Brain-computer Interface (BCI) is a rapidly growing area of research. Despite the impressive expansion in the recent years, none of the BCI systems described in the literature are sufficiently mature for the daily use out of the laboratory.
One of the major challenges in BCI is the nonstationary nature of EEG signals. The properties of EEG signals change considerably across sessions and subjects. This challenge has led to the need of calibrating a subject-specific BCI model for each subject at the beginning of each session. This calibration session which takes around 20-30 minutes is very time consuming, and tiring. In this talk, I will discuss a number of novel machine learning algorithms recently developed in our team to deal with the above mentioned challenge. The developed algorithms apply transfer learning techniques in raw EEG, features and classification domains to extract parts of information that are not only discriminative but also in common across different sessions and subjects. These algorithms will be evaluated using motor imagery-based BCI datasets and their advantages and disadvantages will be discussed in detail. Finally, future research directions in this filed will be discussed.