Step-like dependence of memory function on pulse width in spintronics reservoir computing

Nozomi Akashi,Sumito Tsunegi,Tomohiro Taniguchi,Hitoshi Kubota,Kohei Nakajima,Terufumi Yamaguchi

doi:10.1038/s41598-020-76142-x

Abstract

Physical reservoir computing is a type of recurrent neural network that applies the dynamical response from physical systems to information processing. However, the relation between computation performance and physical parameters/phenomena still remains unclear. This study reports our progress regarding the role of current-dependent magnetic damping in the computational performance of reservoir computing. The current-dependent relaxation dynamics of a magnetic vortex core results in an asymmetric memory function with respect to binary inputs. A fast relaxation caused by a large input leads to a fast fading of the input memory, whereas a slow relaxation by a small input enables the reservoir to keep the input memory for a relatively long time. As a result, a step-like dependence is found for the short-term memory and parity-check capacities on the pulse width of input data, where the capacities remain at 1.5 for a certain range of the pulse width, and drop to 1.0 for a long pulse-width limit. Both analytical and numerical analyses clarify that the step-like behavior can be attributed to the current-dependent relaxation time of the vortex core to a limit-cycle state.

Highlights

Physical reservoir computing is a type of recurrent neural network that applies the dynamical response from physical systems to information processing
Physical reservoir computing is a model of recurrent neural network (RNN) where many-body systems, called reservoirs, are used as the networks[2,3,4,5,6,7,8,9], enabling it to bridge the gap between neural science, information science, biology, and physics
Physical reservoir computing using magnetic vortex-core dynamics in a fine-structured ferromagnet was performed by solving the Thiele equation numerically

Summary

Introduction

Physical reservoir computing is a type of recurrent neural network that applies the dynamical response from physical systems to information processing. 3 45 time (μs) limit, where the range of the pulse width depends on the material parameters and input-current strength.

Results

Conclusion