A Demonstration of Analog/Digital Converter-Free Reservoir Computing Using a Series-Connected Resonating Switched MEMS Counter

This paper proposes a reservoir computing system that utilizes the dynamics of series-connected resonators, toward the implementation of energy-efficient machine-learning information processing. The oscillator's mechanical characteristics dynamically change in response to external signal input as a substrate bias voltage, and a contact switch equipped in each resonator detects a large displacement exceeding the initial gap. The system simply counts the contact in a timestep at each resonator as the reservoir output. Using the designed device, reservoir computing with two types of benchmark tasks was experimentally demonstrated, clearly showing that the combination of our device and readout method effectively extracts time-series information useful for computing. The small footprint and simple readout without the need for analog peripheral circuits contribute to practical edge computing.