Deep transfer learning-based decoder calibration for intracortical brain-machine interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Intracortical brain-machine interfaces (iBMIs) enable brain-device communication.
• Signal non-stationarity in iBMIs requires frequent decoder recalibration, demanding extensive new data.
• Current recalibration methods are time-consuming and data-intensive.

Purpose of the Study:

• To develop a novel method for efficient iBMI decoder recalibration.
• To minimize the need for new data during the recalibration process.
• To leverage deep transfer learning for improved iBMI performance.

Main Methods:

• Proposed an active learning domain adversarial neural network (AL-DANN).
• Utilized historical neural data with a small amount of current data (four samples per category).
• Employed domain adversarial and active learning strategies for knowledge transfer.

Main Results:

• AL-DANN outperformed existing state-of-the-art methods in decoder calibration.
• Achieved over 80% reduction in recalibration time.
• Required only four new samples per category for effective recalibration.

Conclusions:

• AL-DANN offers a highly efficient solution for iBMI decoder recalibration.
• Deep transfer learning shows significant potential for advancing iBMI technology.
• This method reduces the burden of data collection for daily iBMI use.