Credal Learning Theory

Statistical learning theory is the foundation of machine learning, providing theoretical bounds for the risk of models learned from a (single) training set, assumed to issue from an unknown probability distribution. In actual deployment, however, the data distribution may (and often does) vary, causing domain adaptation/generalization issues. In this paper we lay the foundations for a `credal' theory of learning, using convex sets of probabilities (credal sets) to model the variability in the data-generating distribution. Such credal sets, we argue, may be inferred from a finite sample of training sets. Bounds are derived for the case of finite hypotheses spaces (both assuming realizability or not), as well as infinite model spaces, which directly generalize classical results.

Paper

References (80)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer SbDN6/10 · confidence 2/52024-07-04

Summary

This paper extended traditional statistical learning theory which usually considers a fixed underlying data-generating distribution to a case where the underlying distribution is assumed to be from a convex set of distributions. Several excess risk bounds (finite realizable, finite unrealizable, and infinite hypothesis space) were derived.

Strengths

- This paper aims to analyze the uncertainty and divination of training/test data distributions in a principled manner, which is very appreciated. - Using convex sets of distributions is a reasonable approach. - This paper considered different assumptions on the hypotheses.

Weaknesses

- The writing can be improved. For example, the first paragraph is basically the notation and a review of ERM. The second paragraph mentioned DA and DG, but it is still unclear what problems existing theories and problem settings have and why it is important to solve them. Then, the author implied that the existing assumptions are unrealistic and too strong so the resulting theories are not generalizable. However, these statements are too vague so it is difficult to verify or falsify them. - Partly due to the writing, I cannot fully understand the proposed theorems, let alone their proofs. For example, I do not understand what exactly a _well-defined_ quantity $\epsilon^\star(\delta)$ in Theorem 4.1 (and the following theorems) is. Therefore, the theoretical implications are very unclear to me, and I cannot accurately assess the significance and novelty of this work.

Questions

Minor issues: - ERM, DA, DG: please provide references. - the position of Figure 1. - l. 74: the citing style "Zhou et al. [63]" seems strange (to me)

Rating

6

Confidence

2

Soundness

3

Presentation

2

Contribution

2

Limitations

The author discussed the limitations briefly in the conclusion section.

Reviewer 2ZTx6/10 · confidence 3/52024-07-10

Summary

This is a paper on learning theory with a focus on machine learning. The authors consider a setup with several training sets available. Given an additional (test) set, the authors assume that the distribution generating these new data coincides with one of the distributions generating the training set or is a convex combination of them. Under a few additional assumptions, the authors can find bounds on the expected loss of the empirical risk minimiser.

Strengths

The paper is quite technical and dense. Yet, the proofs gathered in the appendix are relatively easy to follow and, after a preliminary check, correct. I think the generalised setup considered by the authors is quite interesting and challenging. The author also presents an experimental evaluation of their bounds. This is very valuable for such a kind of theoretical paper.

Weaknesses

The setup considered by the authors might appear a bit special and not very common. Most of the results are a generalisation to the multi-dataset setup of results included in Liang's Lecture Notes. In a sense, the work in this paper might appear as a (straightforward?) extension of those results to the case of multiple datasets. It is not very clear whether, relaxing the realizability assumption (something adding a lot of realism to the modelling) might have a strong impact to the bounds.

Questions

- In which sense are the results in the paper not a simple corollary of those in Liang's LNs? - Is it possible to characterise the impact of the realizability assumption on the bounds? What about considering this case even in the experiments? - Is it possible to advocate the multi-dataset setup considered by the authors better? Are there examples of real tasks that cope with such a situation? I found the experiments very good for the paper, while the discussion about the credal set learning was not so crucial. What about having the latter in the appendix and the former in the main body of the paper?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations discussed by the authors wrt the experiments are very reasonable and I see no problems with them.

