Ensembling as Approximate Bayesian Inference for Predictive Uncertainty Estimation in Deep Learning


We view ensembling as an approximate Bayesian inference method, justify why it should be a reasonable approximation for Deep Neural Networks and extensively compare it with other approximate methods in terms of predictive uncertainty estimation quality. We provide experimental results on illustrative toy problems and the real-world computer vision tasks of street-scene semantic segmentation and depth completion.

SSDL, 2019 (Oral)

This extended abstract describes preliminary results from the paper Evaluating Scalable Bayesian Deep Learning Methods for Robust Computer Vision.