Uncertainty in Deep Learning: How to be Bayesian?