A transient neural code for feedback-driven motor corrections during reaching

Abstract

Movement is the result of complex, dynamic interaction between cortical and subcortical circuits. These dynamic interactions implement both feedforward motor control, arising from preparatory states, and feedback control, triggered by unexpected sensory events during movement. We show that the neural responses for feedback-driven control can be transient and small in variance, posing difficulties for unsupervised inference methods. We thus propose the Behavior-Aligned Neural Dynamics (BAND) model, which exploits semi-supervised learning to extract latent dynamics that predict both feedforward planned movement and unplanned feedback corrections. Our analysis suggests that motor corrections during movement 1) are encoded on the population level in small neural variability in primary motor (M1), but not dorsal premotor (PMd) cortex; 2) are transient; and 3) are driven by sensory feedback. Our work highlights the importance of targeted closed-loop aware methods to extract and study neural dynamics underlying complex behavioral phenomena.