Motor error signals provide a rich source of information that can be used to sculpt adaptive changes in motor output that reduce future errors. A key question is how these error signals are constructed from sensory information about motor errors due to external disturbances, internally-generated motor output noise, or a combination of the two. In particular, motor errors arise both from externally-generated contributions, which provide a useful teaching signal for reshaping motor output, and from internally-generated motor output noise which does not. This random internally-generated motor output noise is what prevents us from repeating actions with perfect precision, and its random nature means that noise in future actions is independent of noise in previous actions and thus cannot be predicted, which makes learning from past noise-induced effects futile. Here we show that the teaching signal for motor learning is predictively denoised so that adaptive changes in motor output reflect specific responses to external perturbations unmuddied by the effects of internally-generated errors. This increases the fidelity of motor learning and provides new insight into the acuity of sensorimotor prediction.
