Abstract
One of the serious problems of neural networks is that the number of weights increases explosively as size of the network becomes larger. A lot of studies have been done for the multilayer perceptron type networks. However, there is no such a study for neural network associative memories. This paper proposes two effective algorithms to reduce the number of weights in BAM (Bidirectional Associative Memory): one is a simple algorithm which does not require too much additional computations; the other is designed to reduce the weights as many as possible. Both of the algorithms can guarantee the recall of the stored data. Computer simulations are carried out to confirm the validity of the proposed algorithms and to examine the performance.
