Impaired representation of temporal statistics in cerebellar degeneration

To increase perceptual precision, the brain is hypothesized to perform Bayesian inference, evaluating sensory input within the context of the statistical structure of the environment. The cerebellum is a strong candidate for this computation in interval timing: its integrity is essential for precise timing and supports encoding temporal statistics in animals. However, how the sensory input is combined with prior knowledge within this circuit remains unknown. To examine this, we tested patients with cerebellar degeneration and matched controls on duration reproduction tasks with varying prior distributions. Patients showed the signatures of Bayesian inference, but with greater timing variability, stronger regression toward the mean, and slower updating of the prior. We developed a model of the cerebellar microcircuit in which Purkinje cell plasticity over a granule cell temporal basis set provides a shared substrate for learning the prior and performing the inference. Simulations of cerebellar degeneration showed an output approximated a Bayesian observer with an intact prior and a noisier likelihood. This model provided a good fit of the patients' behavior over a wide range of parameter space. These results suggest a critical role for the human cerebellum in Bayesian inference, supporting optimally timed behavior even under extensive neuropathology.