Improving the Accuracy of Spiking Neural Networks for Radar Gesture Recognition Through Preprocessing

Area of Science:

• Artificial Intelligence
• Edge Computing
• Neuromorphic Engineering

Background:

• Event-based neural networks (SNNs) offer efficient AI solutions for edge devices.
• Optimizing SNNs requires investigating preprocessing techniques alongside network architecture.
• Radar gesture recognition is a key application for low-power embedded systems.

Purpose of the Study:

• To demonstrate a highly accurate and energy-efficient spiking neural network (SNN) for radar gesture recognition.
• To investigate the critical role of signal preprocessing in SNN performance compared to deep neural networks (DNNs).
• To highlight the unique impact of preprocessing on SNNs versus DNNs, challenging conventional design assumptions.

Main Methods:

• Developed a 4-bit weight spiking neural network (SNN) with a simple architecture (one convolutional, two fully connected layers).
• Implemented and evaluated the SNN for radar gesture recognition, comparing its performance against a DNN with identical topology and training.
• Analyzed the influence of signal preprocessing parameters on both SNN and DNN accuracy.

Main Results:

• Achieved state-of-the-art 93% accuracy in radar gesture recognition using the SNN within only four processing time steps.
• Demonstrated that efficient signal preprocessing is significantly more critical for SNNs than for DNNs.
• Showcased antagonistic effects of preprocessing parameters on SNNs and DNNs, indicating distinct design requirements.

Conclusions:

• The developed SNN enables ultralow-power radar-based gesture recognition suitable for extreme-edge devices.
• Signal preprocessing is a vital factor for maximizing SNN performance, often more so than for DNNs.
• Findings necessitate tailored preprocessing strategies for SNNs, as DNN-derived conclusions may not generalize.