Proceedings of the Annual Meeting of the Cognitive Science Society

Several dynamical systems have been previously pTX>posed to give a neural-Iike (i.e. connectionist)description of category formation. These typically either involve supervised training (as in Sutton &.Barto, 1981; Reilly et al., 1982) or identify dense regions ("clusters') in the stimulus distribution asnatural categories (Amari & Takeuchi, 1978; Rumelhart & Zipser, 198S). By combining two existingconnectionist-type learning procedures, one supervised and one unsupervised, a hybrid 'self-supervisedleamingf (SSL) mechanism for concept and category learning has been developed. Each unit in thenetwork comes to represent some concept of the order of complexity of a single word; the activity ofthe unit signals the contribution of its associated concept to the current mental state. A crucialassumption of this i^proach is that every concept unit (C-unit) receives inputs from two or moreinformation streams. The self-supervised learning process is governed by a data-driven dynamical rulewhich results in a two-stage learning process. In the first stage, a C-unit becomes selectively responsiveto a particular pattern •"'• from one of the information streams, ignoring all other patterns in thatstream. This is followed by an associative stage in which the unit develops graded response propertiesto stimulus patterns incident from the other information stream(s). The trigger feature thus becomes akind of prototype for the concept to be formed by the C-unit. Populations of C-units display interesting representational properties; these are seen to have attributes of both local and distributed representations.