Reservoir computing model using a single nonlinear nanoelectromechanical resonator at atmospheric conditions

Abstract

Reservoir computing is an alternative method to conventional systems using computationally expensive recurrent neural networks (RNNs). In this method, the training is performed only at the final layer of a nonlinear physical system functioning as a black box substituted instead of the hidden layers in RNNs requiring intensive training. This study suggests using a small nanoelectromechanical systems (NEMS) resonator with intrinsic nonlinearities for reservoir computing instead of relying on complicated feedback loops or spatially extended reservoirs as used in the earlier works. The linear classification is made possible by trans-forming the input data into a higher dimensional space, which is accomplished by utilizing the combination of the nonlinearity of the NEMS resonator and its fading memory behavior stemming from its transient response. Compared to reservoir computing using micromechanical resonators, the use of nanoelectromechanical resonators results in faster information processing, enabled by their rapid decay times arising from their small dimensions. Moreover, the implementation of the proposed NEMS reservoir computing architecture is more practical since it can operate at atmospheric conditions and occupies less space than its MEMS counterparts. This study emphasizes the efficient and feasible information processing potential of the suggested approach for a range of applications by the evaluation of its performance with the MNIST handwritten digit recognition task.