Reviewer sJhN5/10 · confidence 3/52024-07-12

Summary

This paper introduces a novel learning framework termed “Credal Learning Theory” which extends the traditional statistical learning theory to handle variations in data distributions, especially the topic “domain generalization”. The authors propose using credal sets, which are convex sets of probability distributions, to model the variability in data generating processes. By leveraging multiple training sets generated from different distributions, the framework infers a credit set that allows the model to capture uncertainty and have guarantees in domain generalization. The paper also provides theoretical bounds on the expected risk of models learned within this credal framework under various conditions, including finite and infinite hypothesis spaces, and both realizable and non-realizable scenarios.

Strengths

This paper introduces the novel concept of Credal Learning Theory to account for data distribution variability using credal sets. This approach presents a new perspective for analyzing the domain generalization field. The authors innovatively apply techniques from Imprecise Probabilities to infer models from limited multiple sample sets, which is practically significant. The definition of credal sets as the convex closure of a family of distributions is also quite natural. Additionally, the paper provides complete mathematical proofs to support the derived generalization bounds. The use of credal sets to model epistemic uncertainty is robust and well-justified, providing theoretical guarantees.

Weaknesses

My main concern is whether the use of the credal set introduced in this paper is sufficient for analyzing domain generalization. As the paper describes, the model is inferred from a finite number of datasets, and we are studying cases where each dataset is sampled from one of a potential family of distributions. Domain generalization aims to provide guarantees for generalization across different distributions within this potential family. The paper provides generalization guarantees for all distributions within the credal set constructed by convex combinations of inferred distributions. However, the gap lies in whether the credal set adequately represents the true potential family of distributions. To convincingly address this, the paper should at least consider the following two aspects: 1. Provide real-world examples demonstrating that the constructed credal set encompasses most of the potential distributions we need to consider in these scenarios, thus illustrating the practical relevance of the credal set. 2. Make assumptions about the potential family of distributions, such as assuming the family follows a certain distribution, and prove that the credal set occupies a significant proportion of the support region with high density for this distribution. This would show that the generalization guarantees given for the credal set are meaningful for real-world distribution scenarios. These two points are essential to demonstrate the universality of the credal set. Otherwise, there may be cases where the distributions of multiple datasets used for inferring the model are very close to each other, making their convex combination very small and only covering a tiny part of the support of the true family of distributions. This would mean there is no guarantee for the true potential family of distributions.

Questions

Please refer to Weaknesses. In addition to the issues mentioned in Weaknesses above, are there any analyses of Computational Complexity that are specific to the use of credal sets as an approach to infer models? Can such an approach be implemented for limited but large datasets?

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have discussed the limitation of this paper in Section 5.

Authorsrebuttal2024-08-06

Last Question

**We report here too our answer to Reviewer sJhN's last question** *Are there any analyses of Computational Complexity that are specific to the use of credal sets as an approach to infer models? Can such an approach be implemented for limited but large datasets?* We thank the reviewer for their question, the answer to which we will try to add to the new version of our manuscript. Yes, there are analyses in the literature of computational complexity specific to the use of credal sets, particularly in the context of graphical models and probabilistic inference. Such approaches can be implemented for large datasets, but they often require approximation techniques to be computationally feasible [1, 2, 3 ]. Despite this, credal set approaches can be implemented for large datasets using techniques like parallel processing, distributed computing, and efficient data structures. Similar to Deep Learning-based approaches, utilization of high-performance computing resources, algorithm optimization, and domain-specific adaptations, the computational challenges can be effectively managed. Recent advancements demonstrate the practicality of these approaches. For instance, "Credal-Set Interval Neural Networks" (CreINNs) have shown significant improvements in inference time over variational Bayesian neural networks [4]. Thus, while the computational demands are comparable to those of deep learning-based methods, the robustness and flexibility of credal sets, as demonstrated in recent research, make them a practical and valuable approach [5, 6]. References: [1] Mauá, Denis Deratani, and Fabio Gagliardi Cozman. "Thirty years of credal networks: Specification, algorithms and complexity." International Journal of Approximate Reasoning 126 (2020): 133-157. [2] Lienen, Julian, and Eyke Hüllermeier. "Credal self-supervised learning." Advances in Neural Information Processing Systems 34 (2021): 14370-14382. [3] Mauá, Denis D., et al. "On the complexity of strong and epistemic credal networks." arXiv preprint arXiv:1309.6845 (2013). [4] Wang, Kaizheng, et al. "CreINNs: Credal-Set Interval Neural Networks for Uncertainty Estimation in Classification Tasks." arXiv preprint arXiv:2401.05043 (2024). [5] Marinescu, Radu, et al. "Credal marginal map." Advances in Neural Information Processing Systems 36 (2024). [6] Wang, Kaizheng, et al. "Credal Wrapper of Model Averaging for Uncertainty Estimation on Out-Of-Distribution Detection." arXiv preprint arXiv:2405.15047 (2024).

