Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning

Pietro Carlotti, Nevena Gligić, Arya Farahi · 2026

Read Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning by Pietro Carlotti, Nevena Gligić, Arya Farahi (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-2602-01477.txt.

Description

Pietro Carlotti and co-authors give a new reading of evidential deep learning. They prove its training amounts to amortized variational inference inside a hierarchical Bayesian model built on pseudo-likelihood. That lens exposes a flaw: standard evidential learning mixes epistemic and aleatoric uncertainty, breeding steady overconfidence on OOD inputs. Their fix, Density-Informed Pseudo-count EDL, splits prediction from uncertainty size. It gauges the conditional label distribution apart from the marginal input density. The split keeps evidence in high-density regions while pulling OOD predictions toward a uniform prior.

Catalog record

CREATORPietro Carlotti, Nevena Gligić, Arya Farahi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE78572
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602.01477

Browse the full library · About the Tower of Babel