Abstract

The problem of categorization is examined, with an interest in assessing the impact on theories of mental architecture. The point of departure is Fodor's modular model of mind, within a classical symbolic First Order system of Knowledge Representation. Weaknesses in the classical model are presented, both with respect to computational complexity and failure to explain concept acquisition and plasticity. Current associationistic alternatives are presented, including Bayesian model-based reasoning and connectionist Boltzmann machines. Arguments from the classical perspective for the unacceptability of connectionist architectures for modeling higher cognitive functions are examined. Evidence for the feasibility of incorporating recursive data structures in connectionist nets considerably weakens the classical position. Implementation of a modularist architecture within a connectionist framework is recommended to adequately model normal human higher cognitive functions and to explain the phenomena of informational encapsulation. Adoption of a revisionist view of symbol processing within a connectionist framework rather than an eliminativist view is recommended, using localized regions to explain the modularity existing between perceptual and higher cognitive functions. Finally, adoption of an instrumentalist view towards the question of realism with respect to epistemological constructs is recommended in light of the relativity of the competence theories guiding the various functional modules within human cognition