Robust Implementation of Memristive Reservoir Computing with Crossbar Based Readout Layer

Abstract

Reservoir computing (RC) is a neural network paradigm that aims to significantly reduce the training cost of recurrent neural networks. Physical reservoir computing is the hardware realization of this system that aims to achieve faster data processing with low training costs. This paper presents a robust memristor-based liquid state machine implementation of reservoir computing. A generalized hardware framework for this system's implementation is proposed that can be used to realize any given network topology. Also, a readout layer based on simple spike-timing-dependent plasticity (STDP) based online learning is proposed for the system. Such an implementation will lead to reduced training cost for the system. Additionally, this readout layer is shown to be robust against device level switching asymmetries that are prevalent in contemporary memristor devices.