Not Just How Much, But Where: Decomposing Epistemic Uncertainty into Per-Class Contributions
Researchers have developed a novel method to decompose epistemic uncertainty in Bayesian deep learning models into per-class contributions. This new metric, termed $C_k(x)$, allows for a more nuanced understanding of model ignorance, particularly in safety-critical applications where the cost of failure is asymmetric. By decomposing mutual information (MI) into a vector that weights uncertainty by class, the method improves selective prediction accuracy and provides better out-of-distribution detection compared to traditional scalar MI. AI
IMPACT This research could lead to more reliable AI systems in safety-critical domains by providing a clearer understanding of model uncertainty.