Riemannian geometry-based transfer learning for reducing training time in c-VEP BCIs.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) require extensive training data for calibration before each use, limiting their practical application.
• Transfer learning offers a potential solution to reduce this calibration burden.
• Code modulated visual evoked potential (c-VEP) based BCIs are a specific type of BCI that can benefit from improved training efficiency.

Purpose of the Study:

• To propose and evaluate a novel Riemannian geometry-based transfer learning algorithm for c-VEP BCIs.
• To significantly reduce the calibration time of BCIs without compromising classification accuracy.
• To enhance the real-world applicability of c-VEP BCIs.

Main Methods:

• Development of a Riemannian geometry-based transfer learning algorithm incorporating log-Euclidean data alignment (LEDA) and training accuracy-based subject selection (TSS).
• Implementation of super-trial construction, covariance matrix estimation, and minimum distance to mean classification.
• Evaluation using leave-one-subject-out (LOSO) cross-validation on data from sixteen subjects in a c-VEP BCI experiment.

Main Results:

• The proposed algorithm demonstrated significantly higher classification accuracy compared to the subject-specific baseline algorithm.
• The transfer learning approach effectively reduced the required training time for BCI calibration at equivalent performance levels.
• LEDA was shown to minimize data distribution differences between subjects, while TSS enhanced target-source subject similarity.

Conclusions:

• The Riemannian geometry-based transfer learning algorithm is highly effective for c-VEP BCIs.
• This method substantially decreases BCI calibration time and improves classification performance.
• The findings facilitate the practical deployment of BCIs in real-world scenarios.