Multi-Modeshape Reservoir Computing Using a Continuous MEMS Microbeam
Abstract Delay-based Reservoir computing (RC) offers great potential in time-series problems, especially when applied in hardware due to its low computational power and its compact nature. However, this approach suffers from a large computational delay because of the serial probing of virtual nodes. To address this disadvantage, this paper presents the use of a continuous MEMS arch for Delay-based RC. This novel approach reduces the computational delay by using fewer virtual nodes through maintaining sufficient virtual node coupling and nonlinear complexity. As a demonstration, we show that a single MEMS arch is capable of performing a binary waveform classification task of a multi-frequency square-and-triangle waveform problem with a success rate > 96% using only 10 virtual nodes compared to 40 virtual nodes in a typical implementation. The reduction in the number of virtual neurons is achieved by biasing the MEMS device using an AC source around its second modeshape.