Abstract

A two-layer high-content addressable memory, based on the K-nearest neighbor and the winner take all networks, as applied to autoassociation and the heteroassociation models, is given. The architecture is an efficient neural network in which pattern classification can be obtained very easily. In this two-layer model, the total number of interconnections is N_i^* M + N_o^* M, where N_i and N_o are the input neurons and the output neurons respectively, and M is the number of exemplars, which is less than the conventional CAMs. Since the storage capacity can be as high as 2^Ni, the neural network needs only one feedforward training time, such that no feedback iteration is required. Fast convergence is therefore an apparent feature, and the noise performance is better than that of the Hopfield model and the inter-pattern association (IPA) model. Since this network would converge to the nearest exemplar, it would not produce spurious outputs. Computer simulation and experimental results confirm our findings and are provided.

© 1992 Optical Society of America