Feature-Selection-Based Transfer Learning for Intracortical Brain-Machine Interface Decoding.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Intracortical brain-machine interfaces (iBMIs) face challenges in decoder calibration time and computational demands from high-dimensional neural data.
• Existing methods for decoder optimization or computational reduction often address only one problem, not both simultaneously.

Purpose of the Study:

• To develop a method that simultaneously optimizes decoder calibration and reduces computational burden in iBMIs.
• To introduce a novel transfer learning approach combined with feature selection for enhanced iBMI performance.

Main Methods:

• Proposed a symmetrical-uncertainty-based transfer learning (SUTL) method.
• Utilized symmetrical uncertainty to quantify feature stationarity, importance, and redundancy for selection.
• Applied the SUTL method to neural data from non-human primates for decoding motor intentions.

Main Results:

• SUTL significantly reduced the need for current data samples (ten per category) and computational load (less than 10% features, 30% channels).
• Achieved superior decoding performance compared to traditional methods.
• Feature selection via SUTL yielded a 6.6% higher average decoding accuracy than neuron selection.

Conclusions:

• The SUTL method effectively addresses key challenges in iBMI development, namely data requirements for calibration and computational complexity.
• SUTL offers a promising approach for improving the efficiency and performance of iBMIs.
• Feature selection proves more effective than neuron selection for enhancing decoding accuracy in iBMIs.