Transfer Learning Approach in Classification of BCI Motor Imagery Signal

The paper presents application of a transfer learning-based, deep neural network classification model to the brain-computer interface EEG data. The model was initially trained on the publicly available dataset of motor imagery EEG data gathered from BCI experienced users. The final fitting was performed on the set of six participants for whom it was the first contact with a BCI system. The results show that initial training affects classification accuracy positively even in case of inexperienced participants. In the presented preliminary study five participants were examined. Data from each participant were analysed separately. Results show that the transfer learning approach allows to improve classification accuracy by even more than 10% points in comparison to the baseline deep neural network models, trained without transfer learning.