On the Complexity of Bayesian Generalization