Z(2) Gauge Neural Network and its Phase Structure

TL;DR

This paper explores the phase structures of neural network models with Z(2) gauge symmetry, analyzing how reverberation effects influence different models through numerical simulations.

Contribution

It introduces and compares three Z(2) gauge neural network models, revealing the separation of Higgs and spin-glass phases due to reverberation effects.

Findings

Higgs phase splits into two stable spin-glass phases in quenched models.

Reverberation term significantly affects phase stability.

Phase structures depend on connectivity and system size.

Abstract

We study general phase structures of neural-network models that have Z(2) local gauge symmetry. The Z(2) spin variable Si = \pm1 on the i-th site describes a neuron state as in the Hopfield model, and the Z(2) gauge variable Jij = \pm1 describes a state of the synaptic connection between j-th and i-th neurons. The gauge symmetry allows for a self-coupling energy among Jij's such as JijJjkJki, which describes reverberation of signals. Explicitly, we consider the three models; (I) annealed model with full and partial connections of Jij, (II) quenched model with full connections where Jij is treated as a slow quenched variable, and (III) quenched three-dimensional lattice model with the nearest-neighbor connections. By numerical simulations, we examine their phase structures paying attention to the effect of reverberation term, and compare them each other and with the annealed 3D lattice model which has been studied beforehand. By noting the dependence of thermodynamic quantities upon the total number of sites and the connectivity among sites, we obtain a coherent interpretation to understand these results. Among other things, we find that the Higgs phase of the annealed model is separated into two stable spin-glass phases in the quenched cases (II) and (III).