Design and Analysis of a Neuromemristive Reservoir Computing Architecture for Biosignal Processing

Area of Science:

• Neuromorphic Engineering
• Biomedical Signal Processing

Background:

• Reservoir computing (RC) offers advantages over recurrent neural networks (RNNs) in signal processing due to its non-linear computation and simpler training.
• Previous research focused on software-based RCs or hardware implementations using integrated circuits and reconfigurable platforms for improved power and latency.

Purpose of the Study:

• To propose and validate a novel neuromemristive RC architecture for biosignal processing applications.
• To leverage device mismatch for random weight generation and employ mixed-signal subthreshold circuits for energy efficiency.

Main Methods:

• Developed a neuromemristive RC architecture with a doubly twisted toroidal structure.
• Utilized device mismatch for implementing random reservoir weights.
• Designed mixed-signal subthreshold circuits for energy-efficient realizations.
• Compared digital (reconfigurable) and subthreshold mixed-signal implementations.
• Validated the architecture using Electroencephalogram (EEG) and Electromyogram (EMG) biosignal benchmarks.

Main Results:

• The neuromemristive RC architecture achieved 90% accuracy for epileptic seizure detection using EEG data.
• The architecture demonstrated 84% accuracy for Electromyogram (EMG) based prosthetic finger control.
• The study provided a comprehensive analysis comparing digital and mixed-signal subthreshold realizations.

Conclusions:

• The proposed neuromemristive RC architecture is effective for biosignal processing tasks.
• Mixed-signal subthreshold circuits offer energy-efficient implementation of RC.
• The novel architecture shows significant potential for real-world applications in healthcare and prosthetics.