Smoothness as a Failure Mode of Bayesian Mixture Models in Brain–Machine Interfaces

Various recursive Bayesian filters based on reach state equations (RSE) have been proposed to convert neural signals into reaching movements in brain-machine interfaces. When the target is known, RSE produce exquisitely smooth trajectories relative to the random walk prior in the basic Kalman filter. More realistically, the target is unknown, and gaze analysis or other side information is expected to provide a discrete set of potential targets. In anticipation of this scenario, various groups have implemented RSE-based mixture (hybrid) models, which define a discrete random variable to represent target identity. While principled, this approach sacrifices the smoothness of RSE with known targets. This paper combines empirical spiking data from primary motor cortex and mathematical analysis to explain this loss in performance. We focus on angular velocity as a meaningful and convenient measure of smoothness. Our results demonstrate that angular velocity in the trajectory is approximately proportional to change in target probability. The constant of proportionality equals the difference in heading between parallel filters from the two most probable targets, suggesting a smoothness benefit to more narrowly spaced targets. Simulation confirms that measures to smooth the data likelihood also improve the smoothness of hybrid trajectories, including increased ensemble size and uniformity in preferred directions. We speculate that closed-loop training or neuronal subset selection could be used to shape the user's tuning curves towards this end.