Feasibility and advantage of reservoir computing on single-electron devices

Abstract

In this paper, applying reservoir computing (RC) to a single-electron (SE) device/circuit is discussed. RC, as one of the recurrent neural networks (NNs), consists of the input, reservoir, and output layers. It is known that NNs have redundancy for fluctuation. In addition, in RC, the weight functions of some of the connections are not necessarily updated, although it is needed for conventional NNs. Moreover, the number of connections in the reservoir layer can be reduced. Generally, nano-electronic devices, including SE devices, suffer from fluctuation and the large number of connecting paths between elements. By applying RC to the devices, these problems are expected to be overcome. In this study, the SE-RC circuit was designed and its operation was simulated by Monte Carlo simulation. The results indicate that the approach of applying RC to the SE circuits is very effective and can be a candidate for circuit architecture.