Invited talk: Using Loss Surface Geometry for Scalable Bayesian Deep Learning