Two iterative encoding schemes for bidirectional associative memory

Two out-product-type encoding schemes for the bidirectional associative memory (BAM) are introduced which are based on the application of orthogonalization techniques for the stored pattern pairs. With symmetric synaptic connections, one of the schemes overcomes the well-known shortcoming of collapse in capacity which occurrs in most existing symmetric BAMs. The emerged nice saturation property of the scheme can be applied to estimate the storage capacity of Kosko´s scheme (Kosko,1987,1988) numerically rather than statistically. Another encoding scheme yields asymmetric synaptic connections, and assures the complete recall of all linearly independent pattern pairs. Both theoretical analysis and simulations demonstrate that the proposed schemes outperform the existing BAMs in storage capacity, complexity and error-correction capability