Preventing combinatorial explosion in a localist, neural network architecture using temporal synchrony

Abstract

I evaluate the pros and cons of a cognitive modelling approach based strictly on localist representations by examining a sequential decision-making architecture developed to model the neural basis of problem-solving. This architecture, comprising layers of graded-activation processing units interleaved with layers of approximately binary units, seems not only amenable to distributed representations in its graded layers, but actively in need of them in order to support levels of model complexity on the scale of symbolic systems. The architecture employs a key connectionist principle: that the semantics of a representation is defined by what is connected to what. Because of this choice, however, the architecture requires a combinatorially explosive number of localist units as problem complexity increases. I discuss the possibility of preventing combinatorial explosion by binding low-level representations into high-level representations through temporal synchrony, using the same dynamics that underlie decision-making in the architecture.