Intra- and Inter-Subject Transfer Learning for Non-InvasiveBrain-Computer Interface
Alexandre Bleuzé · theses.fr (ABES) · 2023
A brain-computer interface (BCI) is a direct link between a brain and a computer, enabling an individual to perform tasks without the need for peripheral nerves or muscles. In recent years, BCIs have become increasingly interesting, especially in the healthcare sector, because of their potential to help patients. They have been used to help some people recover their motor functions after a stroke or spinal cord injury, or to help people with degenerative diseases such as amyotrophic lateral sclerosis, who gradually lose the ability to control their limbs and then communicate. Another factor adding to the appeal of BCIs is their potential to enhance the capabilities of healthy people in areas such as video games. Today, thanks to technological advances in the healthcare field, the tools needed to set up BCIs, such as electroencephalograms, are becoming more affordable, enabling the multiplication of experiments and clinical tests, giving access to a vast amount of data, sometimes freely available on the Internet. This data could make it possible to create models that have been trained from the data of many people, thereby increasing the performance of future systems while reducing their calibration time. This would also enable the use of less expensive hardware for equivalent performance, making BCIs more affordable. The main problem today is that the data available in open access is very heterogeneous, whether in terms of quality, paradigm or even simply hardware. For these reasons, it is very difficult to exploit all this data to extract common features. The aim of this thesis is to find methods for adapting and using open-access data to create machine learning models that are highly robust because they are trained on data from a wide range of subjects. To this end, we are focusing on Riemannian geometry, the use of which in brain-computer interfaces has recently shown to be highly effective. More specifically, this original work focuses on the development of transfer learning methods in the tangent space of the Riemannian variety. The proposed methods have been evaluated on a large number of databases covering several paradigms: motor imagery, P300 evoked potentials and steady-state visual evoked potentials. The work carried out in this thesis has led to the development of a method called Tangent Space Alignment (TSA), which achieves an overall improvement in accuracy of 2.7% over a previously published Riemannian method, Riemannian Procrustes Analysis (RPA). Another contribution of this thesis to the scientific community is research into the use of mathematical arbitrary sources in BCI transfer learning. The work carried out in this thesis shows that little information is lost when aligning to this arbitrary source, and studies the impact on accuracy between subjects, enabling new alignment possibilities to be explored and mathematically normalized alignment sources to be sought, rather than existing subject data which may not possess the right mathematical properties to serve as a quality source.