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

Researchers have developed a method to distinguish between two types of uncertainty in facial expression recognition: ambiguity from human disagreement versus errors from distribution shift. The Uncertainty-Aware Routing system uses deep ensembles to separate aleatoric and epistemic uncertainty, enabling more intelligent handling of ambiguous faces versus out-of-distribution inputs.

Facial expression recognition systems face a fundamental challenge that extends beyond typical computer vision problems. When humans annotate facial expressions, they frequently disagree about the correct label, reflecting genuine ambiguity in the data itself. Simultaneously, models deployed in real-world environments encounter images that differ significantly from training data due to lighting, angles, or other variations. The critical insight here is that these two failure modes require different responses: ambiguous in-distribution samples should be flagged with their uncertainty for human review, while out-of-distribution samples should be rejected entirely. Traditional uncertainty quantification conflates these distinct phenomena into a single score, making it impossible to route decisions appropriately.

This work builds on established uncertainty decomposition theory, separating aleatoric uncertainty (data ambiguity) from epistemic uncertainty (model knowledge gaps). By validating each component against independent signals—human annotator disagreement for aleatoric uncertainty and corruption-induced distribution shift for epistemic uncertainty—the researchers provide empirical proof that meaningful separation is achievable. The use of DINOv2 ensemble models demonstrates strong practical performance, recovering annotator disagreement with 0.66 Spearman correlation and detecting distribution shifts with 0.699 AUROC.

The routing mechanism proves more practical than existing approaches, retaining 1.8 times more ambiguous in-distribution samples at matched rejection rates. This has direct implications for real-world deployment: medical imaging systems, security applications, and emotion-aware interfaces all benefit from distinguishing genuine label uncertainty from model failure. The dual-validation protocol establishes a replicable methodology for evaluating uncertainty decomposition methods beyond facial expression recognition, potentially influencing how practitioners design robust vision systems across domains.

• →Uncertainty decomposition distinguishes between aleatoric uncertainty (label ambiguity) and epistemic uncertainty (distribution shift), requiring different handling strategies
• →Deep ensembles of DINOv2 models achieve 0.66 correlation with human annotator disagreement and 0.699 AUROC for detecting corruption-induced shifts
• →Uncertainty-Aware Routing retains 1.8x more ambiguous in-distribution samples than single-uncertainty baselines at equivalent out-of-distribution rejection rates
• →The proposed dual-validation protocol provides a replicable methodology for evaluating uncertainty decomposition across computer vision tasks
• →Practical deployment benefits include more intelligent handling of ambiguous facial expressions in medical, security, and human-computer interaction applications