Advancing Bayesian deep learning : Sensible priors and accelerated inference