Interpretable Uncertainty Routing Separating Emotion Ambiguity from Distribution Shift in Facial Expression Recognition

The study introduces Uncertainty-Aware Routing (UAR) for facial expression recognition (FER), which decomposes uncertainty into aleatoric and epistemic components to address ambiguity and distribution shift. Utilizing a Deep Ensemble of fully fine-tuned DINOv2 models, the approach achieves a Spearman correlation of 0.66 for annotator disagreement and an AUROC of 0.699 for detecting corruption-induced shifts. This separation allows for more interpretable actions in model deployment, enabling better handling of ambiguous in-distribution faces while effectively rejecting out-of-distribution inputs, which is crucial for practitioners aiming to improve FER systems in real-world applications.