Uncertainty Gating for Cost-Aware Explainable Artificial Intelligence
Summary: arXiv:2603.29915v1 Announce Type: new
Abstract
Post-hoc explanation methods are widely used to interpret black-box predictions, but their generation is often computationally expensive and their reliability is not guaranteed. We propose epistemic uncertainty as a low-cost proxy for explanation reliability: high epistemic uncertainty identifies regions where the decision boundary is poorly defined and where explanations become unstable and unfaithful. This insight enables two complementary use cases: ‘improving worst-case explanations’ (routing samples to cheap or expensive XAI methods based on expected explanation reliability), and ‘recalling high-quality explanations’ (deferring explanation generation for uncertain samples under constrained budget).
Key Findings
Across four tabular datasets, five diverse architectures, and four XAI methods, we observe a strong negative correlation between epistemic uncertainty and explanation stability. Further analysis shows that epistemic uncertainty distinguishes not only stable from unstable explanations, but also faithful from unfaithful ones. Experiments on image classification confirm that our findings generalize beyond tabular data.
Introduction
The increasing reliance on artificial intelligence in various sectors has necessitated the development of methods that can explain AI decisions. Traditional post-hoc explanation methods, while useful, often come with significant computational costs and variable reliability. These limitations pose challenges in practical applications where both accuracy and efficiency are paramount.
Understanding Epistemic Uncertainty
Epistemic uncertainty refers to the uncertainty in a model’s predictions due to a lack of knowledge. By leveraging this type of uncertainty, we can gain insights into the reliability of explanations generated by AI systems. High levels of epistemic uncertainty signal areas where the model’s confidence is low, indicating potential instability in the explanations provided.
Proposed Methodology
- Improving Worst-Case Explanations: By routing samples based on their epistemic uncertainty, we can allocate computational resources more efficiently. This means directing simpler, cost-effective explanation methods to uncertain cases and reserving more complex methods for cases with higher confidence.
- Recalling High-Quality Explanations: For uncertain samples, we propose deferring the generation of explanations until more resources are available, ensuring that only reliable explanations are produced within budget constraints.
Experimental Results
Our experiments encompassed four distinct tabular datasets, utilizing five different machine learning architectures and four explanation methods. The results consistently demonstrated a negative correlation between epistemic uncertainty and the stability of explanations. Additionally, our analysis indicated that high epistemic uncertainty often coincided with unfaithful explanations, further affirming the utility of this approach.
Generalization to Image Classification
To validate the robustness of our findings, we extended our analysis to image classification tasks. The results confirmed that the relationship between epistemic uncertainty and explanation reliability holds true beyond tabular data, indicating a broader applicability of our proposed methodology.
Conclusion
In conclusion, our research introduces a novel approach to enhance the reliability and efficiency of explainable AI methods. By incorporating epistemic uncertainty as a guiding principle, we can not only improve the quality of explanations but also optimize resource allocation in AI systems. This advancement has significant implications for the deployment of AI in critical areas where understanding decision-making processes is essential.