Riemannian Locality Preserving Method for Transfer Learning With Applications on Brain-Computer Interface.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) show promise for medical rehabilitation and commercial use.
• Transfer learning is crucial for addressing inter-subject variability in electroencephalography (EEG) signals within BCIs.
• Existing methods struggle to effectively adapt EEG data across different subjects or sessions.

Purpose of the Study:

• To propose a novel transfer learning method for BCIs that preserves Riemannian data structure and adapts joint distributions.
• To enhance the effectiveness of transfer learning in BCIs by addressing domain shifts in EEG signals.
• To introduce the Riemannian locality preserving-based transfer learning (RLPTL) method.

Main Methods:

• Constructing a Riemannian graph based on Riemannian distance to capture geometric data information.
• Embedding the Riemannian graph within the joint distribution adaptation (JDA) framework.
• Developing the Riemannian locality preserving-based transfer learning (RLPTL) algorithm.

Main Results:

• The proposed RLPTL method achieved the highest accuracies in both multi-source domain (MSD) and single-source domain (SSD) experiments.
• RLPTL outperformed eight baseline methods across three motor imagery datasets.
• The method demonstrated superior performance in handling inter-subject variability in EEG signals.

Conclusions:

• The RLPTL method offers a feasible and efficient approach for transfer learning in BCIs.
• Preserving Riemannian locality and adapting joint distributions significantly improves BCI performance.
• This work advances the application of transfer learning for robust EEG signal analysis in BCIs.