Multilayer feed forward neural networks for non-linear continuous bidirectional associative memory

Abstract

The study of Bidirectional associative memory (BAM), with recurrent neural networks and symmetric as well as asymmetric weights, has already been undertaken in various different ways. Using two phases of learning for multilayer neural network architecture in the present paper, a multilayer feed forward neural network model has been proposed to construct the non-linear continuous BAM for pattern association. In the first phase an input pattern is presented to input layer and back propagation learning rule is used to train the network in the feed forward direction for the corresponding associated output pattern. In second phase the output pattern is presented to output layer as input and again the back propagation learning rule is used to train the same network in feedback direction for the corresponding associated input pattern. In these two passes i.e. forward pass and backward pass, the interconnection weights are considered asymmetric. This training process continues till the network does not converge to the final optimal weights by minimizing the mean square errors in both the directions simultaneously. At this convergence of weights the input and output layers exhibit the stability and the performance of such type of BAM is evaluated for the test pattern set while the simulation results exhibit the better performance of associative memory for the proposed method. The storage capacity of network is also increased due to the non-linear mapping between input-output pattern pairs and also the occurrence of spurious states reduces during the recalling process.