Riemannian Procrustes Analysis: Transfer Learning for Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalographic (EEG) signals exhibit statistical variability across sessions and subjects.
• This variability poses a significant challenge for Brain-Computer Interface (BCI) systems, hindering data reuse and necessitating lengthy calibration.
• Existing transfer learning methods struggle to effectively account for this inherent data variability.

Purpose of the Study:

• To develop a novel transfer learning approach to mitigate statistical variability in EEG signals.
• To enable BCI systems to effectively reuse data from previous recordings, thereby reducing or eliminating the need for calibration.
• To improve the robustness and generalizability of BCI models across different users and sessions.

Main Methods:

• Proposed Riemannian Procrustes Analysis (RPA), a geometry-aware method utilizing Procrustes analysis.
• Employed symmetric positive definite (SPD) matrices as statistical features for EEG signals.
• Applied geometrical transformations (translation, scaling, rotation) on the SPD manifold to match data distributions.
• Validated the method on simulated data and eight public BCI datasets (243 subjects).

Main Results:

• RPA demonstrated superior classification accuracy compared to existing geometry-aware methods.
• Significant improvements were observed in transfer learning performance across diverse BCI datasets and experimental paradigms.
• Ensemble classification strategies benefited from dataset statistics matching via RPA.

Conclusions:

• RPA offers a simple yet powerful solution for matching statistical distributions of EEG datasets.
• The method effectively addresses the challenge of EEG signal variability in transfer learning.
• RPA paves the way for more efficient and user-friendly BCI systems by reducing calibration requirements.