Improving Transfer Performance of Deep Learning with Adaptive Batch Normalization for Brain-computer Interfaces.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCIs) utilize deep learning and transfer learning to address user variability.
• Generalization of BCIs across different datasets remains a challenge.

Purpose of the Study:

• To compare transfer learning methods for BCIs in cross-dataset scenarios.
• To evaluate the effectiveness of different preprocessing strategies with transfer learning models.
• To introduce and assess AdaBN for domain adaptation in BCIs.

Main Methods:

• Compared manifold embedded knowledge transfer and pre-trained EEGNet.
• Investigated three preprocessing strategies.
• Implemented AdaBN for target domain adaptation.
• Separated source and target batch normalization layers.

Main Results:

• EEGNet combined with Riemannian alignment and AdaBN yielded the highest transfer accuracy.
• Achieved approximately 65.6% accuracy on the target dataset.
• Demonstrated the benefit of domain adaptation techniques.

Conclusions:

• The proposed EEGNet with Riemannian alignment and AdaBN shows promising results for cross-dataset BCI generalization.
• Separating batch normalization layers is a potentially valuable technique in domain adaptation for BCIs.
• This study offers insights for designing more robust transfer neural networks for BCIs.