Ultimate "SIR" in Autonomous Linear Networks with Symmetric Weight Matrices, and Its Use to Stabilize the Network - A Hopfield-like network

TL;DR

This paper introduces a novel Hopfield-like nonlinear network with a symmetric weight matrix that stabilizes itself using an 'ultimate SIR' variable, demonstrating improved performance in associative memory tasks over traditional Hopfield networks.

Contribution

It proposes a new stabilization method for Hopfield-like networks using the 'ultimate SIR' concept, applicable in both continuous and discrete-time systems, with proven convergence properties.

Findings

'Ultimate SIR' converges to a system-specific constant.

The network stabilizes and performs better in associative memory tasks.

The approach applies to both continuous and discrete-time systems.

Abstract

In this paper, we present and analyse two Hopfield-like nonlinear networks, in continuous-time and discrete-time respectively. The proposed network is based on an autonomous linear system with a symmetric weight matrix, which is designed to be unstable, and a nonlinear function stabilizing the whole network thanks to a manipulated state variable called``ultimate SIR''. This variable is observed to be equal to the traditional Signal-to-Interference Ratio (SIR) definition in telecommunications engineering. The underlying linear system of the proposed continuous-time network is $\dot{{\mathbf x}} = {\mathbf B} {\mathbf x}$ where {\bf B} is a real symmetric matrix whose diagonal elements are fixed to a constant. The nonlinear function, on the other hand, is based on the defined system variables called ``SIR''s. We also show that the ``SIR''s of all the states converge to a constant value, called ``system-specific Ultimate SIR''; which is equal to $\frac{r}{\lambda_{max}}$ where $r$ is the diagonal element of matrix ${\bf B}$ and $\lambda_{max}$ is the maximum (positive) eigenvalue of diagonally-zero matrix $({\bf B} - r{\bf I})$, where ${\bf I}$ denotes the identity matrix. The same result is obtained in its discrete-time version as well. Computer simulations for binary associative memory design problem show the effectiveness of the proposed network as compared to the traditional Hopfield Networks.