Variation-tolerant Computing with Memristive Reservoirs

As feature-size scaling in integrated CMOS circuits further slows down, attention is shifting to computing by non-von Neumann and non-Boolean computing models. In addition, emerging devices are expected to behave in time-dependent non-linear ways, beyond a simple switching behavior, and will exhibit extreme physical variation, heterogeneity and unstructuredness. One solution path to address this challenge is to use a dynamical information processing approach that harnesses the intrinsic dynamics of networks of emerging devices. In this paper we employ an approach inspired by reservoir computing, a machine learning technique, to perform computations with memristive device networks that show variation and unstructuredness. Reservoir computing harnesses the nonlinear transient dynamics of such networks and is thus ideally suited for our memristive devices. We, for the first time, apply the reservoir computing approach to a regular structured reservoir and show, on a simple signal classification problem, that this architecture is highly tolerant towards device variation. Furthermore we prove that, compared to unstructured random reservoirs, regular structured reservoirs lead to better average performance as well as to higher variation tolerance. Based on our results of the signal classification task, we argue that harnessing the intrinsic non-linear and time-dependent properties of memristive device networks has the potential to lead generally to more efficient, cheaper, and more robust nanoscale electronics.