Understanding uncertainty in deep neural networks is essential for building trustworthy machine learning systems, especially in safety-critical applications where incorrect predictions carry high costs. Traditional uncertainty quantification methods offer scalar measures of model confidence, but they provide limited insight into which spatial regions of an input contribute to different types of uncertainty. A new research paper from a team of computer scientists introduces a visualization framework that addresses this gap by generating spatial uncertainty activation maps.
The Problem with Current Uncertainty Quantification
Existing methods typically output a single number representing model confidence, such as probability scores or entropy. According to the paper, this approach fails to indicate whether uncertainty stems from missing evidence or from conflicting evidence between competing hypotheses. In safety-critical domains, knowing the source of uncertainty can guide further data collection or model refinement. Without spatial context, developers cannot pinpoint which parts of an input—such as specific regions in an image—cause the model to be uncertain.
Introducing Uncertainty Activation Maps (UAM)
The proposed framework, termed Uncertainty Activation Map (UAM), combines two established techniques: Evidential Deep Learning (EDL) and Full-Gradient Class Activation Mapping (FullGrad). EDL provides principled uncertainty quantification by treating model outputs as evidence collected for each class. FullGrad offers complete gradient decomposition, enabling attribution of predictions to input features. The paper explains that by leveraging the complete gradient decomposition property of FullGrad and the principled uncertainty quantification of Subjective Logic, the method produces theoretically grounded visualizations that highlight specific image regions responsible for model uncertainty.
Two Types of Uncertainty: Vacuity and Dissonance
UAM distinguishes between two fundamental uncertainty types:
| Uncertainty Type | Definition | Implication |
|---|---|---|
| Vacuity | Lack of evidence | The model has seen insufficient data to make a confident prediction – areas where more training examples are needed. |
| Dissonance | Conflicting evidence between competing hypotheses | The model is uncertain because different pieces of evidence support different outcomes – areas where ambiguity exists. |
"Our approach distinguishes between two fundamental types of uncertainty: vacuity, representing lack of evidence, and dissonance, capturing conflicting evidence between competing hypotheses."
The authors report that vacuity and dissonance activation maps are generated by computing belief-weighted attributions, enabling identification of where models lack knowledge versus where they encounter ambiguous evidence.
How It Works: Combining EDL and FullGrad
The Uncertainty Activation Map framework works by first training a neural network using evidential deep learning, which outputs evidence values for each class. Using FullGrad, the method computes gradient-based attributions with respect to the evidence values. Then, through belief-weighted aggregation, separate spatial maps are produced for vacuity and dissonance. The paper describes this as a theoretically grounded visualization that provides intuitive visual feedback to assess model reliability in complex visual recognition tasks.
Validation and Benchmark Performance
According to the research, extensive evaluations across multiple benchmark datasets demonstrate that the UAM framework effectively addresses the critical gap between uncertainty quantification and explainability. While the paper does not specify which datasets, it claims that the framework produces interpretable spatial uncertainty maps that help developers understand model behavior at a granular level.
Implications for Enterprise AI
For technology leaders deploying deep learning in production, the UAM framework offers a practical tool for auditing model reliability. By revealing whether a model is uncertain due to lack of training data (vacuity) or conflicting signals (dissonance), developers can prioritize targeted data collection or adjust model architecture. The spatial nature of the maps also helps identify systematic weaknesses—for example, a model that is consistently vacuous on certain object textures or lighting conditions.
While the research is currently academic, its application to safety-critical domains such as autonomous systems, medical imaging, and industrial quality control is direct. Enterprises investing in AI for these areas should monitor developments in uncertainty visualization to ensure their models can be trusted when decisions matter most.
The paper was authored by Jeong, Dong Hyun, Chen, Feng, Cho, Jin-Hee, Kaplan, Lance M., Jøsang, Audun, and Soo-Yeon, and is available on arXiv under the title "Visualizing Uncertainty: Spatial Maps of Missing and Conflicting Evidence in Deep Learning."
