A BAM with increased information storage capacity

The commonly encountered version of B. Kosko's bidirectional associative memory (BAM) (1988) is generalized to include nonzero thresholds. This modified neural network is referred to as the nonhomogeneous BAM. It is shown that the nonhomogeneous BAM can have any number of stable states between 1 and 2/sup n/, where n is the lowest dimension of either layer of the BAM. Nonhomogeneous BAMs with 1,2, . . ., 2/sup n/-1, 2/sup n/ stable states are constructed. The storage capacity of the BAM is the number of stable states, selected by the user at random from some specified subset of stable states, that can be implemented by the BAM for some matrix selection and some set of thresholds. It is shown that the capacity of the nonhomogeneous BAM can greatly exceed the number of processing elements in the network. This is accomplished without violating information theory, because these stable states utilize a sparse code in which little information is stored in each state.<<ETX>>