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

The article presents the Uncertainty Activation Map (UAM), a novel visualization framework that integrates Evidential Deep Learning (EDL) with Full-Gradient Class Activation Mapping (FullGrad) to produce spatial uncertainty maps in deep learning models. UAM distinguishes between vacuity (lack of evidence) and dissonance (conflicting evidence), offering interpretable visualizations that indicate specific regions contributing to model uncertainty. This framework enhances the understanding of uncertainty in deep learning, which is essential for developing reliable AI systems in safety-critical applications, by bridging the gap between uncertainty quantification and explainability through belief-weighted attributions.