Multiclass Informative Instance Transfer Learning Framework for Motor Imagery-Based Brain-Computer Interface.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery tasks using noninvasive electroencephalography (EEG) for brain-computer interfaces (BCIs) demand extensive user training data.
• User fatigue is a significant issue due to prolonged data collection sessions.

Purpose of the Study:

• To investigate and enhance transfer learning strategies for multiclass motor imagery BCI.
• To reduce the substantial training data requirements for novel users.

Main Methods:

• Direct instance transfer from previous users was applied to multiclass motor imagery BCI.
• Active learning (AL) driven informative instance transfer was explored.
• An optimal ensemble of direct and informative transfer methods was designed and implemented.

Main Results:

• Informative instance transfer outperformed direct transfer, reducing training data by 49% for most subjects.
• The optimized ensemble method surpassed individual transfer approaches for nearly all subjects.
• The ensemble achieved benchmark performance with an average of 75% less training data across 8 out of 9 subjects.

Conclusions:

• A generic transfer learning framework using an optimized ensemble significantly reduces data requirements for new BCI users.
• This approach enhances the efficiency and accessibility of motor imagery BCI systems.
• The findings pave the way for more practical and less demanding BCI applications.