Can someone explain to me the "Bayesian methods don't overfit" argument