State-Reification Networks: Improving Generalization by Modeling the Distribution of Hidden Representations