A three-stage architecture for bidirectional associative memory

Kosko (1988) proposed bidirectional associative memory (BAM), where pairs of patterns are encoded. When one of the pair is presented, the other is expected to be recalled. Wang et al. (1990, 1991) have shown that BAM cannot correctly recall a pattern pair, if it is not at a local minima of the energy function. In this paper we propose a method to compute exact amount of noise in recalling every encoded pattern. This can be used to find vectors for construction of augmented matrices for both Wang et al., and Wang-Lee (1993) methods. The computation of noise also gives, for the first time, a method for obtaining minimum dimension of augmented vectors. We use the noise information to construct augmented matrix for guaranteed recall of every encoded pattern. The novel encoding method proposed here requires only a few additional elements. We also propose a three layer BAM (TLBAM) that requires smaller size correlation matrices. The TLBAM when encoded using the novel augmentation method guarantees correct recall of all stored pairs. The encoded network also requires much fewer elements than any other method