Evidential Deep Learning

Follow this topic Ask about this topic

Evidential deep learning trains neural networks to predict evidence as well as class probabilities, commonly using Dirichlet distributions for classification and Normal–Inverse-Gamma distributions for regression. Methods regularize evidence growth, derive objectives from Bayesian or variational principles, fuse evidence across modalities, or add post-hoc uncertainty heads; estimates can then guide abstention, data acquisition, and feature updates. The field is applied to settings such as medical imaging, autonomous sensing, object detection, and out-of-distribution recognition, with evaluation often measuring calibration, error detection, and risk as well as accuracy. A central open tension is whether evidence-based uncertainty is genuinely meaningful and reliable under distribution shift: evidence can be miscalibrated or poorly identified, and metrics such as vacuity can be confounded by differences in class structure.

Latest papers

61 papers9 papers in the last four weeks, up 125% on the four weeks before. 0.1% of all new papers.

Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning
Feb 1, 2026Pietro Carlotti, Nevena Gligić, Arya FarahiEvidential Deep Learning OOD Detection
Department of Statistics and Data Sciences University of Texas at Austin

Feb 1, 2026J/K move · Enter open · S save

Density-Informed Pseudo-Counts for Calibrated Evidential Deep Learning

Pietro Carlotti, Nevena Gligić, Arya Farahi

Department of Statistics and Data Sciences University of Texas at Austin

Evidential Deep Learning OOD Detection Density Estimation Uncertainty Calibration Epistemic Uncertainty

Open paper AskPDF arXiv

Evidential Deep Learning (EDL) is a popular framework for uncertainty-aware classification that models predictive uncertainty via Dirichlet distributions parameterized by neural networks. Despite its popularity, its theoretical foundations and behavior under distributional shift remain poorly understood. In this work, we provide a principled statistical interpretation by proving that EDL training corresponds to amortized variational inference in a hierarchical Bayesian model with a tempered pseudo-likelihood. This perspective reveals a major drawback: standard EDL conflates epistemic and aleatoric uncertainty, leading to systematic overconfidence on out-of-distribution (OOD) inputs. To address this, we introduce Density-Informed Pseudo-count EDL (DIP-EDL), a new parametrization that decouples class prediction from the magnitude of uncertainty by separately estimating the conditional label distribution and the marginal covariate density. This separation preserves evidence in high-density regions while shrinking predictions toward a uniform prior for OOD data. Theoretically, we prove that DIP-EDL achieves asymptotic concentration. Empirically, we show that our method enhances interpretability and improves robustness and uncertainty calibration under distributional shift.