STUDY OF DIFFERENT" STRATEGIES FOR RECALL OF ALL TRAINING PAIRS' IN BIDIRECTIONAL ASSOCIATIVE MEMORY

Abstract

:Ttie bidirectional associative memory (BAM) is heteroassociative ; that is , it accepts an input vector on one set of neurons . and 'produces a related, , but different output vector on another. set. BAM is capable of generalization, producing-correct outputs despite corrupted inputs. -BAM encoding (the process of training the BAM to recognize. a set of N vector pairs) was first introduced by • Kosko. But Kosko's encoding method does not ensure the recall of any stored pair. Multiple training encoding strategy is the enhancement of Kosko's method and ensures recall of a particular trained pair and in some cases .a number of pairs. - In this dissertation, study . of necessary and sufficient conditions is done so that recall of all training pairs is guaranteed. Here two methods, to find the weights of generalised correlation matrix of bidirectional associative memory are presented. One is sequential multiple training (SMT) method which yields integers for the weights where the weights are multiplicites of training of the tranining pairs. Another Is linear programming/ -multiple training method -which determine weights that satisfy the conditions when a solution is feasible. Computer simulations for the two methods have been done .Observation has been made that it is possible to recall all the training pairs using themselves as the initial condition provided the solution exists. Recall of all traning pairs has been applied to recognition of different patterns