Accurate neuroprosthetic control through latent state transition training

Abstract

AbstractHand movements are an essential way that primates interact with the environment and they comprise some of our most complex behaviors. Despite remarkable recent progress in intracortical brain-computer interfaces, current neuroprostheses still lack the fine control required to interact with objects of daily living. We present a BCI training approach to develop neural activity latent variables for accurate hand shape control. The method shapes neural activity by daily training to target patterns and provides incremental control of degrees of freedom. When tested in two rhesus monkeys implanted in key areas of the cortical grasping circuit, our approach provides hand configuration accuracy comparable to native grasping. This training method was advantageous in a collision task, can potentially extend to control in higher-dimensions, and facilitates learning of novel actions in the kinematic latent space. Using our paradigm, neural activity patterns evolved to match the trained trajectories over time, making our approach a potential tool to study cortical learning.