Reproduce “Bayesian OOD detection with aleatoric uncertainty and outlier exposure” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xi Wang, Laurence Aitchison

The authors model dataset curation by treating annotator consensus as a defined label and disagreement as an undefined one. They train a Bayesian classifier with proxy outliers assigned an undefined-label likelihood, combining aleatoric and epistemic uncertainty. On CIFAR-10, the method reports 97.27% AUROC and 12.28% FPR95 on SVHN, outperforming Bayesian and deterministic baselines in tested undefined-label OOD settings.

Key insights

  • Aleatoric uncertainty provides a principled signal for OOD detection when inputs inherently lack a well-defined class label, as captured by the generative model of data curation.
  • A formal likelihood for "undefined class labels" enables the seamless and principled incorporation of Outlier Exposure into Bayesian models, moving beyond heuristic objectives.
  • Combining both aleatoric and epistemic uncertainties within a unified Bayesian framework leads to a more robust and effective OOD detection system.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.