Cross-Subject Event-Related Potential Classification via Multi-View Based Contrastive Learning

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Event-related potentials (ERPs) are crucial for brain-computer interfaces (BCIs), providing feedback and error signals.
• Existing BCI models struggle with inter-subject variability, limiting their generalization to new users.
• Acquisition noise further degrades the performance of BCI models across different subjects.

Purpose of the Study:

• To develop a multi-view contrastive learning domain generalization (MVCLDG) method for enhanced cross-subject ERP recognition.
• To improve the discriminative feature extraction and learn domain-invariant representations for BCIs.
• To address the limitations of inter-subject variability and acquisition noise in ERP-based BCIs.

Main Methods:

• MVCLDG fuses raw electroencephalography (EEG) with phase information using multi-scale inception blocks for comprehensive feature extraction.
• Domain-alignment and contrastive learning constraints are applied to minimize distributional discrepancies and enhance class separability.
• The method was validated on ERN and semantic-syntactic violation datasets in cross-subject settings.

Main Results:

• MVCLDG significantly outperformed baseline and existing domain generalization methods in cross-subject ERP recognition.
• No additional target-domain adaptation was required, demonstrating robust generalization.
• Ablation studies confirmed the efficacy of individual components, and visualizations supported neurophysiological interpretability.

Conclusions:

• MVCLDG presents an effective strategy for ERP recognition by combining multi-view feature mining with contrastive domain generalization.
• The method yields improved and interpretable cross-subject ERP recognition, advancing closed-loop BCIs.
• This approach enhances the feasibility of user-generalizable ERP-based BCIs.