Understanding Approximation for Bayesian Inference in Neural Networks