Several associative memory models, such as Hopfield Associative Memory (HAM) by Hopfield, Bidirectional Associative Memory (BAM) by Kosko and Multidirectional Associative Memory (MAM) by Hagiwara, have been proposed. We have proposed the Multidirectional Associative Memory with a hidden layer. There have been many researches for the storage capacity of HAM. We denote the number of neurons as N. For correlation learning, 0.15N was derived by Hopfield, 0.14N by Amit and and 0.16N by Amari. Moreover, it is known that the storage capacity by the pseudo-inverse matrix is N and that by error-correcting is 2N. Therefore the storage capacity of HAM is proportional to N. It is known that the storage capacity of BAM depends on the number of neurons of the layer which has less neurons than the other. Let M be the number of neurons in a MAM except the layer which has most neurons. Then we can expect that the storage capacity of the MAM is proportinal to M owing to the similarity between correlation learning of the HAM and that of the MAM. We investigate this conjecture by computer simulation for correlation learning, error-correcting learning, pseudo-relaxation learning and Multidirectional Associarive Memory with a hidden layer.
展开▼
