Trust Me, I Know the Way: Predictive Uncertainty in the Presence of Shortcut Learning

TL;DR

The paper investigates how predictive uncertainty should be represented when shortcut learning drives spurious correlations. It leverages a Bayesian ensemble framework and the entropy decomposition $H(Y)=H(Y| ext{Θ})+I(Y; ext{Θ})$ to separate aleatoric and epistemic components and to interpret EU as model disagreement. Through MNIST-based experiments with CMNIST3 and PMNIST3 shortcuts, the authors show that shortcuts can provoke disagreement among ensemble members, yielding high EU on OOD data while shortcut-free data yield more diffuse uncertainty. The work provides a step toward reconciling ignorance and disagreement perspectives on EU and highlights the significant impact of distribution shifts on uncertainty quantification in practical settings.

Abstract

The correct way to quantify predictive uncertainty in neural networks remains a topic of active discussion. In particular, it is unclear whether the state-of-the art entropy decomposition leads to a meaningful representation of model, or epistemic, uncertainty (EU) in the light of a debate that pits ignorance against disagreement perspectives. We aim to reconcile the conflicting viewpoints by arguing that both are valid but arise from different learning situations. Notably, we show that the presence of shortcuts is decisive for EU manifesting as disagreement.

Figures (5)

• Figure 1: Examples of 3-class CMNIST3 data (left) with full-image color shortcut and PMNIST3 (right; cropped for better visibility) with $1 \times 1$ colored patches (shortcuts in 95% of images).
• Figure 2: Uncertainty estimates (20% least-confident predictions) based on data with varying $s$; ensemble size 3; average over 3 independent runs (error bars: $\pm$ 1 standard error).
• Figure 3: More epochs. Uncertainty estimates (20% least-confident predictions) based on data with varying $s$; ensemble size 3; average over 3 independent runs (error bars: $\pm$ 1 standard error).
• Figure 4: Larger ensemble. Uncertainty estimates (20% least-confident predictions) based on data with varying $s$; ensemble size 5; average over 3 independent runs (error bars: $\pm$ 1 standard error).
• Figure 5: In-distribution test data. Uncertainty estimates (20% least-confident predictions) based on data with varying $s$, where train and test classes coincide but shortcuts are only present in the training data; ensemble size 3; average over 3 independent runs (error bars: $\pm$ 1 standard error).