Local-learning-based neuron selection for grasping gesture prediction in motor brain machine interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• High-dimensional neural recordings pose computational challenges for motor brain-machine interfaces (mBMI).
• Not all neural activity directly correlates with specific movement tasks, necessitating efficient neuron selection.

Purpose of the Study:

• To propose and evaluate a local-learning-based method for neuron selection in gesture prediction for reaching and grasping tasks.
• To reduce computational burden in portable mBMI systems by identifying essential neurons.

Main Methods:

• Nonlinear neural activity decomposition into a weighted feature space.
• Definition of a margin to distinguish inter-class and intra-class neural patterns.
• Minimization of a margin-based exponential error function with 1-norm regularization for sparse weights, identifying important neurons.

Main Results:

• A small subset of 10 neurons achieved over 95% of the full recording's decoding accuracy for gesture prediction.
• Selected neurons exhibited visually distinguishable temporal patterns correlated with hand states.
• The proposed method outperformed other algorithms in eliminating irrelevant neurons and identifying the optimal subset for decoding.

Conclusions:

• The algorithm effectively identifies neuronal importance without assuming a coding model, offering high performance across different decoding models.
• The method demonstrates robustness with noisy signals and fast convergence, indicating feasibility for portable BMI systems.
• Reduced computational load and maintained performance enhance the practicality of mBMI systems.