Our new paper A Data-Driven Measure of Relative Uncertainty for Misclassification Detection has been accepted to appear at ICLR 2024. In this paper we proposed a data-driven method, powered by a statistical diversity and dissimilarity metric, to detect incorrect classification at test time by assessing the uncertainty of a given model.
A huge shout-out to my great colleagues Eduardo, Georg, and Pablo for their fantastic work!
A preliminary version of this paper has appeared at the NeurIPS 2023 Workshop on Mathematics of Modern Machine Learning.