dissociations also observed when the critical stimuli (what,where and when) are employed in more conventional elementaland configural discriminations (cf. Coutureau et al., 2002; Cowellet al., 2010); or when the different variants of SPC involve stimulithat are not closely allied to episodic memory? Leaving these ques-tions to one side, the extant evidence begins to place some con-straints on the nature of any hybrid elemental-configuralassociative system. We have argued, admittedly on the basis of rel-ative sparse and indirect evidence, that the two systems are unli-kely either to operate in parallel or to be fully interactive.Instead, the clear-cut nature of the dissociations suggests thatthe two systems participate in an asymmetric interaction, wherea configural system is especially likely to be recruited when theelemental system is error prone.

Acknowledgments

The research reported in this article involving the authors wassupported by grants from the BBSRC UK. Correspondence concern-ing this article should be addressed to: R.C. Honey, School of Psy-chology, Cardiff University, Tower Building, Park Place, Cardiff,CF10 3AT, UK.

References

Allman, M. J., & Honey, R. C. (2006). Transfer of configural learning between thecomponents of a preexposed stimulus compound: Implications for elementaland configural models of learning. Journal of Experimental Psychology: AnimalBehavior Processes, 32, 307–313.

Alvarado, M. C., & Rudy, J. W. (1992). Some properties of configural learning: Aninvestigation of the transverse-patterning problem. Journal of ExperimentalPsychology: Animal Behavior Processes, 18, 145–153.

Atkinson, R. C., & Estes, W. K. (1963). Stimulus sampling theory. In R. D.Luce, R. B.Bush, & E. Galanter (Eds.), Handbook of mathematical psychology (Vol. 3, pp.121–268). New York: Wiley.

Brogden, W. J. (1939). Sensory pre-conditioning. Journal of Experimental Psychology,25, 323–332.

Clayton, N. S., & Dickinson, A. (1998). Episodic-like memory during cache recoveryby scrub jays. Nature, 395, 272–274.

Coutureau, E., Killcross, A. S., Good, M., Marshall, V. J., Ward-Robinson, J., & Honey, R.C. (2002). Acquired equivalence and distinctiveness of cues: II. Neuralmanipulations and their implications. Journal of Experimental Psychology:Animal Behavior Processes, 28, 388–396.

Cowell, R. A., Bussey, T. J., & Saksida, L. M. (2010). Components of recognitionmemory: Dissociable cognitive processes or just differences in representationalcomplexity? Hippocampus, 20, 1245–1262.

Eacott, M. J., Easton, A., & Zinkivskay, A. (2005). Recollection in an episodic-likememory task in the rat. Learning & Memory, 12, 221–223.

Good, M., & Honey, R. C. (1991). Conditioning and contextual retrieval inhippocampal rats. Behavioral Neuroscience, 105, 499–509.

Good, M., & Honey, R. C. (1997). Dissociable effects of selective lesions tohippocampal sub-systems on exploratory behavior, contextual learning andspatial learning. Behavioral Neuroscience, 111, 487–493.

Good, M. A., Barnes, P., Staal, V., McGregor, A., & Honey, R. C. (2007). Context- butnot familiarity-dependent forms of object recognition are impaired followingexcitotoxic hippocampal lesions in rats. Behavioral Neuroscience, 121, 218–223.

Honey, R. C., Close, J., & Lin, T. E. (2010). Acquired distinctiveness and equivalence: Asynthesis. In C. J. Mitchell & M. E. Le Pelley (Eds.), Attention and associativelearning: From brain to behaviour (pp. 159–186). Oxford: Oxford UniversityPress.

Honey, R. C., & Good, M. (1993). Selective hippocampal lesions abolish thecontextual specificity of latent inhibition and conditioning. BehavioralNeuroscience, 107, 23–33.

Honey, R. C., & Good, M. (2000a). Associative components of recognition memory.Current Opinion in Neurobiology, 10, 200–204.

Honey, R. C., & Good, M. (2000b). Associative modulation of the orienting response:Distinct effects revealed by hippocampal lesions. Journal of ExperimentalPsychology: Animal Behavior Processes, 26, 3–14.

R.C. Honey et al. / Neurobiology of Learning and Memory 108 (2014) 96–103 103

Honey, R. C., & Ward-Robinson, J. (2002). Acquired equivalence and distinctivenessof cues: I. Exploring a neural network approach. Journal of ExperimentalPsychology: Animal Behavior Processes, 28, 378–387.

Honey, R. C., Good, M., & Manser, K. L. (1998). Negative priming in associativelearning: Evidence from a serial-habituation procedure. Journal of ExperimentalPsychology: Animal Behavior Processes, 24, 229–237.

Honey, R. C., Watt, A., & Good, M. (1998). Hippocampal lesions disrupt anassociative mismatch process. Journal of Neuroscience, 18, 2226–2230.

