Towards Bayesian Deep Learning: A Framework and Some Existing Methods