Credal Concept Bottleneck Models: Structural Separation of Epistemic and Aleatoric Uncertainty

Tanmoy Mukherjee, Marius Kloft, Pierre Marquis, Zied Bouraoui · 2026

Read Credal Concept Bottleneck Models: Structural Separation of Epistemic and Aleatoric Uncertainty by Tanmoy Mukherjee, Marius Kloft, Pierre Marquis, Zied Bouraoui (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-11219.txt.

Description

Doubt is often split into reducible and irreducible parts, yet both splits come from one spread and stay correlated. Tanmoy Mukherjee and three co-authors propose structural separation instead: knowledge and aleatoric doubt get disjoint setting paths and distinct supervision targets.

They build the idea into a credal concept bottleneck model and a credal self-explaining network, and prove that the two doubt heads avoid coupling through shared gradients. Across five ambiguity-aware benchmarks, separation cuts knowledge-aleatoric link while keeping predictive results. Aleatoric scores track annotator ambiguity, while knowledge scores track error and data scarcity. The result is doubt that is operationally distinct, not just decorrelated.

Catalog record

CREATORTanmoy Mukherjee, Marius Kloft, Pierre Marquis, Zied Bouraoui
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE84 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-11219

Browse the full library · About the Tower of Babel