Transfer Learning: A Riemannian Geometry Framework With Applications to Brain-Computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) often require extensive user-specific calibration.
• Transfer learning aims to reduce calibration by leveraging existing data.
• Cross-session and cross-subject variability pose significant challenges for BCI transfer learning.

Purpose of the Study:

• To develop a method for improving transfer learning in electroencephalogram (EEG)-based BCIs.
• To address the challenges of cross-session and cross-subject classification.
• To enable calibration-less BCI operation through effective data transfer.

Main Methods:

• Representing EEG data using spatial covariance matrices.
• Utilizing Riemannian geometry for analysis of symmetric positive definite (SPD) matrices.
• Proposing an affine transformation to normalize covariance matrices across sessions/subjects.
• Implementing classification using minimum distance to mean and probabilistic Riemannian Gaussian distributions.

Main Results:

• Demonstrated significant improvements in BCI classification performance.
• Validated the effectiveness of the affine transformation on two BCI datasets.
• Showcased the ability to make data from different sessions and subjects comparable.

Conclusions:

• The proposed affine transformation effectively addresses BCI transfer learning challenges.
• This method enhances the comparability of EEG data across diverse sessions and subjects.
• The approach offers a promising solution for calibration-less BCI systems.