[Paper Review] Bayesian OOD detection with aleatoric uncertainty and outlier exposure

Xi Wang, Laurence Aitchison|arXiv (Cornell University)|Feb 24, 2021

Adversarial Robustness in Machine Learning39 references4 citations

TL;DR

This paper proposes a principled Bayesian framework for out-of-distribution (OOD) detection by integrating aleatoric uncertainty and outlier exposure (OE). It models dataset curation as a consensus process where ambiguous, ill-defined class-label images (UCL points) cause annotator disagreement, signaling aleatoric uncertainty; this uncertainty is formalized as a likelihood for OOD points, enabling Bayesian inference with OE. The method outperforms both standard Bayesian and aleatoric uncertainty approaches on OOD detection benchmarks.

ABSTRACT

Typical Bayesian approaches to OOD detection use epistemic uncertainty. Surprisingly from the Bayesian perspective, there are a number of methods that successfully use aleatoric uncertainty to detect OOD points (e.g. Hendryks et al. 2018). In addition, it is difficult to use outlier exposure to improve a Bayesian OOD detection model, as it is not clear whether it is possible or desirable to increase posterior (epistemic) uncertainty at outlier points. We show that a generative model of data curation provides a principled account of aleatoric uncertainty for OOD detection. In particular, aleatoric uncertainty signals a specific type of OOD point: one without a well-defined class-label, and our model of data curation gives a likelihood for these points, giving us a mechanism for conditioning on outlier points and thus performing principled Bayesian outlier exposure. Our principled Bayesian approach, combining aleatoric and epistemic uncertainty with outlier exposure performs better than methods using aleatoric or epistemic alone.

Motivation & Objective

• To address the lack of principled Bayesian integration of outlier exposure (OE) in OOD detection.
• To provide a formal likelihood for out-of-distribution points with undefined class-labels (UCL points) using aleatoric uncertainty.
• To unify epistemic uncertainty, aleatoric uncertainty, and OE within a single Bayesian framework.
• To improve OOD detection performance by modeling human annotator consensus during dataset curation.

Proposed method

• Models dataset curation as a consensus process among multiple annotators, where agreement implies well-defined class-labels and disagreement implies undefined labels (UCL points).
• Formalizes the disagreement in annotator responses as aleatoric uncertainty, which is linked to the model's predictive probabilities.
• Derives a likelihood function for UCL points based on the probability of consensus failure, enabling principled Bayesian inference with outlier data.
• Integrates this likelihood into Bayesian neural network (BNN) training, allowing outlier exposure to be treated as a likelihood term.
• Uses the generative model to condition on OOD points with undefined labels, enabling Bayesian OE without ad-hoc objectives.
• Combines epistemic uncertainty (from BNNs) and aleatoric uncertainty (from consensus failure) for improved OOD detection.

Experimental results

Research questions

• RQ1 Can aleatoric uncertainty be formally justified as a signal for OOD detection in a Bayesian framework?
• RQ2 How can outlier exposure be integrated into Bayesian OOD detection using principled likelihoods?
• RQ3 Can a generative model of dataset curation explain why aleatoric uncertainty detects certain types of OOD data?
• RQ4 Does combining aleatoric uncertainty, epistemic uncertainty, and Bayesian outlier exposure improve OOD detection performance?

Key findings

• The proposed method achieves superior AUROC and FPR95 scores compared to baseline Bayesian methods without outlier exposure.
• It outperforms standard aleatoric uncertainty-based methods that use outlier exposure without principled likelihoods (e.g., Hendrycks et al., 2018).
• The integration of outlier exposure via a likelihood for UCL points leads to measurable performance gains in OOD detection.
• Aleatoric uncertainty effectively signals OOD points with undefined class-labels, as confirmed by the consensus-based generative model.
• The method provides a formal Bayesian interpretation for OE, resolving the ambiguity in existing ad-hoc OE objectives.