Guidelines to use Transfer Learning for Motor Imagery Detection: an experimental study

Brain-Computer Interfaces (BCI) based on Motor imagery (MI) shown promising results for motor recovery, intraoperative awareness detection or assistive technology control. However, they suffer from several limitations due to the high variability of electroencephalographic (EEG) signals, mainly lengthy and tedious calibration times usually required for each new day of use, and a lack of reliability for all users. Such problems can be addressed, to some extent, using transfer learning algorithms. However, the performance of such algorithms has been very variable so far, and when they can be safely used is still unclear. Therefore, in this article, we study the performance of various state-of-the-art Riemannian transfer learning algorithms on a MI-BCI database (30 users), for various conditions: 1) supervised and unsupervised transfer learning; 2) for various amount of available training EEG data for the target domain; 3) intra-session or inter-session transfer; 4) for both users with good and less good MI-BCI performances. From such experiments, we derive guidelines about when to use which algorithm. Re-centering the target data is effective as soon as a few samples of this target set are taken into account. This is true even for an intra-session transfer learning. Likewise, re-centering is particularly useful for subjects who have difficulty producing stable motor imagery from session to session.