Unsupervised nonmotor learning in the human cerebellum

A core motor function of the cerebellum is error-based learning: it uses prediction errors-mismatches between predicted and actual sensory feedback-to refine actions. One provocative hypothesis is that the cerebellum performs similar computations in cognitive domains. Here, we ask whether error-based learning computations in the cerebellum extend to a passive statistical learning task that requires neither action nor decision-making. Human participants viewed sequences of visual stimuli with varying probabilistic stimulus-stimulus transitions while undergoing fMRI. Subjects reported no explicit knowledge of the transitional probabilities after the scan. Nonetheless, putative cognitive regions of the cerebellum encoded prediction errors during statistical learning. Moreover, error signals in the cerebellum were encoded in a different manner than error signals in the anterior hippocampus, a well-known substrate of statistical learning: cerebellar activity covaried with trial-by-trial prediction errors that evolved slowly over time (echoing cerebellar computations in motor learning), whereas the hippocampus responded in a more fixed manner to the underlying transition structure. Computational modeling formalized this dissociation, with cerebellar activity explained by a delta-rule model that incrementally adjusted predictions based on recent experience, and hippocampal activity characterized as a form of Bayesian updating of a transition distribution held in memory. Our findings demonstrate that the cerebellum encodes prediction errors during passive nonmotor learning, and thus that it likely plays a domain-general role in error-based learning.