Keeping Neural Networks Simple by Minimizing the Description Length of the Weights