Optimizations of reservoir computing in a nonlinear nanomechanical resonator - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Optimizations of reservoir computing in a nonlinear nanomechanical resonator

Xin Zhou
Connectez-vous pour contacter l'auteur

Résumé

Reservoir computing is a branch of artificial intelligence derived from recurrent neuron network theory. It has the advantage of being efficient in many applications like speech recognition, image classification and time series prediction [1].Recently, studies aimed at using new physical systems for reservoir computing have emerged, and microelectromechanical systems (MEMS) have often been considered [2]. This new kind of reservoir computer consists in creating a network of virtual nodes from a single physical node formed by the MEMS device. The input signal is modulated by a temporal mask, each value of which is equal to the weight corresponding to each virtual node of the reservoir. To add memory, a delayed response of each virtual node is reintroduced into the reservoir. A readout trained weight matrix is then used with the response of each virtual node to obtain the output.In this study, we consider a reservoir computer based on nanoelectromechanical system (NEMS). The device is composed of a circular silicon nitride membrane which is electrostatically actuated [3]. When performing the computation, several parameters are involved in the experimental setup, and our work consists in optimizing them in order to improve the performance of the reservoir computer. We first develop a model of the considered device, which is numerically solved with Matlab ordinary differential equation solver. The parameters, used for modeling, are obtained from fitting the experimental results. Once we have these parameters, we launch simulation of the reservoir computer. To evaluate its performance, the NARMA benchmark has been chosen [4]. It consists in predicting a series of number from a given input. After training the reservoir computer, we see in Figure 1 a comparison between the target and the predicted values. This result shows how effective reservoir computing can be in time series prediction. [1] Tanaka, G., Yamane, T., Héroux, J. B., Nakane, R., Kanazawa, N., Takeda, S., ... & Hirose, A. (2019). Recent advances in physical reservoir computing: A review. Neural Networks, 115, 100-123. [2] Dion, G., Mejaouri, S., & Sylvestre, J. (2018). Reservoir computing with a single delay-coupled non-linear mechanical oscillator. Journal of Applied Physics, 124(15). [3] Zhou, X., Venkatachalam, S., Zhou, R., Xu, H., Pokharel, A., Fefferman, A., ... & Collin, E. (2021). High-q silicon nitride drum resonators strongly coupled to gates. Nano Letters, 21(13), 5738-5744. [4] Atiya, A. F., & Parlos, A. G. (2000). New results on recurrent network training: unifying the algorithms and accelerating convergence. IEEE transactions on neural networks, 11(3), 697-709.
Fichier non déposé

Dates et versions

hal-04270783 , version 1 (05-11-2023)

Identifiants

  • HAL Id : hal-04270783 , version 1

Citer

Toky Harrison Rabenimanana, Xin Zhou. Optimizations of reservoir computing in a nonlinear nanomechanical resonator. GdR MecaQ 8th Annual Meeting 2023, Oct 2023, Lille (France), France. ⟨hal-04270783⟩
67 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More