EEG-Based Emotion Recognition via Knowledge-Integrated Interpretable Method

Abstract

Despite achieving success in many domains, deep learning models remain mostly black boxes, especially in electroencephalogram (EEG)-related tasks. Meanwhile, understanding the reasons behind model predictions is quite crucial in assessing trust and performance promotion in EEG-related tasks. In this work, we explore the use of representative interpretable models to analyze the learning behavior of convolutional neural networks (CNN) in EEG-based emotion recognition. According to the interpretable analysis, we find that similar features captured by our model and state-of-the-art model are consistent with previous brain science findings. Next, we propose a new model by integrating brain science knowledge with the interpretability analysis results in the learning process. Our knowledge-integrated model achieves better recognition accuracy on standard EEG-based recognition datasets.