Rotating neurons for all-analog implementation of cyclic reservoir computing

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Hardware Accelerators

Background:

• Reservoir computing (RC) is a promising paradigm for efficient processing of time-varying data.
• Hardware implementations of RC are crucial for energy-efficient neuromorphic systems.
• Existing hardware approaches primarily focus on the reservoir layer, lacking end-to-end architectures.

Purpose of the Study:

• To propose and validate a versatile method for implementing cyclic reservoirs using rotating elements.
• To demonstrate an end-to-end reservoir computing architecture.
• To explore the potential of hardware physics for high-performance, resource-efficient computation.

Main Methods:

• Mathematical proof of equivalence between the proposed rotating neuron reservoir and standard cyclic reservoir algorithms.
• Simulations for benchmark testing in nonlinear system approximation.
• Development of a hardware prototype integrating rotating elements, signal-driven dynamic neurons, and a memristor-based output layer.

Main Results:

• The rotating neuron reservoir achieved record-low errors in a nonlinear system approximation benchmark.
• The hardware prototype demonstrated high performance in near-sensor computing, chaotic time-series prediction, and handwriting classification.
• The all-analog system achieved 94.0% accuracy in handwriting classification with over 1000x lower system-level power compared to prior works.

Conclusions:

• The proposed rotation-based architecture offers a novel and effective approach for hardware reservoir computing.
• This work demonstrates the successful integration of physical rotating elements as computational resources.
• The developed system showcases significant advancements in energy efficiency and performance for neuromorphic applications.