Deep multimodal emotion recognition using modality-aware attention and proxy-based multimodal loss

Abstract

Emotion recognition based on physiological signals has garnered significant attention across various fields, including affective computing, health, virtual reality, robotics, and content rating. Recent advancements in technology have led to the development of multi-modal bio-sensing systems that enhanced the data collection efficiency by simultaneously recording and tracking multiple bio-signals. However, integrating multiple physiological signals for emotion recognition presents significant challenges due to the fusion of diverse data types. Differences in signal characteristics and noise levels significantly deteriorate the classification performance of a multi-modal system and therefore require effective feature extraction and fusion techniques to combine the most informative features from each modality without causing feature conflict. To this end, this study introduces a novel multi-modal emotion recognition method that addresses these challenges by leveraging electroencephalogram and electrocardiogram data to classify different levels of arousal and valence. The proposed deep multimodal architecture exploits a novel modality-aware attention mechanism to highlight mutually important and emotion-specific features. Additionally, a novel proxy-based multimodal loss function is employed for supervision during training to ensure the constructive contribution of each modality while preserving their unique characteristics. By addressing the critical issues of multi-modal signal fusion and emotion-specific feature extraction, the proposed multimodal architecture learns a constructive and complementary representation of multiple physiological signals and thus significantly improves the performance of emotion recognition systems. Through a series of experiments and visualizations conducted on the AMIGOS dataset, we demonstrate the efficacy of our proposed methodology for emotion classification.