Fast and scalable bayesian deep learning by weight-perturbation in Adam
accepted at
ICML2018
!