Hysteretic reservoir

Physical-reservoir computing (PRC) is an information-processing framework where the nonlinear physics of a reservoir can be leveraged to compute a task. Since the readout consists of linear and static weights, most of the computation, requiring nonlinear transformation and memory of the inputs, is done by the natural dynamics of the reservoir. Furthermore, training those readout weights using linear regression can be computationally fast and lightweight to implement. In this study, we propose a method to utilize the hysteresis effect, traditionally considered a hindrance in computation, to generate nonlinearity and memory in PRC. Specifically, we developed a model that emulates hysteresis properties by introducing latency in the reservoir states, and we demonstrate its effectiveness through dynamical analysis. This approach allows us to relax the requirement for a recurrent weight matrix as implemented in traditional echo state networks. Furthermore, we validate the practical applicability of this hysteretic reservoir framework with a hardware implementation using a simple coil transformer. Our findings indicate that materials exhibiting hysteresis effects can be harnessed to unlock a new class of computational materials, heralding a novel framework for PRC.