INTACT ARTIFICIAL GRAMMAR LEARNING IN AMNESIA: Dissociation of Classification Learning and Explicit Memory for Specific Instances

The present study investigates whether the ability to classify' on the basis of rules can be learned independently of memory for the specific instances used to teach the rules. Thirteen amnesic patients and 14 control subjects studied letter strings generated by an artificial grammar. Subjects were then shown new letter strings and were instructed to classify them as grammatical or nongrammaticai. Amnesic patients performed as well as normal subjects. However, amnesic patients performed more poorly than control subjects on a recognition test of the exemplars that had been presented. Amnesic patients also performed more poorly than control subjects when the instructions were to base the classification on explicit comparison with the original exemplars. The results show that classification learning based on exemplars of an artificial grammar can develop normally despite impaired memory for the exemplars themselves. Whereas exemplar memory depends on interactions between neocortex and the limbic system, classification learning may depend on interaction between neocortex and the neostriatum. Twenty-five years ago, Reber (1967) suggested that normal subjects can learn to classify letter strings correctly without developing explicit knowledge about the basis for the classification. The "correct" letter strings were generated by an artificial, finite-state grammar. The key finding was that after inspecting a group of exemplars that adhered to the grammatical rules, subjects were able to classify new items as either "grammatical" or "nongrammaticai" at well above chance levels, even when the existence of rules underlying the exemplars was not mentioned until after the exemplars had been presented. In a series of papers (see Reber, 1989, for review), Reber suggested that successful classification is based on implicit memory and that subjects acquire the basis for making correct classifications without having explicit, conscious access to their knowledge. This view has been challenged with the proposal that performance on this task is based on the conscious application of explicit (declarative) memory strategies that are imperfectly formed and only partially correct (Dulany, Carlson, & Dewey, 1984; Perruchet & Pacteau, 1990). This issue thus concerns the question of how the ability to classify based on a fixed set of rules arises from specific experiences. One view is that the ability to classify develops gradually as instances are presented, and that the acquired knowlAddress correspondence to Larry R. Squire, Department of Psychiatry, Veterans Affairs Medical Center, 3350 La Jolla Village Dr., San Diego. CA 92161, edge is implicit and distinct from explicit memory for the individual instances. Another view is that the information that supports correct classification has no special status. It is constructed out of remembered instances and is available as explicit, conscious memory. Similar viewpoints have been expressed regarding the learning of natural categories, that is, categories not defined by a fixed set of rules. In the first view, category knowledge is distinct from memory for individual exemplars (Franks & Bransford, 1971; Hayes-Roth & HayesRoth, 1977; Homa & Chambliss, 1975; Posner & Keele, 1968, 1970; Reed, 1972). This view allows for (but does not require) the possibihty that category level knowledge could develop entirely independently of exemplar memory. In the other view, knowledge about category membership is derived from and is based directly on the exemplars stored in memory (Hintzman, 1986; McClelland & Rumelhart, 1986; Medin & Schaffer, 1978; for review, see Estes, 1988). By this view, performance on tests directed at category level knowledge depends on and should always be associated with performance on tests of exemplar memory. Studies of amnesic patients could decide between these two views. Amnesic patients, despite being severely impaired on conventional tests of recall and recognition, are fuUy intact on many other tasks of learning and memory (Hintzman, 1990; Mayes, 1988; Squire, 1987; Weiskrantz, 1987). These facts can be understood by supposing that amnesia impairs the ability to acquire one kind of memory, that is, memory for facts and events (termed declarative or explicit memory). What is spared in amnesia is a heterogeneous group of other abilities (collectively termed nondeclarative or implicit memory), which depend on structures not damaged in amnesia, including neocortex, striatum, cerebellum, and amygdala. These other abilities have been described as skillful behaviors, conditioning and habit formation, and the phenomenon of priming (Squire, 1987; Tulving & Schacter, 1990). Among the tasks that amnesic patients have been shown to acquire normally are perceptuomotor skills (Brooks & Baddeley, 1976; Nissen & BuUemer, 1987), perceptual skills like mirror reading (Cohen & Squire, 1980), perceptual and semantic priming (Graf, Squire, & Mandler, 1984; Schacter, Cooper, Tharan, & Rubens, 1991; Shimamura & Squire, 1984), adaptation level effects (Benzing & Squire, 1989), and changes in preference and judgment that are likely based on priming (Johnson, Kim, & Risse, 1985; Squire & McKee, 1992). Nondeclarative (implicit) memory is nonconscious. Thus, performance changes as the result of experience without providing conscious access to specific prior episodes or to any memory content. In the present study, we asked whether artificial grammar 172 Copyright © 1992 American Psychological Society VOL. 3, NO. 3, MAY 1992 PSYCHOLOGICAL SCIENCE Barbara J. Knowlton, Seth J. Ramus, and Larry R. Squire learning is an example of the kind of implicit learning that is spared in amnesia. If the ability to make grammaticality judgments after studying exemplars is based on implicit memory that is not accessible to consciousness, and if explicit memory does not materially contribute to these judgments, then amnesic patients should be able to make grammaticality judgments as well as normal subjects. Alternatively, if the ability to make such judgments depends on the explicit use of imperfectly formed rules or direct comparisons with stored exemplars, then amnesic patients should perform poorly.