Application of a Multi-Source Multi-Task Weight Adaptation Framework for Cross-Domain EEG Emotion Recognition(MS-MWA)

Abstract

In the practical implementation of the Electroencephalography (EEG) emotion recognition system, the development of a universal model applicable to different subjects and sessions is crucial. Previous studies have primarily focused on employing domain adaptation methods to address the differences in EEG distributions between different subjects and sessions.However, traditional approaches overly rely on source domain data, resulting in sensitivity to changes and limited generalization and adaptability. The introduction of multi-task domain adaptation, through diverse training, enhances generalization, making the model more flexible and adaptable to the complex and dynamic nature of real-world environments. Therefore, we propose a Multi-Source Multi-Task Weight Adaptation Framework (MS-MWA). To enhance model generalization, we introduce multiple tasks at different levels, enabling the model to learn hierarchical feature representations. At shallow layers, the model learns domain-invariant features, while at deep layers, it learns domain-specific features, thereby improving the model’s understanding of complex abstractions.To address the challenge of difficulty in adjusting relative weights among various task losses, we propose a multi-task deep learning approach that balances different loss functions using homoscedastic uncertainty between tasks. This method not only reduces the need for continual adjustment of weights for optimal performance but also enhances overall model performance.We evaluated our proposed model on the publicly available datasets SEED and SEED-IV. In cross-subject experiments, our model achieved accuracy rates of 91.35% and 72.20% on SEED and SEED-IV, respectively. In cross-session experiments, our model demonstrated accuracy rates of 92.27% on SEED and 77.47% on SEED-IV.