A Continuous-time Learning Rule for Memristor–based Recurrent Neural Networks

Abstract

Among the recent disruptive technologies, volatile/nonvolatile memory-resistor (memristor) has attracted the researchers' attention as a fundamental computation element. It has been experimentally shown that memristive elements can emulate synaptic dynamics and are even capable of supporting spike timing dependent plasticity (STDP), an important adaptation rule for competitive Hebbian learning. The overall goal of this work is to provide a novel analogue computing platform based on memristor devices and recurrent neural networks that exploit the memristor device physics to implement the backpropagation algorithm. Back propagation for recurrent neural networks requires a side network for the propagation of error derivatives. The use of memristor-based synaptic weights permit to propagate the error signals in the network via the nonlinear dynamics without the need of a digital side network. Experimental results show that the approach significantly outperforms conventional architectures used for pattern reconstruction. Further results will be reported in an extended work.