Reviewer hWdx6/10 · confidence 2/52024-07-12

Summary

The paper develops a so called credal learning theory that uses convex sets of probability distributions (also known as credal sets) to model the uncertainty of the data-generating distribution. As in the classical statistical learning theory, the paper derives new theoretical bounds on the risk of the models learned from a collection of datasets instead of a single dataset. These datasets do not necessarily correspond to the same distribution. The new results (i.e., bounds) can be viewed as a generalisation of the classical results.

Strengths

- The paper looks at a core problem in machine learning and the credal sets based approach to bounding the expected risk of a machine learning model appears to be quite novel.

Weaknesses

- In my opinion, the presentation is quite dense, the notation appears to be quite heavy in places and the paper is not easy to follow. I think it's important to illustrate some of the key concepts introduced in sections 3 and 4 with some examples. Tables 1 and 2 as well as Figure 2 are good but they seem somewhat disconnected from the rest of the paper.

Questions

- In my understanding, the results derived in the paper rely on the availability of the credal set that is supposed to contain the true data generating process. How is that credal set elicited? - In principle, the credal set can be arbitrarily large in the sense that it can have arbitrarily many extreme points. Is there a bound on its size?

Rating

6

Confidence

2

Soundness

3

Presentation

2

Contribution

3

Limitations

The limitations of the proposed approach are clearly discussed in the paper.

Authorsrebuttal2024-08-09

Thank you!

If our answers contributed to solve the doubts that the reviewer had (which we strongly hope), we would greatly appreciate it they could raise their score. Sincerely, The Authors

Reviewer 2ZTx2024-08-12

I thank the reviewers for their detailed feedback and for the additional work. I am happy to confirm my initial positive opinion about the paper.

Reviewer SbDN2024-08-12

Thank you for your detailed reply. I've read other reviews and responses. I'm now more positive about this paper but still not so sure about my understanding, so I raised my score while keeping my confidence. My previous rating was mainly due to my insufficient understanding of this work. I appreciate the author's explanation of the problem setting, their technical contributions, and the practical relevance (e.g., the availability of the credal set, continual learning, distribution misspecification). I do not doubt that it is difficult to explain such a topic clearly to every reader in 9 pages, as I also faced similar issues before. The current presentation may be sufficient for some researchers, but as someone who only vaguely remembers those important results in statistical learning theory, I have to agree with `Reviewer hWdx` and `Reviewer 2ZTx` that this paper is quite dense. Maybe a 10-page conference paper is not the best form to present this theory. I'm looking forward to the author's other expositions of this theory (journal papers, lectures, tutorials, etc.), if possible, so that I may be able to use it in other settings. Regarding the citing style, do not worry and please ignore it. It's just a personal preference.

Authorsrebuttal2024-08-12

Thank you!

We thank the reviewer for the time they spent understanding more deeply our paper, for understanding our struggles with the page limit, and for raising their score. As they correctly suggest, we are currently preparing a journal version where we also plan to extend our results, and we are thrilled about their interest in possibly using our findings in their research. Once again, we thank the reviewer for their comments, The Authors

Reviewer hWdx2024-08-12

Thanks for the clarifications.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC