On the Expressiveness of Approximate Inference in Bayesian Neural Networks

While Bayesian neural networks (BNNs) hold the promise of being flexible, well-calibrated statistical models, inference often requires approximations whose consequences are poorly understood.