Riemannian transfer learning based on log-Euclidean metric for EEG classification

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Brain-computer interfaces (BCI) enable direct brain-device communication, bypassing peripheral nerves.
• Decoding brain intentions for real-time control is a critical BCI challenge.
• Existing Riemannian geometry methods for EEG decoding struggle with data distribution shifts, impacting cross-session and cross-subject performance.

Purpose of the Study:

• To develop a Riemannian transfer learning method for BCI that accounts for data distribution drift.
• To improve the robustness and accuracy of EEG signal decoding in BCI applications.

Main Methods:

• Proposed two Riemannian transfer learning methods utilizing the log-Euclidean metric.
• Leveraged historical data (source domain) to train Riemannian decoders for current tasks.
• Utilized data from other subjects to enhance decoder training for a target subject.

Main Results:

• Verified proposed methods on BCI competition datasets (III, IIIa, and IV 2a).
• Demonstrated superior classification performance compared to baseline methods without transfer learning.
• Achieved comparable performance to affine invariant Riemannian metric methods but with significantly higher computational efficiency.

Conclusions:

• The proposed transfer learning method enhances Riemannian classifier performance in BCI.
• The unsupervised and time-efficient nature of the transfer learning process shows potential for online BCI learning scenarios.