Information-theoretic Generalization Analysis for Expected Calibration Error

While the expected calibration error (ECE), which employs binning, is widely adopted to evaluate the calibration performance of machine learning models, theoretical understanding of its estimation bias is limited. In this paper, we present the first comprehensive analysis of the estimation bias in the two common binning strategies, uniform mass and uniform width binning. Our analysis establishes upper bounds on the bias, achieving an improved convergence rate. Moreover, our bounds reveal, for the first time, the optimal number of bins to minimize the estimation bias. We further extend our bias analysis to generalization error analysis based on the information-theoretic approach, deriving upper bounds that enable the numerical evaluation of how small the ECE is for unknown data. Experiments using deep learning models show that our bounds are nonvacuous thanks to this information-theoretic generalization analysis approach.

Paper

References (42)

Scroll for more · 30 remaining

Similar papers

Peer review

Reviewer kktU5/10 · confidence 4/52024-07-10

Summary

This paper analyzes the estimation bias and generalization error of the expected calibration error (ECE). Specifically, in a binary classification setting, the authors provide an upper bound for the total bias with an improved convergence rate, applicable to both uniform mass and uniform width binning strategies. They also determine the optimal number of bins to minimize the total bias. Furthermore, the authors utilize the information-theoretic generalization framework, particularly the Conditional Mutual Information (CMI) framework, to characterize the generalization of ECE.

Strengths

1. This paper achieves a tighter bound for total bias compared to previous works. 2. The optimal number of bins is determined using the upper bound of the total bias.

Weaknesses

1. As the authors themselves note, a significant limitation is that the analysis in this work is only applicable to binary classification. 2. Some assumptions (e.g., Assumption 2) are not well justified. 3. The writing has significant room for improvement; several arguments are unclear or misleading. Please find more details in the questions below.

Questions

1. Does Assumption 2 hold true in practice? Is there a way to verify it? Additionally, what is the motivation behind assuming $n_e\geq 2B$? How is this assumption utilized? If $n_{te}\leq 2B$, will Theorems 2 and 3 still be valid? 2. In the proof sketch of Theorem 2, you mention that $\mathrm{ECE}$ and $\mathrm{TCE}$ could be re-written. While they seem correct to me, could you elaborate on how $\mathrm{TCE}(f_\mathcal{I})$ is obtained in the form shown in Line 152? I did not find the details in the complete proof. 3. According to Theorem 5, do the upper bounds indicate that UWB is a better binning strategy than UMB, given that UMB has an additional $\mathrm{fCMI}$ bias term? It seems that only UMB's expected binning bias is sensitive to the training data, which might be seen as a disadvantage in terms of the upper bound. 4. The writing can be significantly improved. For example, in Line 244, you mention "Our theory might guarantee the ECE under test dataset for them." Do you mean your theory might guarantee low ECE under the test dataset? Additionally, in Lines 251-252, "This implies that if the model generalizes well, evaluating the ECE using the training dataset may better reduce the total bias than that using test dataset." Why does evaluating ECE reduce total bias? What we really care about is ECE on unseen/test data. How does evaluating ECE on training data affect this purpose? 5. Why is the metric entropy method only used for UWB? It seems that you upper bound $\mathrm{eCMI}$ by the $\mathrm{fCMI}$ term first in your proof. What prevents you from giving a similar result for UMB? 6. In Line 339-340, you mention that "a notable trend towards acquiring relatively stable nonvacuous bounds can be observed when adopting $B =\lfloor n^{1/3} \rfloor$", but according to Figure 1, it seems $B=52$ is tighter than $B=B =\lfloor n^{1/3} \rfloor$ in most cases. Could you clarify this? 7. Since $\mathrm{eCMI}$ and $\mathrm{fCMI}$ terms are key components in both standard generalization error and expected total bias of calibration error, do you have any new insights into the relationship between calibration and generalization from this perspective?

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

The authors have thoroughly discussed some limitations of this work.

Reviewer 8W416/10 · confidence 4/52024-07-11

Summary

This paper investigates the estimation bias in expected calibration error (ECE) for binary classification models, focusing on uniform mass binning (UMB) and uniform width binning (UWB). The authors present a comprehensive theoretical analysis, establishing upper bounds for the bias and the generalization error. Based on the convergence rates of binning and statistical bias, they identify the optimal number of bins to minimize the total estimation bias.

