Handling Few Training Data: Classifier Transfer Between Different Types of Error-Related Potentials.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) often require extensive training data for classifier calibration.
• Limited training data in real-world or rehabilitation settings poses a significant challenge for BCI usability.
• Error-related potentials (ErrPs) are crucial neural signals for feedback in BCIs, but their collection can be time-consuming.

Purpose of the Study:

• To propose and validate an application-oriented approach for transferring trained classifiers in BCIs.
• To address the challenge of insufficient training data in complex application or rehabilitation scenarios.
• To reduce the calibration time required for ErrP-based BCIs.

Main Methods:

• Developed a classifier transfer approach applicable across different scenarios and even different types of event-related potentials.
• Transferred a classifier trained on observation error-related potentials (ErrPs) to detect interaction ErrPs within the same subject.
• Compared the proposed transfer method against a cross-subject transfer approach using the same ErrP type.

Main Results:

• The proposed classifier transfer approach is feasible and effective.
• Transferring a classifier between different types of ErrPs within the same subject outperformed cross-subject transfer.
• The method successfully detected a different brain pattern (interaction ErrPs) using a classifier trained on a similar pattern (observation ErrPs).

Conclusions:

• The developed transfer approach offers a promising solution for handling limited training data in ErrP-based BCIs.
• This method can significantly reduce the calibration time needed for BCI systems.
• The approach facilitates the deployment of BCIs in practical and clinical settings where data collection is constrained.