On Information-Theoretic Measures of Predictive Uncertainty

Kajetan Schweighofer, Lukas Aichberger, Mykyta Ielanskyi, Sepp Hochreiter

Introduction

In high-stakes machine learning applications such as autonomous driving or medical diagnostics, a model's ability to quantify its own uncertainty is as critical as the accuracy of its predictions. While the entropy of the posterior predictive distribution is often treated as the default metric, it frequently fails in complex scenarios and lacks a rigorous theoretical justification. This paper addresses this gap by introducing a unifying framework that categorizes information-theoretic measures of predictive uncertainty, showing that most existing metrics are different approximations of the same fundamental quantity: the cross-entropy between the predicting model and the true data-generating distribution.

The 3x3 Framework

The authors propose a structured 3x3 framework to classify uncertainty measures based on two primary axes:

  1. The Predicting Model (I): This axis identifies whether the prediction is generated by a single pre-selected model (A), a Bayesian model average (B), or a model sampled anew from the posterior for each prediction (C).
  2. Approximation of the True Distribution (II): This axis defines how the unknown "true" distribution is approximated—using a single reference model (1), the average of the posterior (2), or by considering the entire ensemble of models in the posterior (3).

By crossing these factors, the authors derive nine distinct measures of Total Uncertainty (TU), each of which can be decomposed into Aleatoric Uncertainty (inherent data noise) and Epistemic Uncertainty (model-related lack of knowledge).

The Alignment Principle

A key finding of this research is the Alignment Principle: to achieve optimal performance in tasks like selective prediction (deciding when to abstain from a forecast), the uncertainty measure must be aligned with the specific model used for prediction. For instance, when using a single model to predict, a measure tailored to that single model (such as TU (A3)) significantly outperforms standard Bayesian average measures.

Performance comparison of uncertainty measures across different prediction strategies (Single Model, Average Model, and According to Posterior). The results highlight the importance of aligning the measure with the predicting model.

Challenging the Epistemic Uncertainty Dogma

A common heuristic in machine learning is that Epistemic Uncertainty (EU) is the most effective tool for detecting out-of-distribution (OOD) data. However, this study reveals a surprising counter-trend: Total Uncertainty (TU) and Aleatoric Uncertainty (AU) often perform significantly better for OOD detection. While AU and EU are highly correlated on clean, in-distribution data, they become distinct on OOD data, suggesting that the value of disentangling these two types of uncertainty is highly dependent on the data the model is processing.

Experimental Results and Robustness

The researchers evaluated their framework across standard image classification benchmarks, including CIFAR-10, CIFAR-100, SVHN, and Tiny-ImageNet. The results were consistent across various neural architectures, such as ResNet-18 and DenseNet-169:

  • Selective Prediction: For single-model predictions, TU (A3) achieved the highest performance. When using an average model, TU (B/C3) was superior, reaching AUROC scores above 90% in several configurations.
  • OOD Detection: Measures like TU (B/C2) and TU (B/C3) consistently outperformed EU-based measures, which were often among the worst performers for distinguishing between in-distribution and OOD samples.
  • Active Learning: While standard mutual information (EU (C2)) was effective for simple datasets like MNIST, simpler aleatoric measures proved more robust for complex datasets like Fashion-MNIST.

Conclusion and Future Work

This paper provides a principled foundation for choosing uncertainty measures, moving beyond heuristics toward a theoretically grounded approach. While the current framework focuses on classification, future work aims to extend these insights to regression and autoregressive models. Applying this framework to Large Language Models (LLMs) could be particularly impactful for distinguishing between factual errors and "confabulations" in sequence generation.

Summary updated

BibTeX
@inproceedings{schweighofer2025information,
  author = {Kajetan Schweighofer and Lukas Aichberger and Mykyta Ielanskyi and Sepp Hochreiter},
  title = {{On Information-Theoretic Measures of Predictive Uncertainty}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=hLv7yZ268J},
}
Figure