Pluralization of Learning in Physical Reservoir Computing using MEMS Nonlinear Resonator Array
Kosuke SHIMA, Hiroki Takemura, Masaki Shimofuri, Amit BANERJEE, Jun Hirotani, Toshiyuki Tsuchiya · The Proceedings of Mechanical Engineering Congress Japan · 2024
In this report, “pluralization” of learning is demonstrated using an array of MEMS nonlinear resonators as a physical reservoir. The recent spread of IoT devices has led to better efficiency in factory production and more convenience in daily life. On the other hand, there are concerns about rapid increase of data amount processed on the cloud that causes communication delays and large power consumption. To solve this problem, we proposed to implement a new machine learning architecture, physical reservoir computing with a triple MEMS nonlinear resonator array. The array of doubly supported beam resonators is made of single crystal silicon using a silicon-on-insulator wafer. The actuation, detection, and coupling between resonators are conducted electrostatically. In this report, we theoretically modeled the coupled resonators system and solved it numerically using the fourth-order Runge-Kutta method. As a result, we confirmed that the learning performance is improved in both Short-Term Memory and Parity Check tasks when the state matrix used for learning are obtained from all the resonators (pluralization).