Regularization Continual Learning Based on Bayesian Uncertainty Modeling for P300 Brain-Computer Interface

The P300 is a specific component of the event-related potentials (ERPs) and has been extensively utilized in brain-computer interface (BCI) applications. Numerous neural network models have been implemented to detect P300 and achieve outstanding results in intra-subject scenarios. However, in real-world situations where data streams from different subjects arrive sequentially, intra-subject models are time-consuming and resource-intensive. This necessitates research into cross-subject scenarios. To address this issue, we propose a continual learning (CL) method, named Elastic Weight Consolidation with Bayesian Neural Network (EWC-BCNN), for cross-subject P300 decoding. Specifically, EWC-BCNN comprises two main modules: EWC and BCNN. EWC employs a regularization term to penalize significant changes in parameters deemed important. Additionally, BCNN quantifies parameter uncertainty, identifying and protecting crucial knowledge. By integrating BCNN, EWC can make more informed decisions about which weights to preserve and to what extent, thus enhancing its ability to balance the retention of past knowledge with the adaptation to new information. We evaluated our method on two public P300 datasets. Our experimental results demonstrate that EWC-BCNN achieves better P300 detection performance than point-estimate networks. Furthermore, EWC-BCNN outperforms other state-of-the-art CL methods.