Robust neural decoding by kernel regression with Siamese representation learning.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-machine interfaces (BMIs) offer potential for motor function restoration.
• Current BMIs suffer from unstable performance due to neural signal variability.
• Fixed decoding functions are inadequate for dynamic neural activity.

Purpose of the Study:

• To develop a robust kernel regression framework for BMIs.
• To address neural signal variability and improve decoding accuracy.
• To enhance the stability and reliability of BMI systems.

Main Methods:

• Proposed a novel nonparametric kernel regression framework.
• Utilized Siamese networks to learn neural signal representations.
• Constrained representations with kinematic parameters for end-to-end learning.
• Handled neural variations caused by tuning function changes and intrinsic noise.

Main Results:

• The proposed approach outperformed existing methods on two datasets.
• Demonstrated significant improvements in robustness with limited samples and missing channels.
• Showcased effective tackling of neural variations.

Conclusions:

• The novel framework offers robust performance across different conditions.
• Provides a new perspective for developing robust BMI control.
• Highlights the potential of adaptive decoding for improved motor function restoration.