Exploiting Multiple EEG Data Domains with Adversarial Learning.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) is crucial for assessing mental states but faces challenges like poor signal-to-noise ratio and domain dependency.
• Machine learning models trained on EEG data often exhibit poor generalization due to subject-specific and equipment-related variations.
• Existing transfer learning methods for EEG signals require significant domain calibration.

Purpose of the Study:

• To propose a multi-source learning framework using domain-invariant representations for EEG signals.
• To address the challenge of poor generalization in machine learning models applied to multi-modal EEG data.
• To enable robust EEG-based brain-computer interfaces by reducing domain-specific information.

Main Methods:

• An adversarial inference approach was developed to learn domain-invariant representations from multiple EEG data sources.
• EEG recordings from diverse emotion recognition datasets (SEED, SEED-IV, DEAP, DREAMER) were unified.
• The method aimed to suppress data-source-relevant information while preserving essential discriminative features.

Main Results:

• The proposed invariant representation learning approach successfully suppressed data-source-relevant information leakage by 35%.
• Stable performance in EEG-based emotion classification was maintained despite unifying data from different domains.
• The method demonstrated the feasibility of multi-source learning for EEG data.

Conclusions:

• Multi-source learning via domain-invariant representations is a viable alternative for EEG signal analysis.
• The adversarial inference approach effectively learns representations that generalize across different EEG data sources.
• This technique enhances the robustness and applicability of EEG-based brain-computer interfaces in real-world scenarios.