Learning is a functional state of the brain that should be understood as a continuous process, rather than being restricted to the very moment of its acquisition, storage, or retrieval. The cerebellum operates by comparing predicted states with actual states, learning from errors, and updating its internal representation to minimize errors. In this regard, we studied cerebellar interpositus nucleus (IPn) functional capabilities by recording its unitary activity in behaving rabbits during an associative learning task: the classical conditioning of eyelid responses. We recorded IPn neurons in rabbits during classical eyeblink conditioning using a delay paradigm. We found that IPn neurons reduce error signals across conditioning sessions, simultaneously increasing and transmitting spikes before the onset of the unconditioned stimulus. Thus, IPn neurons generate predictions that optimize in time and shape the conditioned eyeblink response. Our results are consistent with the idea that the cerebellum works under Bayesian rules updating the weights using the previous history.