Replica symmetry breaking in dense Hebbian neural networks (Q2675357)

The aim of the article is an investigation on the mechanism on Replica Symmetry Breaking (RSB) in dense Hebbian networks, i.e., generalized Hopfield neural networks where neurons interact in P-ples, in high-storage regime (also called Baldi and Venkatesh regime). In the first section, details on dense Hebbian neural networks are provided. One defines the Hamiltonian of the dense Hebbian network (DHN), the control parameters, the partition function related to the Hamiltonian and the quenched statistical pressure of the DHN. In the next sections one provides the mathematical instruments to address the stated problem. Two approaches are considered: on one side the results available in the PDE theory and on the other one a probabilistic approach, the Guerra's interpolation techniques. The second section is devoted to the first approach and develops a methodology based on transport PDE. One obtains a set of self-consistent equations for the order parameters which further allow the construction of phase diagrams for the networks. In the first subsection, under the assumption of replica symmetry (RS) and some additional assumptions, one shows that the maximum storage that network can handle is the Baldi-Vekatesh storage and is achieved for \(a=P-1\). One obtains the quenched statistical pressure of the DHN and further the self-consistency equations. In the second subsection, one concentrates on the solution of the quenched statistical pressure of the dense associative networks under the first step of replica symmetry breaking (1-RSB). The corresponding set of self-consistency equations is shown. The authors mention that the achieved results coincide with already predicted results in earlier papers, in the literature. In the third section, the results achieved by the transport PDE method, are obtained by a suitable generalization of Guerra's interpolation technique, under RS as well as 1-RSB assumptions. In the fourth section, results regarding the zero-temperature self-consistency equation for the Mattis magnetization, under RS as well as 1-RSB assumptions, are presented. The fifth section is devoted to the discussion on Glassiness, to the structure of the Glassiness. One proves a set of representation theorems that allow to decompose Hebbian networks into combinations of pure spin glasses. Conclusions and Outlooks are to be found in the sixth section and some additional proofs in Appendices A--D.