Abstract:

Because of the structural limitation of traditional two-layer binary bidirectional associative memory (BAM) network, there are some defects such as limited storage capacity, insufficiency to distinguish small differences in patterns, and lack of capacity to store non-orthogonal patterns. It is a solution that extending it to three-layer network. However, the learning of three-layer binary BAM network is a difficult problem, and three-layer continuous BMA network is inconvenient to deal with binary problems. In order to solve these problems, a three-layer binary BAM network is proposed. The network takes advantage of the MR Ⅱ learning algorithm for multilayer binary feedforward neural networks to perform the learning. The experimental results show that the three-layer binary BAM network based on MR Ⅱ algorithm can improve the storage capability effectively and retain the advantages of binary network, so it has relative high theoretical and practical values.

Key words: three-layer binary bidirectional associative memory network, bidirectional associative memory, pattern storage, binary neural network, MR Ⅱ learning algorithm, minimal disturbance principle