Neural and Conceptual Interpretations of Parallel Distributed Processing Models

Abstract : Mind and brain provide two quite different perspectives for viewing cognition. Yet both perspectives are informed by the study of parallel distributed processing. This duality creates a certain ambiguity about the interpretation of a particular PDP model of a cognitive process: Is each processing unit to be interpreted as a neuron? Is the model supposed to relate to the neural implementation of the process in some less direct way? A closely related set of questions arises when it is observed that PDP models of cognitive processing divide broadly into two classes. In local models, the activity of a single unit represents the degree of participation in the processing of a known conceptual entity a word, a word sense, a phoneme, a motor program. In distributed models, the strength of patterns of activity over many units determine the degree of participation of these conceptual entities. In some models, these patterns are chosen in a deliberately arbitrary way, so that the activity of a single unit has no apparent "meaning" whatever -- no discernible relation to the conceptual entities involved in the cognitive process. On the surface, at least, these two types of models seem quite different. Are they as different as they seem? How are they related? This chapter begins with a brief consideration of the neural interpretation of PDP models of cognition. These considerations serve mostly to lay out a certain perspective on the PDP modeling world, to make some distinctions I have found to be valuable, to introduce some terminology and to lead into the main question of this chapter: How are distributed and local PDP models related? The chapter ends with a discussion of how, using the framework of PDP models, we might forge a mathematical relationship between the principles of mind and brain.