Statistical Coding and Short-Term Synaptic Plasticity: A Scheme for Knowledge Representation in the Brain

Abstract

This work is a theoretical investigation of some consequences of the hypothesis that transmission efficacies of synapses in the Central Nervous System (CNS) undergo modification on a short timescale. Short-term synaptic plasticity appears to be an almost necessary condition for the existence of activity states in the CNS which are stable for about 1 sec., the timescale of psychological processes. It gives rise to joint "activity-and-connectivity" dynamics. This dynamics selects and stabilizes particular high-order statistical relationships in the timing of neuronal firing; at the same time, it selects and stabilizes particular connectivity patterns. In analogy to statistical mechanics, these stable states, the attractors of the dynamics, can be viewed as the minima of a hamiltonian, or cost function. It is found that these low-cost states, termed synaptic patterns, are topologically organized. Two important properties of synaptic patterns are demonstrated: (i) synap-tic patterns can be "memorized" and later "retrieved", and (ii) synaptic patterns have a tendency to assemble into compound patterns according to simple topo-logical rules. A model of position-invariant and size-invariant pattern recognition based on these two properties is briefly described. It is suggested that the scheme of a synaptic pattern may be more adapted than the classical cell-assembly notion for explaining cognitive abilities such as generalization and categorization, which pertain to the notion of invariance. We are still in almost complete ignorance of how the brain works: there are some very good cues that the Central Nervous System (CNS) is a probabilistic distributed highly non-linear dynamical system-or asynchronous network of "automata"-but we have practically no conceptual tools for studying such knowledge representation devices. The remarkable performances of our cognitive apparatus have received little convincing explanation on the basis of neuronal functioning. One of the most intriguing aspects of cognition is perhaps best described by the notion of an "invariant". Many of the outstanding abilities of the brain, such as categorization and generalization, are directly related to it and could probably be better understood if we had a solid theory of how the brain generates and manipulates invariants. This would require as a first step clarifying the notion of invariance itself. The studies of perception by the Gestalt psychologists at the turn of the century have popularized the idea that NATO ASI Series, Vol. F20 Disordered Systems and Biological Organization Edited by E. Bienenstock et al.