BASIN OF ATTRACTION IN NETWORKS OF MULTISTATE NEURONS

Abstract

The retrieval dynamics of an extremely diluted network of three-state neurons designed by the projection rule is investigated. Depending on the parameters of the model four different fixed points with different stability can exist. Only one of them corresponds to full retrieval. It is discussed how the parameters of the model must be chosen in order to get perfect retrieval and large basins of attraction. Finally it is shown that a network of three-state neurons storing binary patterns has better retrieval properties than a network of binary neurons storing the same patterns.