A Relevant Prototype Domain Gradient Projection Continual Learning Method for Cross-Subject P300 Brain-Computer Interfaces

Brain-computer interfaces (BCIs) enable communication between the human brain and external devices. However, different factors such as subject specificity lead to excessive individual dependency on conventional BCI systems and deterioration of the usability of the system. Therefore, we propose a relevant prototype domain gradient projection continual learning method to address this challenge. Specifically, we use a prototype domain construction strategy to construct a prototype domain for each subject. Then, the gradient projection method is utilized during continual learning to construct a new prototype domain by selecting previous relevant prototype domains for the current subject. We applied our method to a P300-based cross-subject BCI spelling system and obtained good performance. An average accuracy of 95% was obtained in the offline test of 15 subjects. In addition, the simulated online test for 10 subjects had an average accuracy of 85%, and more than half of the subjects were able to reach 90%. At the same time, the phenomenon of catastrophic forgetting can be avoided in our method. Many experimental results demonstrate the effectiveness of our method.