Neural Architecture Search of Time-to-First-Spike-Coded Spiking Neural Networks for Efficient Eye-based Emotion Recognition

A research team from Zhejiang University and the National University of Singapore has introduced TNAS-ER, a novel framework that automates the design of ultra-efficient AI models for emotion recognition via smart glasses. The system specifically targets Time-to-First-Spike (TTFS) Spiking Neural Networks (SNNs), a brain-inspired computing paradigm where neurons fire at most one binary spike, leading to extremely sparse and energy-efficient computation. TNAS-ER employs a neural architecture search (NAS) to find the optimal network structure, a critical factor often overlooked, as spike timing in TTFS networks is tightly linked to architectural design.

TNAS-ER's key innovation is an ANN-assisted search strategy. It uses a standard ReLU-based Artificial Neural Network (ANN) as a guide to stabilize the training and optimization of the more complex TTFS SNN. The framework uses an evolutionary algorithm, optimizing for both recognition accuracy (measured by weighted/unweighted average recall) and efficiency. This co-design approach results in compact, high-performance networks tailored for the constrained compute and power budgets of wearable eyewear.

Extensive experiments validate the framework's effectiveness. The TNAS-ER-designed networks achieve strong emotion recognition performance while demonstrating a massive reduction in computational cost compared to conventional deep learning models. The researchers further validated the models on neuromorphic hardware, confirming their superior energy efficiency and strong potential for real-world, always-on applications in next-generation emotion-aware wearable devices.