Knowledge-guided quantization-aware training for EEG-based emotion recognition

Abstract

Emotion recognition is of paramount importance in various domains. In recent years, the use of models that employ electroencephalogram data as input has seen substantial achievements. However, the increasing complexity of these EEG models presents substantial challenges that hinder their deployment in resource-limited environments. This situation emphasizes the critical need for effective model compression. However, extreme compression often leads to significant degradation in model performance. To address this issue, we propose a novel Knowledge-Guided Quantization-Aware Training method for EEG-based emotion recognition task. This method integrates knowledge from emotional neuroscience into the quantization process, emphasizing the importance of the prefrontal cortex part in the EEG sample selection process to construct the calibration set and successfully enhance the performance of Quantization-Aware Training techniques. Experimental results demonstrate that our proposed framework achieves quantization to 8 bits, which leads to surpassing SOTAs in EEG-based emotion recognition. The source code is made available at: https://github.com/Stewen24/KGCC .