Generalized hetero-associative neural networks

Abstract

While auto-associative neural networks (e.g., the Hopfield model implementing the standard Hebbian prescription for learning) play as the reference setting for pattern recognition and associative memory in statistical mechanics, hetero-associative extensions (despite much less investigated) display richer emergent computational skills. Here we study the simplest generalization of the Kosko's Bidirectional Associative Memory (BAM), namely a Three-directional Associative Memory (TAM), that is a tripartite neural network equipped with generalized Hebbian weights. We study its information processing capabilities analytically (via statistical mechanics and signal-to-noise techniques) and computationally (via Monte Carlo simulations). Confined to the replica symmetric description, we provide phase diagrams for this network in the space of the control parameters, highlighting the existence of a region where the machine can successful perform recognition as well as other tasks. For instance, it can perform pattern disentanglement, namely when inputted with a mixture of patterns, the network is able to return the original patterns, namely to disentangle the signal's components. Further, they can also perform retrieval of (Markovian) sequences of patterns and they can also disentangle mixtures of periodic patterns: should these mixtures be sequences that combine patterns alternating at different frequencies, these hetero-associative networks can perform generalized frequency modulation by using the slowly variable sequence of patterns as the base-band signal and the fast one as the information carrier.