EEG Transfer Learning Algorithm Based on Weighted Multi-Source Domain Adaptation

• 1、School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing, 100876

Abstract：In recent years, electroencephalogram (EEG) transfer learning algorithms have demonstrated significant potential in the research of Brain computer interfaces (BCI). EEG signals exhibit variability between individuals due to differences in psychological states and cognitive levels. This variation in data distribution leads to a significant drop in performance when models trained on EEG data from a single individual are applied to others. Current EEG transfer algorithms face challenges in effectively capturing semantic correlations across different data domains, indicating significant potential to improve the efficiency of transferring semantic knowledge. This paper proposes a weighted adaptive multi-source domain EEG transfer learning algorithm to address these challenges. The proposed algorithm aims to narrow the distance and align the decision boundaries between different domains to eliminate the effects of domain-shift. It assigns different transfer weights to source domains based on the normalized mutual information between the source and target domains, indicating the significance of samples. An attention module is employed to effectively capture semantic information correlations across different EEG channels, enabling the model to deeply learn the relationships between domains and thoroughly align source and target domain data, thus achieving EEG transfer from source to target domains. Experimental validation on datasets demonstrates that the proposed algorithm significantly enhances the performance of target domain data on source domain models, proving the algorithm's effectiveness.

Keywords： Deep Learning Brain Computer Interfaces Transfer Learning Domain Adaptation