Distribution Learnability and Robustness

We examine the relationship between learnability and robust (or agnostic) learnability for the problem of distribution learning. We show that, contrary to other learning settings (e.g., PAC learning of function classes), realizable learnability of a class of probability distributions does not imply its agnostic learnability. We go on to examine what type of data corruption can disrupt the learnability of a distribution class and what is such learnability robust against. We show that realizable learnability of a class of distributions implies its robust learnability with respect to only additive corruption, but not against subtractive corruption. We also explore related implications in the context of compression schemes and differentially private learnability.

Paper

References (86)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 2DNr6/10 · confidence 3/52023-06-19

Summary

This paper studies the interaction between learnability and several notions of "robust" learnability, wherein candidate distributions can be corrupted in various ways. The authors show an equivalence between learnability and a robust version of learnability in which a contaminated distribution can be added to the nominal distribution. The authors study how differential privacy interacts with learning, showing that DP can imply learnability under certain contamination models, but not others. Sample compression schemes are also explored. **Overall assessment.** This paper is strong on several fronts. It considers a wide selection of problems related to distribution learnability. And the technical quality of the paper is high. On the negative side, I think the biggest drawback is the writing. The clarity with which a paper like this is written is crucial, as the results are of a technical nature, thus requiring careful, clear explanations so that the reader can follow along. In its current form, I think this paper falls short of the mark. If the authors can demonstrably improve the writing and presentation, I will raise my score accordingly.

Strengths

**Breadth of settings.** This paper considers numerous problem settings spanning additive (and subtractive) contamination, differential privacy, learnability (in the sense of PAC learning), and sample compression. One imagines that the scope of this paper could have been reduced, i.e., to the interaction of DP and learnability, and the paper would still constitute a solid contribution. In an age when it can be advantageous to increase one's throughput by spreading content over multiple publications, I appreciated that the authors endeavored to give the reader the full picture capturing how learnability manifests on all of these different axes. **Soundness.** From what I can tell, the paper is quite sound. The paper presents informal discussions and in some cases informal proofs in the main text. The appendix is fully rigorous, with each main result proved in its own subsection.

