Reliable learning in challenging environments

The problem of designing learners that provide guarantees that their predictions are provably correct is of increasing importance in machine learning. However, learning theoretic guarantees have only been considered in very specific settings. In this work, we consider the design and analysis of reliable learners in challenging test-time environments as encountered in modern machine learning problems: namely `adversarial' test-time attacks (in several variations) and `natural' distribution shifts. In this work, we provide a reliable learner with provably optimal guarantees in such settings. We discuss computationally feasible implementations of the learner and further show that our algorithm achieves strong positive performance guarantees on several natural examples: for example, linear separators under log-concave distributions or smooth boundary classifiers under smooth probability distributions.

Paper

References (94)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer nCeL5/10 · confidence 3/52023-06-30

Summary

The paper develops learning methods that provide theoretical guarantees for point-wise predictions in challenging scenarios at test-time. In particular, the methods presented address situations affected by adversarial attacks and distribution shifts. The paper provides a theoretical contribution to a relevant line of work on reliable learning.

Strengths

The paper presents a significant theoretical contribution on reliable learning, addressing important scenarios such as those affected by adversarial attacks and distribution shifts. In addition, the paper covers several different notions of losses and reliability. The results presented can lead to a better understanding of the possibilities and limitations of such scenarios.

Weaknesses

The main limitation I can see in the paper is the reliance on the realizable case. I guess that reliance cannot be easily avoided but it may be worth to further discuss and emphasize that issue. For instance, the guarantees in the paper are contingent to the fact that we are in a realizable setting, which is not possible to know in most practical cases. Such assumption is even more significant for the distribution shift case, since it constraints the possible shifts. The paper contains many theoretical results for multiple losses, scenarios, etc. In general, it is good to have many interesting results but the paper is quite dense, and it is difficult to get the main ideas. For instance, in Section 4, the authors mention "The optimal robustly-reliable learner described above may be implemented" but the authors hardly describe such learner before Section 4 besides stating that it exists (it is described in the proofs). The paper would be significantly more readable (at least for readers not very familiar with the topic) if the authors first describe the main ideas and results over the simplest scenario and then generalize them to other cases. It is not clear that the probabilities lower bounded in Theorem 3.3. are not very small or that the abstention probabilities are not very high. I think this point is important because the learners would not be very useful in other cases. It would be good if the authors can quantify numerically those probabilities. The bounds provided show that the probabilities increase when the number of samples increases and the dimension decreases but it is not clear in what scenarios those probabilities are sizable or even the lower bound is larger than 0.

Questions

Why is that the robustly-reliable region does not contain the instances x such that r(x)=0? I guess there is a typo in Theorem 3.1 and the subscript CA in "RR^\set{L}_CA(S,\eta)" should be omitted as in Definition 4 or be a generic W? English usage can be improved in few places, e.g., "Given a 0-1 loss function" should read "Given the 0-1 loss function", "provide bounds the probability mass" should read "provide bounds for the probability mass"

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors correctly describe the limitations of the results in the paper. In that sense, it would be good if the reliance on the realizable case is further discussed.

Reviewer 8X7Z7/10 · confidence 3/52023-06-30

Summary

The authors explore advesarial test-time attacks and distribution shift. They propose a learning algorithm with performance guarantees.

Strengths

The paper is well written with many intuitive illustrations. The authors use refusal as a means to guarantee reliability. As noted by the authors, the trivial classifier that always refuses is reliable. Therefore, I particularly liked section 3.1, which characterizes the probability of refusal (or non-refusal). The concept of reliability radius is readily understandable. The algorithms provided in section 4 can be efficient enough for many real-world applications.

Weaknesses

The authors' analysis is limited to the realizable case; that the true target function is a member of the hypothesis class. While this simplifies analysis, it limits the applicability of the authors' results since in most real-world applications we do not know whether this assumption is valid.

Questions

Can slack be added to the quadratic program in section 4 in the same manner as it is done for SVMs? For the regularized objective in section 4, what good is a lower bound on the reliability radius? It can guarantee an unreliable prediction is unreliable, but cannot guarantee a reliable prediction is reliable. Yet it's the later case that is needed to not refuse for reliable test points.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Limitations are adequately addressed

Reviewer uQQ57/10 · confidence 3/52023-07-06

Summary

This paper presents robustly-reliable learners with optimal guarantees for environments where the training and test data are not drawn from the same distribution, e.g., natural distribution shift and adversarial attacks during test time. The main idea is that for a given point, the robustly-reliable learner either outputs a prediction and a reliability region, or abstains from prediction. The prediction is guaranteed to be correct as long as the test-time perturbation is constrained to this reliability region.

Strengths

* Designing reliable learners with guarantees for challenging environments is highly important and relevant to both the theoretical and applied machine learning community. * It seems likely that other researchers will find relevant the reliability criterions and tools developed in the paper. * The approach appears to be novel and technically sound. The claims of the paper are well-supported by extensive proofs. * The submission is well organized and clearly written overall.

Weaknesses

* A computationally efficient method is presented only for the case of linear separators. It is not clear how easily the presented tools can be used to obtain practical algorithms for more general cases (e.g., neural networks) in practice. * An evaluation on a simple synthetic scenario to demonstrate the empirical effectiveness of the approach would make the paper more compelling. This is a minor point given that this is clearly a theory paper.

