Non-convex multi-species Hopfield models (Q1990104)

The aim of the paper is an investigation on an extension of the standard Hopfield neural network, the multi-species model, having multiple groups of neurons and different Hebbian-like interactions. After a brief review on the standard Hopfield model, the multi-species Hopfield model is introduced. The neurons are allowed to belong to different groups and are characterized by different inter-group and intra-group couplings. The Hamiltonian describing the multi-species Hopfield model is defined and rewritten in terms of the Mattis magnetizations. One introduces the interaction matrix, partition function and pressure of the model. The main result concerns the variational principle for the free energy and the equations for the optimal parameters. Further, the discussion focuses on two particular cases for the multi-species Hopfield model, namely, the bi-directional associative memory model (BAM) and the three-layer restricted Boltzmann machine (RBM).