| dc.description.abstract | This thesis presents a transfer learning method based on Riemannian
geometry for improving the classification accuracy of Electroencephalographic
(EEG) signals. Non-stationarities are ubiquitous in EEG signals, which means the
statistical characteristics of EEG signals alter from time to time. The nonstationarities
of the EEG signals may be caused by different environmental factors
(e.g. user’s fatigue level, the mental and physical state of user, the location of
electrodes placement, etc.). Typically, classic Motor Imagery-based Brain-
Computer Interface (MI-based BCI) requires a calibration session in each run,
even for recorded subjects. During the calibration session, the subjects requested
to perform various Motor Imagery (MI) tasks repeatedly, which will be timeconsuming
and make user feel exhausted. As a consequence, we proposed a
transfer learning method, namely Common Spatial Pattern Riemannian Procrustes
Analysis (CSP-RPA), to shorten the calibration time while keeping MI-based BCI
work optimally. CSP-RPA is based on the tangent space of Riemannian geometric
spaces and combines the Common Spatial Pattern (CSP) method to modify the
Riemannian Procrustes Analysis (RPA) architecture. To alleviate the overfitting
in high-dimensional Riemannian manifold, the tree-based feature selection is
adopted to reduce the dimensionality after mapping data from Riemannian
manifold to tangent space. The framework was validated by the publicly available
EEG dataset 2a of the BCI competition IV. In addition, we used t-SNE (t-
Distributed Stochastic Neighbor Embedding) to visualize and prove the
effectiveness of feature domain adaptation after CSP-RPA algorithm. To sum up,
the experimental results indicate that CSP-RPA is superior to other methods, e.g.,
Re-center, Parallel Transport, RPA, under cross-sessions and cross-subjects
conditions. | en_US |