Multimodal Knowledge Distillation for Emotion Recognition.

Multimodal emotion recognition has emerged as a prominent field in affective computing, offering superior performance compared to single-modality methods. Among various physiological signals, EEG signals and EOG data are highly valued for their complementary strengths in emotion recognition. However, the practical application of EEG-based approaches is often hindered by high costs and operational complexity, making EOG a more feasible alternative in real-world scenarios. To address this limitation, this study introduces a novel framework for multimodal knowledge distillation, designed to improve the practicality of emotion decoding while maintaining high accuracy, with the framework including a multimodal fusion module to extract and integrate interactive and heterogeneous features, and a unimodal student model structurally aligned with the multimodal teacher model for better knowledge alignment. The framework combines EEG and EOG signals into a unified model and distills the fused multimodal features into a simplified EOG-only model. To facilitate efficient knowledge transfer, the approach incorporates a dynamic feedback mechanism that adjusts the guidance provided by the multimodal model to the unimodal model during the distillation process based on performance metrics. The proposed method was comprehensively evaluated on two datasets based on EEG and EOG signals. The accuracy of the valence and arousal of the proposed model in the DEAP dataset are 70.38% and 60.41%, respectively. The accuracy of valence and arousal in the BJTU-Emotion dataset are 61.31% and 60.31%, respectively. The proposed method achieves state-of-the-art classification performance compared to the baseline method, with statistically significant improvements confirmed by paired