Sensory predictions are embedded in cortical motor activity

While moving through the world, such as biking down a bumpy trail, our bodies encounter external forces that cannot be predicted by our own motor output. Delayed sensory feedback is generally too slow to allow accurate state estimation during responses to external forces, so the nervous system may solve this problem by using prior information to make rapid sensory predictions during feedback control. However, it is unknown how flexibly these priors can be set nor how neural activity underlies these fundamental abilities. To answer this, we designed a reaching task in which subjects (human or macaque) were visually cued about the probability distribution from which the direction of upcoming elbow perturbations would be drawn. On single trials and without extensive training humans integrate these probability cues into sensory priors that bias their muscle activity within 70 ms of perturbation onset, faster than the time it takes to initiate a ‘voluntary’ movement. In a similar experiment, a monkey showed the same ability to use sensory predictions during rapid feedback responses. High-density recordings revealed a clear signature of sensory priors in dorsal premotor (PMd) and motor cortex (M1), but not in somatosensory cortex (S1). To understand the computational basis of this signal, we trained artificial neural networks to control a realistic biomechanical model of the arm and perform similar tasks. The model readily learned to take advantage of sensory priors, showing remarkably similar behavior, muscle and neural activity to humans and monkeys. These combined results uncover a link between sensory predictions and movement – neural dimensions responsible for detecting the direction of a perturbation are biased by sensory priors during motor planning, allowing the brain to rapidly act when the perturbation is detected but before the perturbation direction is fully resolved.