Cognitive Systems Research 3 2002 45 55 www elsevier com locate cogsys An attractor network model of serial recall Action editors Wayne Gray and Christan Schunn Matt Jones Thad A Polk Department of Psychology Cognition and Perception University of Michigan 525 East University Ann Arbor MI 48109 USA Received 1 March 2001 accepted 1 September 2001 Abstract We present a neural network model of verbal working memory which attempts to illustrate how a few simple assumptions about neural computation can shed light on cognitive phenomena associated with the serial recall of verbal material We assume that neural representations are distributed that neural connectivity is massively recurrent and that synaptic efficacy is modified based on the correlation between pre and post synaptic activity Hebbian learning Together these assumptions give rise to emergent computational properties that are relevant to working memory including short term maintenance of information time based decay and similarity based interference We instantiate these principles in a specific model of serial recall and show how it can both simulate and explain a number of standard cognitive phenomena associated with the task including the effects of serial position word length articulatory suppression and its interaction with word length and phonological similarity 2002 Elsevier Science B V All rights reserved Keywords Verbal working memory Serial recall Learning Memory Attractor network Recurrent neutral network Phonological similarity Primacy Recency Articulatory suppression Word length 1 Introduction Working memory is among the most intensively studied cognitive processes in both cognitive psychology and neuroscience and yet results from the two fields have not made as much contact with each other as one might hope For example cognitive psychology has discovered a host of robust empirical phenomena associated with verbal working memory and has developed elegant theoretical models such as Baddeley s phonological loop that can explain the empirical results Baddeley 1986 Nevertheless the details of how these psychological hypotheses are Corresponding author E mail address mattj umich edu M Jones instantiated in the brain is an open question but see Burgess Hitch 1999 for one recent proposal Similarly there is a substantial body of neuroscientific research investigating the neural substrates of working memory in both animals Fuster 1973 Funahashi Bruce Goldman Rakic 1989 and humans Smith Jonides 1999 but this work has typically only addressed a small subset of the rich behavioral data and theories available in cognitive psychology In this paper we attempt to illustrate that a simple and independently motivated model of neural computation can make contact with and even shed light on the cognitive psychology of verbal working memory We begin by describing a few widely accepted assumptions about neural computation 1389 0417 02 see front matter 2002 Elsevier Science B V All rights reserved PII S1389 0417 01 00043 2 46 M Jones T A Polk Cognitive Systems Research 3 2002 45 55 Next we discuss some of the emergent computational properties of these assumptions that are relevant to verbal working memory e g maintenance decay interference We then illustrate how these assumptions can be instantiated in a specific computational model that simulates and explains many of the major psychological phenomena associated with the serial recall task 2 A simple model of neural computation We begin with three simple and widely accepted assumptions about neural computation The first is that representations in the cortex are generally distributed across a population of neurons rather than being localized to individual cells The second is that there is massive connectivity among neurons within local areas of cortex and that this connectivity is recurrent rather than unidirectional The third assumption is that synaptic efficacy is modified based on the correlation between pre and postsynaptic activity Hebbian learning Taken together these assumptions give rise to networks with interesting emergent properties many of which are relevant to working memory For example such networks are known to be capable of maintaining an activation pattern via internal reverberatory activity even after the input to the network has been removed Hopfield 1982 Those patterns which the network can maintain in this way are termed attractors and hence the networks themselves are known as attractor networks and under the Hebbian learning rule they tend to emerge as those patterns to which the network is repeatedly exposed Furthermore when presented with a noisy or incomplete version of a previously trained pattern an attractor network will tend to converge its activity upon that attractor state which is most similar to the input thereby retrieving the original pattern Another property of attractor networks that is relevant to working memory is that they naturally exhibit similarity based interference Attractor networks are capable of storing multiple patterns as attractor states but if those patterns are similar to each other overlap substantially then there is a greater likelihood of error In particular we have found that with noisy input these networks can often retrieve an incorrect attractor pattern but one which tends to be similar to the correct one Finally we have also found that attractor networks can be easily extended to exhibit time based decay In the original formulation of attractor networks each unit was binary either ON or OFF and activation patterns could be maintained for indefinite periods of time Hopfield 1982 Hopfield 1984 subsequently showed that networks using more realistic continuous valued units could also exhibit similar computational properties Our investigations have shown that by appropriately increasing the threshold of the individual units continuous valued attractor networks can be made to exhibit time based decay once external input is removed 1 3 The serial recall task In the standard serial recall task a subject is presented either visually or auditorially with a sequence of items most often words letters or digits Once presentation of the list has been completed the task of the subject is to repeat back the list in its original order either by speaking or by writing This task has been intensively studied and a large number of robust behavioral phenomena have been identified Below are some of the major phenomena which we will address in
View Full Document
Unlocking...