ABSTRACT
A number of computational theories of human memory that appear in the current literature share three essential characteristics (e.g., Anderson, Silverstein, Ritz, & Jones, 1977; Eich, 1982; Kohonen, 1988; McClelland & Rumelhart, 1985; Murdock, 1982). First, they view memory as being composed of a large number of relatively simple units operating in parallel. Second, they retain information by distributing the representation of each to-be-remembered item over a large number of these units such that each unit can be involved in the representation of many items. Third, each theory proposes a scheme for the learning of representations through modification of unit values, or of connections between units, during item presentation. The major advantage of these theories is their correspondence to certain properties of human concept formation. Through experience with individual instances, they naturally form abstract categories based on family resemblance and generalize these categories to new instances.
