Cellular Neural Networks for Associative Memory with Multiple Memory Tables

Abstract

It is reported that the CNNs (Cellular Neural Networks) have a high ability of associative memory. However, the number of memory patterns and their similarity influence the recall capability of conventional CNNs. Hence, it is considered that the appropriate number of memory patterns which gives the maximum of the recall capability exists. In this paper, we investigate the influence of memory patterns on the recall capability in CNNs, and estimate the range of the appropriate number of memory patterns.Furthermore, we propose the CNN for associative memory with Multiple Memory Tables (MTT-CNN) in order to apply to practical fields. Thereupon, in order to verify the usefulness of the MMT-CNN, we performed the experiments by using 6×6, 12×12, 15×15 MMT-CNN, and examined the incomplete recall rate in each experiment. As a result, the incomplete recall rate of the MTT-CNN was far below that of the conventional CNN, and we could confirm an improvement of self-recall function. Moreover, we could avoid the incomplete recall completely by choosing more appropriate divisions.