Proceedings of the Annual Meeting of the Cognitive Science Society

Emotion recognition based on electroencephalogram (EEG) signals is a current focus in brain-computer interface research. However, due to the individual differences, how to build a simple and effective model and quickly adapt to the target subject are significant challenges in cross-subject emotion recognition. In this study, we proposed an approach by combining the Dual-Attention network and Meta-Transfer Learning (MTL) strategy based on k-means clustering for meta-task sampling. The Dual-Attention network extracts EEG features through a channel attention block and a temporal attention block. The MTL strategy trains the model to learn both common and individual features among subjects. The meta-task sampling method based on k-means clustering adaptively groups the source domain samples, sampling support and query sets for meta-tasks from Different Groups(DG sampler). The DG sampler allows the model to ”grow in diversity”, further enhancing its generalization capabilities. Binary classification experiments were conducted on the DEAP dataset, achieving accuracies of 72.35% and 71.77% in the arousal and valence dimensions, respectively. The results have reached the state-of-the art level and demonstrated significant performance enhancement in cross-subject EEG-based emotion recognition.