A new algorithm is proposed for improving the learning capability of bidirectional associative memory (BAM) neural networks. The proposed approach, unlike other methods in the literature, is not based on minimizing the energy function of the stored patterns. The proposed technique is the generalization of an auto-associative learning algorithm that has been developed for Hopfield networks. The BAM network is extremely robust to noise, almost guaranteeing perfect recall of all stored patterns with as much as 49% noise. The learning algorithm when applied to a number of test patterns used by other researchers provided satisfying results.
展开▼
