Late Bayesian inference in sensorimotor behavior

Abstract

Sensorimotor skills rely on performing noisy sensorimotor computations on noisy sensory measurements. Bayesian models suggest that humans compensate for measurement noise and reduce behavioral variability by biasing perception toward prior expectations. Whether the same holds for noise in sensorimotor computations is not known. Testing human subjects in tasks with different levels of sensorimotor complexity, we found a similar bias-variance tradeoff associated with increased sensorimotor noise. This result was accurately captured by a model which implements Bayesian inference after - not before - sensorimotor transformation. These results indicate that humans perform \"late inference\" downstream of sensorimotor computations rather than, or in addition to, \"early inference\" in the perceptual domain. The brain thus possesses internal models of noise in both sensory measurements and sensorimotor computations.