Questions

Is there an efficient method for other types of classifiers and loss functions, such as neural networks?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The limitations are not discussed.

Reviewer LDC74/10 · confidence 3/52023-07-11

Summary

This paper studies the problem of classification under found different kinds of adversarial loss functions: 1) ST which I think is by far the most popular adversarial loss. This is same as the expected sup loss of [Madry et al. 2018](https://arxiv.org/abs/1706.06083). 2) TL which is equivalent to the the "exact in the ball" risk of [Gordeau et al](https://www.jmlr.org/papers/volume22/20-285/20-285.pdf) or the "error region risk" of [Diochnos et al](https://proceedings.neurips.cc/paper_files/paper/2018/file/3483e5ec0489e5c394b028ec4e81f3e1-Paper.pdf) 3) IA which is the same as ST but where the adversary only perturbs points belonging to one of the two labels 4) CA loss which is the same as TL with the additional constraint that the true label does not change at the perturbed point. For each type of loss, the paper establishes the "optimal robustly reliable region". Here, the classification is deemed reliable if the the classification is correct on the _perturbed ppoint_. The papers shows that these regions can be computed efficiently for simple hypothesis classes like linear classifiers. The paper also proves lower bounds on "safely-reliable region" for linear separators under log-concave distributions, where a point is classified safely reliably at $\eta_1, \eta_2$ levels if it can be reliably classified at $\eta_2$ level, even when perturbed in a ball of radius $\eta_1$. Finally, the paper also lower bounds the amount by which the reliability region can change under a distribution shift from P to Q in terms of a $P\to Q$ disagreement coefficient.

Strengths

- Generality: The results of the paper hold for 4 kinds of adversarial loss functions, each with different use-cases. - Nice result on reliable learning under distribution shift: Theorem 5.3 proves a connection between reliable learning under distribution shift, and a previously studied notion called the disagreement coefficient. This connection appears interesting and non-trivial.

Weaknesses

- Writing: The paper is very difficult to read. There are many new definitions but very few illustrations / examples. The paper seems to be written in a hurry. Some of the results that are listed under main contributions in section 1.1 only appear in the appendix. Section 6 is super short - it only presents one new definition, followed by an example satisfying the definition the details of which are relegated to the appendix. - Limited contributions: The paper makes limited contributions on the "safely-reliable" notion of classification (Definition 6), which I think is much more important than the notion of "reliable" classification (Definition 3, 4). Theorem 3.3 on probability mass on reliable region only holds for the special case of linear separators with isotropic log-concave distributions. Section 6 on safely-reliable correctness under distribution shift establishes the safely-reliable correctness again only for the special case of linear separators with isotropic log-concave distributions. Further, the shifted distribution is also assumed to only shift in mean, while the covariance matrix remains as Identity matrix. Section 4 also focuses on simple settings like linear separators. - Limited applicability of the notion of "reliability": I want to stress again that the "safely-reliable" notion is more important than the "reliable" notion. The importance of safely-reliable region over the reliable-region is stated by the authors themselves in lines 216-218: "...we note that the probability mass of the robustly-reliable region may not be a meaningful way to quantify the overall reliability of a learner because a perturbation may lie outside of the support of the natural data distribution and have zero probability mass." Basically, a classifier can be robustly reliable at z with level $\eta$ even if predicts a different label at a slightly perturbed point x such that d(x, z) < $\eta$. Hence, reliability does not ensure robustness against adversarial attacks. Literature on certified robustness (for example, randomized smoothing, interval bound propagation) all focus on something like the "safely-reliable" notion rather than reliable. In lines 91-94, the paper says "Prior works have examined pointwise consistency guarantees [SKL17, CRK19, WLF22], i.e. the classifier’s prediction is guaranteed not to have changed under an attack. In contrast, we study a much more desirable property of reliability—guaranteeing that the prediction of the algorithm is correct." I would like to see the paper make a stronger case advocating for the study of reliability than the one given here.

Questions

- Definition 7 is unclear: What is z in line 296? I think it should be any x' in U(x). - I would recommend cutting down on the size of section 1.1: summary of contributions. - It would be good to have more examples to accompany the definitions, for instance contrasting the reliable vs safely-reliable notions. I would also like the authors to comment on my view of the paper in the "weaknesses" section.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Yes.

Senior Area Chairs2023-08-20

please take a look at the author response

Thank you.

Reviewer uQQ52023-08-13

Thank you for your reply. I read the other reviews and your responses to them. I lean towards keeping my score unchanged.

Reviewer 8X7Z2023-08-14

Acknowledgement

I have read the authors' rebuttal. I feel they have adequately addressed my questions and concerns. I am leaning toward leaving my score as it is.

Reviewer nCeL2023-08-15

I thank the authors for their answers. I believe the paper deserves to be published and it will be improved in the camera ready version. Just an additional comment. I agree with the authors that starting with the realizable case is in general good to get the main intuition. However, for this paper I think it would be useful if the authors discuss such assumption in relation with the distribution shifts, since the realizable assumption somehow constraints the shifts addressed in the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC