Demystifying excessively volatile human learning: A Bayesian persistent prior and a neural approximation