Accounting for surprise means building architectures that integrate the prior probabilities into their model