Versatile and Robust Reservoir Computing with PWM-Driven Heterogenous R-C Circuits.

Area of Science:

• Analog Computing
• Neuromorphic Engineering
• Materials Science

Background:

• Physical reservoir computing (PRC) offers low-latency, energy-efficient processing but faces challenges in flexibility, adaptability, and environmental stability.
• Existing PRC systems often rely on specialized materials like memristors, which can be sensitive to environmental changes and manufacturing variations.

Purpose of the Study:

• To develop a highly versatile and robust PRC system overcoming limitations of current implementations.
• To demonstrate state-of-the-art performance in complex computational tasks using a novel circuit design.

Main Methods:

• Implementation of a PRC system utilizing pulse-width modulation (PWM)-encoded resistor-capacitor (R-C) circuits.
• Leveraging customizable nonlinearities and dynamic timescales within the R-C network.
• Testing performance on chaotic time-series forecasting (Mackey-Glass) and multiclass heartbeat classification.

Main Results:

• Achieved state-of-the-art NRMSE of 0.015 for Mackey-Glass chaotic time-series forecasting.
• Attained 94% accuracy for complex multiclass heartbeat classification.
• Demonstrated a significant reduction (98.4%) in relative errors across device batches and temperature variations compared to memristor-based reservoirs.

Conclusions:

• The PWM-encoded R-C circuit PRC system exhibits exceptional versatility and robustness.
• This approach presents a scalable, adaptive, and energy-efficient solution for edge computing in dynamic environments.
• The developed system paves the way for practical and stable analog computing applications.