Human Learning and Memory

ION Abstraction is the process by which the input links to a chunk node from other nodes are selectively strengthened and weakened through repeated experience with different instances of the concept. These input links should include the specification of constituent, propositional, and procedural meaning of concepts, since the relations between concepts often appear to be learned very early (e.g. chairs are to sit on). This strength learning hypothesis is both reason?ble and consistent with associative net­ work theory. It would be highly unparsimonious to opt for some prototype learning theory that cannot be incorporated easily within the more general associa­ tive network theory. For example, one way to store an average prototype without storing examples is to add the value of each attribute dimension to a counter, add its square to another counter, and add one to a number-of­ instances counter. This permits determining the average value and the variance on each feature dimension. When people speak of storing the average prototype without storing the examples, this must be the sort of process they have in mind. An associative computer could be built to do this, but it is a less accurate way to record the regions in a multidimensional attribute space that are associated with any given concept than the strength mechanism previously described. Such an average prototype abstraction mechanism could learn concepts whose examples cluster about a single prototype, such as "apple" or "cat, " but would be quite inadequate to learn superordinate concepts such as "fruit" and "mammal" that do not cluster about a single prototype. Furthermore, the more adequate instance strength abstraction mechanism is also more natural for an associative memory, while the prototype abstraction mechanism would be much more complex to implement. Recently, evidence has accumulated to support the hypothesis that for adults the best example of a concept is sometimes composed of the modal (most frequently occurring) values on each attribute dimension rather than the average values (Goldman & Homa 1977, Neumann 1977, Strauss 1979), though so far infants have been demonstrated to abstract only averages HUMAN LEARNING AND MEMORY 35 (Strauss 1979). Whether adults abstract modes or averages depends on the discriminability of the values on a dimension-highly discriminable values cause the mode to have the greatest strength of association to the concept, less discriminable values cause the average value to have the greatest strength of association to the concept. As Neumann (1977) and Wickelgren (1979a, p. 310) point out, this can be accounted for by old-fashioned stimu­ lus generalization in an associative network theory. No other unified theory, to my knowledge, explains the abstraction of averages when discriminabil­ ity is low and modes when discriminability is high. Starting most recently with Brooks (1978), there is increasing acceptance of the hypothesis that the learning of a concept derives from encoding each instance more or less faithfully, subject to selective attention in learning and to the interference and facilitation in learning, storage, and retrieval that inevitably results from encoding in an associative memory. The most plausible theory at present is that the abstracted constituent structure of a concept is the result of accumulated association of the attributes of events that cued that con­ cept. In addition, inhibition is also involved in some manner so that the at­ tributes that have the strongest association to a concept are those that discriminate best between this concept and other concepts (Rosch & Mervis 1975). See also Wickelgren (1979a, pp. 3 11-12) for an associative explana_ tion of this using the Spencian overlapping excitatory and inhibitory gener­ alization gradients mechanism. BASIC CONCEPTS If it hadn't been so sad it would have been funny to hear psychologists at one time asserting that the earliest concepts learned by children were the most general ones followed by increasing conceptual differentiation, and at other times asserting that the earliest learned con­ cepts were the most specific ones followed by concepts of increasingly greater generality. Brown (1958) pointed out that neither extreme was true and that the earliest concepts learned by children are of intermediate refer­ ential generality, and, as discussed earlier, subsequent research has con­ firmed this conclusion. Brown suggested that the reason for this was greater frequency of use of these concepts by other humans in the child's environ­ ment, and this is surely an important component of the explanation for why concepts like "cat" and "apple" are basic instead of "mammal" or "fruit" on the one hand or "Siamese cat" or "Jonathan apple" on the other. However, it was not until the elegant insight of Rosch (Rosch et al 1976, Rosch 1977) that we understood this phenomenon adequately. According to Rosch, basic concepts are the most general concepts whose examples are sufficiently similar that they hover around a single prototype. This high degree of example similarity means extensive strength generalization in the