Improving Generalization Based on l₁-Norm Regularization for EEG-Based Motor Imagery Classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Multichannel electroencephalography (EEG) is fundamental to brain-computer interface (BCI) systems.
• EEG classification algorithms often struggle with redundant features, leading to model complexity, overfitting, and reduced generalization.
• Existing methods like Common Spatial Pattern (CSP) can be parameter-intensive.

Purpose of the Study:

• To develop a novel l1-norm-based approach for combining EEG channel decision values.
• To decrease model complexity and improve the generalization of EEG classification methods.
• To enhance the classification accuracy and practical applicability of motor imagery (MI) based BCI systems.

Main Methods:

• An l1-norm regularization technique is applied to extract information from independent frequency bands across different EEG channels.
• The proposed method combines decision values directly from EEG channels.
• An efficient optimization solution is developed to minimize the objective function.

Main Results:

• The l1-norm approach significantly reduces the number of parameters compared to CSP methods, mitigating overfitting.
• High classification accuracy was achieved on BCI competition datasets (IVa and I).
• The method demonstrated robust generalization performance, maintaining accuracy as training data decreased.

Conclusions:

• The proposed l1-norm-based method offers a more efficient and generalizable approach for EEG classification in BCI.
• It effectively addresses overfitting issues common in EEG analysis.
• The findings support the practical application of MI-based BCI systems through improved classification performance.