METASTABLE STATES OF NEURAL NETWORKS INCORPORATING THE PHYSIOLOGICAL DALE HYPOTHESIS

Physiologically it is likely that excitatory and inhibitory neurons are rather clearly distinguished or, in other words, each neuron has in most cases a unique excitatory or inhibitory property (the Dale hypothesis). To study the consequence of the physiological constraint, the author proposes a learning rule for neural networks which incorporates the constraint. Then the distribution of the metastable states is calculated and it is found that the retrieval states form a much larger group in the proposed model than in the Hopfield model. The author also studies the process of retrieval by considering the statistical dynamics of the overlaps. The result suggests that excitatory neurons and inhibitory neurons are preferably balanced in number if the attraction basins of the stored patterns are to be sufficiently large.