Trustworthy Deep Learning via Proper Calibration Errors: A Unifying Approach for Quantifying the Reliability of Predictive Uncertainty

Sebastian Gruber, Florian Buettner

Abstract

With model trustworthiness being crucial for sensitive real-world applications, practitioners are putting more and more focus on improving the uncertainty calibration of deep neural networks. Calibration errors are designed to quantify the reliability of probabilistic predictions but their estimators are usually biased and inconsistent. In this work, we introduce the framework of proper calibration errors, which relates every calibration error to a proper score and provides a respective upper bound with optimal estimation properties. This relationship can be used to reliably quantify the model calibration improvement. We theoretically and empirically demonstrate the shortcomings of commonly used estimators compared to our approach. Due to the wide applicability of proper scores, this gives a natural extension of recalibration beyond classification.

Introduction

Deep neural networks are notoriously overconfident; a model might predict an event with 99% probability but be correct only 80% of the time. To address this, researchers use calibration errors to measure the discrepancy between predicted probabilities and actual frequencies. The industry standard, Expected Calibration Error (ECE), has a major flaw: its estimators are biased and inconsistent. The ECE value depends heavily on test set size and binning strategies, making it difficult to compare models or verify recalibration techniques, especially in "small data" regimes like medical diagnostics.

Proper Calibration Errors and the RBS Metric

This work bridges the gap between proper scores (metrics like Brier score or Log-loss, minimized when a model predicts the true distribution) and calibration errors. The authors provide a theoretical hierarchy showing that most common metrics (ECE, MMCE, KS) are lower bounds of the true calibration error.

They define a new class of "proper calibration errors" derived from the divergence of proper scores. Since true calibration errors are hard to estimate without bias, they propose using an upper bound. For classification, they introduce the Root Calibration Upper Bound (RBS), which is derived from the Brier score. Mathematically, they prove that ECE estimators are strictly convex and monotonically decreasing with respect to data size, explaining why ECE artificially improves as more test samples are added.

Empirical Evaluation

The framework was validated across standard computer vision benchmarks (CIFAR-10/100, ImageNet) and regression datasets.

Robustness to Data Size

A key finding is the stability of the RBS estimator. As shown in Figure 4, while binned estimators like ECE exhibit significant relative bias and drop-offs as test set size decreases, RBS remains consistent. This makes RBS a far more reliable metric for small datasets.

Figure 4: Left: Calibration error estimators vs. test set size. Right: Relative bias of estimators. RBS (red) remains stable and unbiased compared to binned methods.

Tracking Calibration Improvement

When evaluating recalibration methods like Temperature Scaling (TS), RBS provides a robust measure of improvement. Traditional metrics often overestimate or underestimate the effect of recalibration on small samples due to estimation artifacts. Figure 1 illustrates how RBS consistently tracks calibration improvement across different test set sizes, whereas other estimators fluctuate wildly.

Figure 1: Calibration improvement tracked by various estimators across different test set sizes. RBS (red) provides a stable estimate of improvement.

Conclusion

The proposed framework of proper calibration errors addresses the fundamental biases in current uncertainty quantification metrics. By leveraging the properties of proper scores, the RBS metric offers a stable and consistent alternative to ECE. Furthermore, because proper scores exist for various distributions, this approach naturally extends calibration assessment to regression and other complex tasks, providing a unifying tool for building more trustworthy AI systems.

Summary updated

Abstract

With model trustworthiness being crucial for sensitive real-world applications, practitioners are putting more and more focus on improving the uncertainty calibration of deep neural networks. Calibration errors are designed to quantify the reliability of probabilistic predictions but their estimators are usually biased and inconsistent. In this work, we introduce the framework of proper calibration errors, which relates every calibration error to a proper score and provides a respective upper bound with optimal estimation properties. This relationship can be used to reliably quantify the model calibration improvement. We theoretically and empirically demonstrate the shortcomings of commonly used estimators compared to our approach. Due to the wide applicability of proper scores, this gives a natural extension of recalibration beyond classification.
BibTeX
@inproceedings{gruber2022trustworthy,
  author = {Sebastian Gruber and Florian Buettner},
  title = {{Trustworthy Deep Learning via Proper Calibration Errors: A Unifying Approach for Quantifying the Reliability of Predictive Uncertainty}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2022},
  url = {https://openreview.net/forum?id=BWhRs3kpHGP},
}
Figure