A review of memristive reservoir computing for temporal data processing and sensing

Abstract

Reservoir computing (RC) is a promising paradigm for machine learning that uses a fixed, randomly generated network, known as the reservoir, to process input data. A memristor with fading memory and nonlinearity characteristics was adopted as a physical reservoir to implement the hardware RC system. This article reviews the device requirements for effective memristive reservoir implementation and methods for obtaining higher-dimensional reservoirs for improving RC system performance. In addition, recent in-sensor RC system studies, which use a memristor that the resistance is changed by an optical signal to realize an energy-efficient machine vision, are discussed. Finally, the limitations that the memristive and in-sensor RC systems encounter when attempting to improve performance further are discussed, and future directions that may overcome these challenges are suggested.