Transfer Learning for Motor Imagery based Brain Computer Interfaces

Current electroencephalogram (EEG) based brain-computer interface (BCI) systems have limited real-world practicality due to a number of issues, including the long calibration period required before each use. This thesis focuses on reducing the time required to calibrate the BCI system without sacrificing classification accuracy. To address this issue, previously collected EEG data could be potentially mined and reused in calibrating the BCI model for a new user/session. However, this is not a trivial task due to two key challenges. First, there are considerable non-stationarities between the current and previously collected EEG signals. Secondly, due to between-session variations, not all the previously collected EEG signals are helpful in training the new BCI model.

Initially, the thesis explored the application of distribution alignment techniques to reduce the effects of EEG non-stationarity. A novel multiclass data space alignment (MDSA) algorithm was proposed and evaluated. Our results showed that the proposed MDSA alignment algorithm successfully improved the classification accuracy and reduced the effects of non-stationarity.

The thesis then addressed the second challenge by developing a new framework. This framework utilised a new algorithm that identifies whether or not the new session would benefit from transfer learning. If so, a novel similarity measurement, called the Jensen-Shannon Ratio (JSR), was proposed to select one of the past session for training the BCI model. The proposed framework outperformed state-of-the-art algorithms when there were as few as five labelled trials per class available from the new session. Despite success to some extent the proposed framework was limited to a binary selection between only one of the past sessions and current data for training the BCI model.

Finally, the thesis utilised the findings of the previous research in order to address both challenges. A novel transfer learning framework was proposed for long-term BCI users. The proposed framework utilised regularisation, alignment and weighting to train a BCI which outperformed state-of-the-art algorithms even when only two trials per class from the new session were available.