Not Just How Much, But Where: Decomposing Epistemic Uncertainty into Per-Class Contributions

Researchers Mame Diarra Toure and David A. Stephens have published a paper introducing a novel method to decompose the epistemic uncertainty of AI models into per-class contributions. Current Bayesian deep learning summarizes a model's ignorance with a single scalar metric called mutual information (MI), which fails to distinguish whether uncertainty involves a benign or a safety-critical class—a critical flaw for applications like medical diagnosis. Their new framework, detailed in the arXiv preprint 'Not Just How Much, But Where: Decomposing Epistemic Uncertainty into Per-Class Contributions,' addresses this by breaking down MI into a vector C_k(x) for each possible class, derived from the mean and variance of the model's predicted probabilities across posterior samples. This decomposition, based on a second-order Taylor expansion, corrects for boundary suppression and allows for fair comparison across rare and common classes.

The practical impact is significant, especially in safety-critical fields. Validation on a diabetic retinopathy diagnosis task showed that using the critical-class component of C_k reduced 'selective prediction risk'—the cost of errors when the model abstains—by 34.7% compared to using standard MI and by 56.2% compared to simple variance baselines. The method also improved out-of-distribution detection and proved more robust to label noise in controlled studies. The research underscores that for reliable AI, *how* uncertainty is propagated through a model's architecture is as important as the metric used to measure it, pushing the field toward more interpretable and trustworthy uncertainty quantification for real-world deployment.