BaCOUn: Bayesian Classifers with Out-of-Distribution Uncertainty

TL;DR

This paper introduces BaCOUn, a Bayesian framework that enhances deep classifiers with reliable uncertainty estimates by augmenting data with boundary points and applying Bayesian inference, improving robustness under dataset shifts.

Contribution

The paper presents a novel Bayesian method combining data augmentation with boundary points and Bayesian inference to improve uncertainty estimation in deep classifiers.

Findings

Enhanced uncertainty estimates under dataset shift

Improved classifier robustness and calibration

Effective detection of out-of-distribution data

Abstract

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.