Role of Synaptic Asymmetry in Neural Network Models of Associative Memory

摘要

The effects of random synaptic asymmetry on the retrieval properties of neural network models of associative memory are investigated analytically and numerically. We consider a Hopfield model in which a random antisymmetric part is added to the otherwise symmetric synaptic matrix, and an asymmetrically diluted version of a model that uses an optimal learning rule based on linear programming. Results are presented for the number of fixed-point attractors, the typical size of the basins of attraction of the memory states and the time of convergence to attractors. The beneficial effects of introducing synaptic asymmetry are discussed.