A CSP-based retraining framework for motor imagery based brain-computer interfaces

CSP is one of the most widely used signal processing approaches in EEG-based MI classification; however, the CSP optimization objective is not completely consistent with the final classification objective, and hence it does not necessarily lead to the best classification performance. This study has proposed a retraining framework, which retrains a neural network with the same forward computational process and initial parameters as the CSP-based traditional model, and further optimizes it on the labeled training data using gradient descent. Experiments on four MI datasets demonstrated that retraining improved traditional models' classification performance and outperformed several popular deep neural network models, especially when the amount of labeled training data was very small. Our work demonstrates the advantage of integrating knowledge from traditional models and from the training data in EEG-based BCIs.