Enhancing SSVEP-Based BCI Performance via Consensus Information Transfer Among Subjects.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) offer high-speed communication via steady-state visual evoked potentials (SSVEPs).
• Decoding accuracy in SSVEP BCIs is limited by the need for extensive subject-specific data.
• Leveraging existing datasets is crucial for improving SSVEP BCI performance.

Purpose of the Study:

• To introduce ConsenNet, a novel framework for enhancing SSVEP classification accuracy.
• To improve the generalization capability of SSVEP BCIs for new subjects.
• To extract invariant features across subjects for robust BCI performance.

Main Methods:

• Exploiting subject diversity to generate synthetic data that preserves task-related components and variability.
• Utilizing knowledge distillation to transfer structured inter-category relationships from a teacher to a student network.
• Incorporating a small amount of new subject data for final model calibration.

Main Results:

• ConsenNet demonstrated superior performance compared to 19 existing methods across three public SSVEP datasets.
• Offline experiments confirmed the framework's effectiveness in enhancing classification accuracy.
• Online experiments validated the practical feasibility of ConsenNet for real-world BCI applications.

Conclusions:

• ConsenNet effectively enhances SSVEP classification by leveraging diverse subject data and knowledge distillation.
• The framework improves BCI generalization to new subjects, addressing a key challenge in the field.
• ConsenNet shows significant promise for developing more accurate and adaptable SSVEP-based BCIs.