Improving subject transfer in EEG classification with divergence estimation.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) classification models often fail on new subjects due to performance degradation.
• High signal variability between subjects necessitates time-intensive calibration for current EEG models.

Purpose of the Study:

• To improve the performance of EEG classification models on unseen subjects using novel regularization techniques.
• To reduce the need for user-specific calibration in EEG signal modeling.

Main Methods:

• Proposed graphical models to identify ideal statistical relationships in EEG data.
• Developed regularization penalties using Mutual Information and Wasserstein-1 divergences to enforce these relationships.
• Implemented efficient estimation algorithms for these divergences during model training using secondary neural networks.

Main Results:

• Extensive experiments on a large EEG dataset demonstrated significant performance improvements over unregularized models.
• Proposed techniques showed superior performance and stability compared to a baseline adversarial classifier across various hyperparameters.
• The computational cost of the proposed methods was comparable to the baseline.

Conclusions:

• The novel regularization techniques effectively enhance EEG classification generalization, particularly in zero-shot subject transfer scenarios.
• These advancements hold the potential to significantly reduce or eliminate the need for user-specific model calibration in EEG applications.