Source Data Selection for Brain-Computer Interfaces

A brain-computer interface (BCI) controls a device using brain signals rather than conventional input modalities such as a computer mouse. Before use, the BCI must be calibrated. During the calibration process, the user repeatedly performs mental tasks while electroencephalography (EEG) data is collected. The calibration process is often time-consuming and tiresome for the user. One approach to reduce calibration time is to use transfer learning. With transfer learning, preexisting so-called source data can be used instead of collecting all data during calibration. BCI performance depends on the source data used, and careful source data selection is important for achieving good performance. As more source data becomes available, effective source data selection methods become increasingly important. Another approach to reducing calibration time for motor imagery-based BCIs is early task selection, in which promising motor imagery tasks are identified during calibration so that data is collected only from those tasks.

The first aim of the thesis is to present methods for source data selection. Two methods are presented: the transfer performance predictor method and the pole ratio method. Multi-armed bandits are also investigated for source data selection.
To understand why BCI performance depends on the source data used, the second aim is to provide insight and intuition into the limitations of transfer learning for BCIs. Two Riemannian geometry-based transfer learning methods, the Riemannian Procrustes analysis and the tangent space alignment, are examined to investigate this.
The third and final aim of the thesis is to study how to improve individualized task selection for BCIs. Two methods based on class-mean distance, the direct prediction band and the mean-difference approach, are presented, and multi-armed bandits are investigated for this purpose.

The research presented in this thesis can help reduce BCI calibration time, thereby making BCIs more practical to use. The source data selection methods and insights into Riemannian geometry-based transfer learning could be applied to other fields of research as well.