Smoothness as a failure mode of Bayesian mixture models in brain-machine interfaces.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Recursive Bayesian filters, specifically reach state equations (RSE), are used in brain-machine interfaces (BMIs) to translate neural signals into intended movements.
• While RSE provide smooth trajectories for known targets, hybrid models incorporating target uncertainty lead to less smooth movements.

Purpose of the Study:

• To investigate the cause of reduced trajectory smoothness in RSE-based hybrid models for BMIs with unknown targets.
• To identify factors influencing trajectory smoothness and propose methods for improvement.

Main Methods:

• Analysis of empirical spiking data from the primary motor cortex.
• Mathematical analysis focusing on angular velocity as a measure of movement smoothness.
• Simulations to test the impact of various smoothing techniques on trajectory performance.

Main Results:

• Angular velocity in BMI trajectories is directly proportional to changes in target probability.
• The proportionality constant depends on the difference in heading between filters for the most probable targets, indicating that closer targets enhance smoothness.
• Smoothing the data likelihood, increasing ensemble size, and achieving uniformity in preferred directions improve hybrid trajectory smoothness.

Conclusions:

• The loss of smoothness in hybrid BMI models is linked to target probability shifts.
• Strategies like optimizing target spacing, increasing neural ensemble size, and aligning preferred directions can enhance BMI performance.
• Future work may involve closed-loop training or neuronal subset selection to refine user tuning for smoother control.