Strengths

* The paper provides a comprehensive analysis of the estimation bias in ECE, providing upper bounds and optimal bin size choices for both UWB and UMB. * The authors further derive upper bounds for the generalization error between ECE and TCE using an information-theoretic approach. * Numerical experiments on deep learning tasks confirm that the derived bounds are non-vacuous.

Weaknesses

* The provided results only apply to binary classification, and require Lipschitz continuity which may not be necessarily satisfied in deep learning models. Also, these bounds are analyzing the ECE using test data but not training data, making them less applicable since test data are not always available in practice. * The convergence rates of the information-theoretic generalization bounds heavily depend on the actual rate of eCMI and fCMI measures, which are not directly clear in analysis. In theorem 6, the authors show that eCMI scales as O(log n) based on metric entropy, but this bound involves the dimensionality d, and is thus hardly applicable to deep learning models. * For experimental results, only the statistical bias is evaluated but not the total generalization error. It is also hard to see to what extent these bounds are tight in the current results. These bounds are also hard to estimate due to the existence of eCMI or fCMI measures. I would suggest the authors additionally consider some synthetic settings where TCE, eCMI, and fCMI are analytically tractable to show the tightness of the bounds. (maybe Gaussian data points?)

Questions

Recent information-theoretic bounds have shown improved rates of O(1/n) under the interpolating regime, and also direct computational tractability with loss CMI or entropy metrics. It may be worth some discussions on whether these techniques can be adopted to acquire tighter bounds. Tighter Information-Theoretic Generalization Bounds from Supersamples. ICML 2023. Rethinking Information-theoretic Generalization: Loss Entropy Induced PAC Bounds. ICLR 2024.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Authorsrebuttal2024-08-10

Thank your for the reply

Thank you for confirming our response. We believe that the additional experiments address your concerns. If there are any further issues or concerns that need to be addressed to improve the score, please let us know. We are happy to discuss them further.

Reviewer Ckaf6/10 · confidence 4/52024-07-13

Summary

The paper studies the expected calibration error using information-theoretical tools. They derive different tight fCMI and eCMI bounds in this setting. Empirical results show that the results are nonvacuous.

Strengths

1/ The paper is in general well written. Adequate discussions are given in the main body of the paper and the appendices. 2/ The paper provides the first information-theoretic comprehensive analysis of the bias associated with the ECE when using the test and training datasets. 3/ The theoretical results seem sound. I skimmed through most of the proofs (I did not go through all of them in detail) but the proofs are well-structured and easy to follow. 3/ Empirical results show that the bound is tight for deep learning models.

Weaknesses

The only weakness, if any, is perhaps that the paper uses conventional machinery for deriving information-theoretic generalization bounds and that it has not developed novel proof techniques.

Questions

Besides fCMI and eCMI based bounds, is it possible to extend the analysis and derive $\Delta$-L based bounds [37]? These bounds are typically tighter compared to fCMI and eCMI.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The limitations are adequately addressed in the conclusion.

Reviewer 9r8F7/10 · confidence 4/52024-07-15

Summary

This paper presents a comprehensive analysis of the estimation bias for expected calibration error (ECE), focusing on two common binning strategies: uniform mass and uniform width binning. The analysis establishes upper bounds on the bias, resulting in an improved convergence rate. Furthermore, these bounds reveal the optimal number of bins needed to minimize the estimation bias. The study also extends the bias analysis to generalization error analysis using an information-theoretic approach, deriving upper bounds that facilitate numerical evaluation for recalibration methods based on training data. Experiments with deep learning models demonstrate that the bounds are nonvacuous, due to the information-theoretic generalization analysis approach.

Strengths

As the author pointed out, the existing literature lacks a theoretical analysis of the estimated ECE and a more principled approach to estimation. This paper addresses and closes this gap.