Hull, C. L. (1943). Principles of behavior. New York: Appleton-Century-Crofts.Iordanova, M. D., Good, M., & Honey, R. C. (2008). Configural learning without

reinforcement: Integrated memories for what, where and when. QuarterlyJournal of Experimental Psychology, 61, 1785–1792.

Iordanova, M., Burnett, D., Aggleton, J. P., Good, M., & Honey, R. C. (2009). The role ofthe hippocampus in mnemonic integration and retrieval: Complementaryevidence from lesion and inactivation studies. European Journal of Neuroscience,30, 2177–2189.

Iordanova, M. D., Burnett, D., Good, M., & Honey, R. C. (2011a). Pattern memoryinvolves both elemental and configural processes: Evidence from the effects ofhippocampal lesions. Behavioral Neuroscience, 125, 567–577.

Iordanova, M. D., Good, M., & Honey, R. C. (2011b). Retrieval-mediated learninginvolving episodes requires synaptic plasticity in the hippocampus. Journal ofNeuroscience, 31, 7156–7162.

Kehoe, E. J. (1988). A Layered network model of associative learning: Learning tolearn and configuration. Psychological Review, 95, 411–433.

Kim, J. J., & Fanselow, M. S. (1992). Modality-specific retrograde memory for fear.Science, 256, 675–677.

Konorski, J. (1967). Integrative activity of the brain: An interdisciplinary approach.Chicago: University of Chicago Press.

Lin, T. E., & Honey, R. C. (2010). Analysis of the content ofconfigural representations: The role of associatively evoked and tracememories. Journal of Experimental Psychology: Animal Behavior Processes, 36,501–505.

Lin, T. E., & Honey, R. C. (2011). Encoding specific associative memory: Evidencefrom behavioral and neural manipulations. Journal of Experimental Psychology:Animal Behavior Processes, 37, 317–329.

Lin, T. E., Dumigan, N. M., Dwyer, D. M., Good, M. A., & Honey, R. C. (2013). Assessingthe encoding specificity of associations with sensory preconditioningprocedures. Journal of Experimental Psychology: Animal Behavior Processes, 39,67–75.

Mackintosh, N. J. (1974). The psychology of animal learning. London: Academic Press.

McLaren, I. P. L., Forrest, C. L., & McLaren, R. P. (2012). Elemental representation andconfigural mappings: Combining elemental and configural theories ofassociative learning. Learning & Behavior, 40, 320–333.

McNish, K. A., Gewirtz, J. C., & Davis, M. (1997). Evidence of contextual fearconditioning following lesions of the hippocampus: A disruption of freezing butnot fear-potentiated startle. Journal of Neuroscience, 17, 9353–9360.

Nadel, L., & Willner, J. (1980). Context and conditioning: A place for space.Physiological Psychology, 8, 218–228.

Palmer, A., & Good, M. A. (2011). Hippocampal synaptic activity, pattern separationand episodic-like memory: Implications for mouse models of Alzheimer’spathology. Biochemical Society Transactions, 39, 902–909.

Pavlov, I.P. (1935). The Conditioned Reflex. In Lectures on Conditioned Reflexes:Conditioned Reflexes and Psychiatry (1941, Volume 2, p. 167). London: Lawrence& Wishart Ltd.

Pearce, J. M. (1994). Similarity and discrimination: A selective review and aconnectionist model. Psychological Review, 94, 61–73.

Rudy, J. W., & Sutherland, R. J. (1989). The hippocampal formation is necessary forrats to learn and remember configural discriminations. Behavioural BrainResearch, 34, 97–109.

Rudy, J. W. (2009). Context representations, context functions, and theparahippocampal–hippocampal system. Learning & Memory, 16, 573–585.

Rudy, J. W., & O’Reilly, R. C. (2001). Conjunctive representations, the hippocampus,and contextual fear conditioning. Cognitive, Affective, and BehavioralNeuroscience, 1, 66–82.

Schreurs, B. G., & Kehoe, E. J. (1987). Cross-modal transfer as a function of initialtraining level in classical conditioning with the rabbit. Animal Learning &Behavior, 15, 47–54.

Sokolov, E. N. (1963). Perception and the conditioned reflex. London: Pergamon Press.Thorndike, E. L. (1911). Animal intelligence: Experimental studies. New York:

Macmillan.Vinogradova, O. S. (1970). Registration of information and the limbic system. In G.

Horn & R. A. Hinde (Eds.), Short-term changes in neural activity and behaviour(pp. 95–148). Cambridge, England: Cambridge University Press.

Wagner, A. R. (2003). Context-sensitive elemental theory. Quarterly Journal ofExperimental Psychology, 23B, 7–29.

Wagner, A. R., & Rescorla, R. A. (1972). Inhibition in Pavlovian conditioning:Application of a theory. In R. A. Boakes & M. S. Halliday (Eds.), Inhibition andlearning (pp. 301–336). London: Academic Press.

Ward-Robinson, J., & Hall, G. (1996). Backward sensory preconditioning. Journal ofExperimental Psychology: Animal Behavior Processes, 22, 395–404.