Proceedings of the Annual Meeting of the Cognitive Science Society

Brain-computer interface (BCI) research has advanced significantly, yet cross-subject P300 decoding remains challenging due to highly variable EEG signals. To address this, we propose a novel domain adaptation framework integrating transfer learning and recurrent neural networks (RNNs). The framework incorporates a gated recurrent unit (GRU) layer to effectively capture temporal dependencies of EEG signals, thereby mitigating temporal misalignment issues. By calculating transfer loss and applying time-step weighting, the framework enhances classification. Furthermore, feature-space adaptive alignment is employed to reduce inter-subject variability, lowering subject dependency and enabling accurate cross-subject character recognition. Experimental results demonstrate that the proposed method achieves an average character recognition accuracy of 92.38%, significantly surpassing traditional P300 recognition algorithms. This indicates that the method effectively mitigates the impact of temporal misalignment in EEG data as well as the high subject-dependence of BCI systems.