Weaknesses

1. Tightness issue of the upper bound in Corollary 1. It is commendable that the authors included a discussion on the tightness of Equation 12. However, it would be more rigorous to formally establish a minimax lower bound for the estimation bias that applies to all types of estimators. The authors could either use existing results from Tsybakov [33] or construct a worst-case analysis using Le Cam’s method to establish the lower bound. While it is acceptable if the constant does not match the upper bound, it is crucial to demonstrate the rate. 2. A drawback of information-theoretic (IT) bounds is the implicit dependency on the algorithm. For example, Theorem 7 appears very similar to Theorem 4, as the recalibration-induced dependence is encapsulated in the CMI term. The authors should provide more commentary on this aspect and clarify the connection between Theorems 6 and 5, as well as which bound is more practical for use. 3. In the caption of Figure 1, It is said that the ECE gap does not change significantly in B. How can we justify that the selection of $B = n^{1/3}$ is better? Figure 1 primarily plots the bound in (14), but as I mentioned earlier, such a bound can be very loose, and more empirical justification should be provided for the selection of optimal B.

Questions

1. Clarification: what is the ECE gap plotted in Figure 1 and Table 1? Estimated ECE? To my understanding, all the bounds in Figure 1 are quite loose. Shouldn't we plot the left-hand side of Equation 14 for a more accurate comparison? 2. How should the error bars for the bound values in Table 1 be interpreted? What is the source of the randomness? 3. It is not accurate to say that I(S;W)=O(log n) in Line 258, as such Barron’s result assume that samples Z are conditional i.i.d. given model parameter w. However, in the learning context, we always assume that training samples are i.i.d.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The Limitations are well addressed in section 7.

Reviewer 8W412024-08-09

Thank you for the response and additional experiments. I'll maintain my positive rating.

Reviewer Ckaf2024-08-10

Thank you for the response. I'll maintain my positive score.

Authorsrebuttal2024-08-10

Thank you for the comment.

Thank you for confirming our response. If there are any further issues or concerns that need to be addressed to improve the score, please let us know. We are happy to discuss them further.

Reviewer 9r8F2024-08-12

I appreciate the authors' effort in preparing the response. It thoroughly addresses all my comments, and I am pleased to see the matched lower-bound results and additional empirical evidence supporting the claim. I will raise my score to 7—this is a solid paper. Please ensure that the new results are incorporated into the final revision, and make sure the figures in the main body are clear and easy to follow.

Authorsrebuttal2024-08-12

Acknowledgement

We greatly appreciate your acknowledgment of our responses. We will incorporate the theoretical facts and insights, along with their proofs, discussed during the rebuttal period into the revised version. Additionally, we will improve the clarity of the existing experiments and include the additional experiments in the revised version as well. We are delighted that you have decided to increase your score. Thank you again for your insightful feedback.

Reviewer kktU2024-08-12

Thank you for the responses

I would like to thank the authors for their responses and apologize for not engaging in the discussion earlier. I have read all the reviews and the corresponding author responses. My concern regarding Assumption 2 has been adequately addressed, and further clarifications have been provided for other questions. I will increase my score to 5 accordingly. However, I am not assigning a higher score because the analysis in this paper is limited to binary classification, with no clear path to extending it to a multi-classification setting. While I also appreciate the minimax results, I am concerned that the analysis (and potentially the assumptions) might require significant refinement in more general or practical settings.

Authorsrebuttal2024-08-12

Acknowledgement and clarification of the limitation of the binary classification

We sincerely appreciate your thorough review of our responses to the reviewers’ comments. And, we are delighted that you have decided to increase your score. Concerning the limitations inherent in binary classification, we addressed similar points in our response to Reviewer 8W41 (Q.1). Please refer to that section for details. In summary, our analysis can extend to multiclass scenarios involving top-calibration error, such as when using softmax for classification. In these cases, evaluating the calibration of the highest softmax score is standard practice, and these metrics can be reduced to a formulation similar to binary classification, enabling our analysis. Concerning the minimax theorem, additional assumptions are necessary because the theorem deals with worst-case distributions. Worst-case scenarios may occasionally involve settings that are unrealistic or unlikely. However, for establishing upper bounds, as in Theorem 5, Assumption 2 is generally sufficient, making it more practical and likely to be met in real-world scenarios. We hope these answers will further clarify your understanding.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC