Batch Mode Query by Committee for Motor Imagery-Based BCI.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCI) offer potential for neural rehabilitation and enhanced human-in-the-loop systems.
• Current BCI systems are largely confined to laboratory settings due to lengthy user training data collection requirements.
• Reducing training data while maintaining performance is crucial for BCI system advancement.

Purpose of the Study:

• To investigate the efficacy of active learning (AL) methods for reducing BCI training data requirements.
• To apply the query by committee (QBC) active learning strategy using both heterogeneous and homogeneous feature spaces.
• To evaluate the performance and data reduction capabilities of QBC-heterogeneous compared to other methods.

Main Methods:

• Implemented QBC by forming committees in both heterogeneous and homogeneous feature spaces.
• For heterogeneous space, combined three feature extraction methods with linear discriminant analysis.
• For homogeneous space, used random K-fold sampling after single-method feature extraction.

Main Results:

• QBC-heterogeneous achieved baseline performance using only 35% of the total training data.
• QBC-heterogeneous demonstrated statistically significant improvements over QBC-homogeneous, other AL methods, and random selection.
• The study confirmed significant reductions in labeling and data collection efforts compared to random labeling.

Conclusions:

• QBC-heterogeneous effectively reduces the calibration time and data requirements for BCI systems.
• This approach shows significant promise for making BCI technology more accessible and practical.
• QBC is a viable strategy for overcoming key limitations in current BCI system development.