Domain-Incremental Learning Framework for Continual Motor Imagery EEG Classification Task.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Inter-subject variability in electroencephalogram (EEG) signals limits brain-computer interface (BCI) model generalization.
• Traditional transfer learning (TL) suffers from catastrophic forgetting in continual learning, degrading performance over time.
• Existing methods struggle with sustained knowledge transfer in BCI applications.

Purpose of the Study:

• To develop a novel domain-incremental learning framework for continual motor imagery (MI) EEG classification.
• To address the challenge of catastrophic forgetting in BCI models.
• To enhance the long-term generalization ability of BCI systems.

Main Methods:

• Separation of subject-invariant and subject-specific features using adversarial training.
• Implementation of an extensible architecture to preserve vulnerable features.
• Incorporation of a memory replay mechanism for knowledge reinforcement.

Main Results:

• The proposed framework effectively mitigates catastrophic forgetting in continual MI-EEG classification.
• Demonstrated significant improvements in model performance over multiple transfer learning steps.
• Successfully retained knowledge across incremental learning tasks.

Conclusions:

• The novel domain-incremental learning framework offers a robust solution for continual BCI learning.
• The approach enhances the robustness and longevity of BCI models in real-world applications.
• This work advances the field of BCI by tackling the critical issue of knowledge retention.