Critical neuromorphic computing based on transient dynamics

Abstract

There is a widely accepted hypothesis that the maximum computational ability for a system is achieved in the vicinity of a critical point. In order to construct a computational system working with criticality, we use the coupled oscillators and adjust a coupling strength so that they remain near the stage of a simultaneous cessation of oscillations. Once such a collective transient dynamical system is obtained, its reaction to an input signal can be easily trained to transform to a target output. We show that the model can perform various computing tasks efficiently, especially when the oscillators maintain marginal synchronization at a critical point. The simplicity of the model suggests that complex behaviour can be created based on synchronization of the interconnected simple units, without requiring finely-designed deep structures.