Visualizing Uncertainty: Spatial Maps of Missing and Conflicting Evidence in Deep Learning

Researchers have proposed a visualization framework called Uncertainty Activation Map (UAM) that pinpoints the spatial regions in an image where a deep neural network lacks evidence or encounters conflicting information, according to a paper submitted to arXiv on June 14, 2026 [1][2]. The framework, detailed in a preprint by Dong Hyun Jeong, combines Evidential Deep Learning (EDL) with Full-Gradient Class Activation Mapping (FullGrad) to produce interpretable spatial maps of model uncertainty [1][2]. The method distinguishes between two types of uncertainty: vacuity, which represents a lack of evidence, and dissonance, which captures conflicting evidence between competing hypotheses [1][2]. By computing belief-weighted attributions, the UAM generates separate activation maps for each uncertainty type, enabling identification of where models lack knowledge versus where they encounter ambiguous evidence [2]. The approach leverages the complete gradient decomposition property of FullGrad and the uncertainty quantification framework of Subjective Logic to produce theoretically grounded visualizations [2]. The authors argue that existing uncertainty quantification methods provide only scalar measures of model confidence, offering limited insight into which spatial regions of an input contribute to different types of uncertainty [1][2]. The paper states that extensive evaluations across multiple benchmark datasets demonstrate the framework's effectiveness in addressing the gap between uncertainty quantification and explainability [1][2]. The submission file is listed at 17,048 KB [1]. The work arrives as the arXiv repository continues to expand. As of November 2024, the submission rate was about 24,000 articles per month, and the repository surpassed two million articles by the end of 2021 [9]. arXiv, an open-access repository of electronic preprints, has been operating since August 14, 1991, and covers fields including computer science, mathematics, and physics [9]. The paper was posted under the Machine Learning category within Computer Science [1]. The preprint is accessible through arXiv's abstract page, which also features experimental community tools developed through the arXivLabs framework, including the Bibliographic Explorer and CORE Recommender [7][8]. arXivLabs, launched in 2020, provides a formalized framework for collaborations that add value for readers and authors while adhering to arXiv's values of openness, community, excellence, and user data privacy [7].