BaCOUn: Bayesian Classifers with Out-of-Distribution Uncertainty
This work addresses the issue of unreliable uncertainty in deep learning models for applications requiring robustness to out-of-distribution data, representing an incremental improvement in uncertainty estimation methods.
The authors tackled the problem of deep classifiers being overconfident and unreliable under dataset shift by proposing a Bayesian framework that provides reliable uncertainty estimates, achieving improved uncertainty calibration without specifying concrete numerical results.
Traditional training of deep classifiers yields overconfident models that are not reliable under dataset shift. We propose a Bayesian framework to obtain reliable uncertainty estimates for deep classifiers. Our approach consists of a plug-in "generator" used to augment the data with an additional class of points that lie on the boundary of the training data, followed by Bayesian inference on top of features that are trained to distinguish these "out-of-distribution" points.