Learning-based constraints on schemata

Abstract

Schemata are frequently used in cognitive science as a descriptive framework for explaining the units of knowledge. However, the specific properties which comprise a schema are not consistent across authors. In this paper we attempt to ground the concept of a schema based on constraints arising from issues of learning. To do this, we consider the different forms of schemata used in computational models of learning. We propose a framework for comparing forms of schemata which is based on the underlying representation used by each model, and the mechanisms used for learning and retrieving information from its memory. Based on these three characteristics, we compare examples from three classes of model, identified by their underlying representations, specifically: neural network, production-rule and symbolic network models.