Weaknesses

**Motivation/Relevance to practice.** There is little to no motivation for the problem under consideration. The authors begin the introduction with a description of the problem of distribution learning without mentioning why one might be interested in this problem and/or what bearing this problem has on practical ML settings. It's certainly worth noting that this is a more theoretical paper, but even still, I think that the paper would be stronger -- and would appeal to a wider audience -- if the authors motivated the problem in a more verbose way. As written, this paper would only be accessible or feel relevant to those who are already familiar with why distribution learning is a well-studied problem, which is a narrow slice of the ML/statistics community. **Definitions.** I found the definitions on the first few pages hard to parse. The three definitions contained in Def. 1.1 are all almost exactly the same; they each differ by at most a couple of words. I think it would make much more sense to condense this by using some sort of meta-definition, and then listing the three definitions as special cases. This also applies to Definition 1.4. In the case of Definition 1.4, the only difference relative to Def. 1.1 seems to be that the learning algorithm is randomized, which could surely be summarized in a more concise way. Another related point is that the authors state the definitions with relatively little context. It would be help to illustrate each of these definitions with an example or a figure. As I mentioned earlier, the differences between each of these definitions is relatively subtle, and it's asking a lot from the reader to keep all seven definitions on pages 2-3 in their heads at once, along with the fine distinctions that separate them. **Writing.** Much of the prose is notably spare in a way that makes the paper more challenging to read. Results (e.g., Theorem 2.1) are often without any preamble. It would be worth thinking about reorganizing the paper so that each result is accompanied by an intuitive explanation of the relevance of the result and how it fits into the broader context of the paper. While I like the attempt to informally present the theorems before offering a more rigorous treatment after Section 1, I don't think that it works here. This is in part because it asks the reader to consider each result at three levels of abstraction. First, the results are stated informally in Section 1. Then, in the remainder of the main text, the same results are stated formally, but the proofs are informal or sketched. Finally, the full treatment is given in the appendix. I think that asking the reader to jump around at these different levels of abstraction is, in this case, confusing. Perhaps the authors could shorten the intro and stick to the results as stated from Section 2 onwards? Relatedly, the intro seems too long here. The reader has to wait until page 5 to learn what the authors feel the contributions are. This dilutes the standard linear flow of ideas that characterizes papers that are easy to follow. To make the paper easier to follow, I would recommend rewriting the introduction so that (a) the problem of distribution learning is well-motivated, (b) the main results are discussed, but not stated, and (c) the contributions are listed by -- at the very latest -- the end of the second page. **How is this "robust?"** One question that jumped out to me was as follows: How are the definitions in Definition 1.1 "robust?" I.e., what are they robust against? The authors claim that the name "robust learnability" is due to the fact that they are accounting for the ground truth distribution $p$ may not be an element of the class $\mathcal{C}$. Would it not be more accurate to call these definitions "agnostic" rather than "robust?" I can see how one would regard the corruption-based definitions in Def. 1.2 and Def. 1.3 as robust, because the distribution $p$ is mixed with a corrupted distribution. However, in the sense of Definition 1.1, I don't understand the use of the word "robust." **Miscellaneous.** * It looks like a some points the authors write $(\alpha,\eta)-robustly learnable and $\alpha-\eta$-robustly learnable (c.f. Def. 1.1(3) vis-a-vis Def 1.4(3)). It seems like it would be good to pick on notation and stick with it. * Why is the TV distance chosen here? The authors assert that this distance metric makes sense, but they give no argument, leaving the reader with some confusion. * It would be worth adding a discussion section after the main results but before the conclusion. After reading through the results, it would be useful if the authors could piece them together in the broader context of distribution learnability. The authors could also discuss how these results could lead to future research on this problem, and/or whether this could lead to new algorithms that would be of practical use.

Questions

See above.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

2 fair

Contribution

3 good

Limitations

N/A

Reviewer 1ajk6/10 · confidence 3/52023-07-06

Summary

This paper studies distribution learnability, with a focus on the relationship between the realizable and the agnostic/robust settings. This relationship is also studied in the context of sample compression schemes and differential privacy. In particular, the authors show separation results: learnability in the realizable setting implies learnability when an adversary can modify the training data by adding training samples. They show however that this does not hold when the adversary can remove part of training data. In fact the authors identify that an adversary having subtractive power implies robust learnability in the more general case where data can be replaced. In terms of connection to sample compression schemes, the main result is that the existence of sample compression schemes in the realizable setting does not guarantee the existence of robust sample compression schemes. Finally, they show that pure DP learnability implies pure DP learnability when the adversary can replace data, but if one considers approximate DP, the implication breaks for adversaries that can subtract data to the training set (but not for adversaries that can add data).

Strengths

- Good presentation, well written, big picture is given for most proofs and the paper easy to follow. The introduction is particularly thorough - The impact is quite broad, with applications to sample compression schemes and differential privacy - I believe these results would be of interest to the learning theory community - The technical contributions seems appropriate for NeurIPS, given the significance of the conceptual contributions Disclaimer: I am not very familiar with DP, so my assessment for the contributions is mainly based on Sections 1-3.

Weaknesses

- Constant $\alpha$ (Thm 1.5): do you have a sense whether $\alpha=4$ the best one could hope for? This high constant seems to be a limitation of the work, in the sense that the only guarantee is that we are roughly at multiplicative factor of 4 away from the optimal distribution in the class $\mathcal{C}$. It's a bit difficult to get a sense of the significance and the text could use some clarification. General comments/things to fix: - p.2: "total variance distance" -> "total variation distance" (maybe include the mathematical definition too? No need for a new definition environment, just $d_{TV}(p,q)=...$) - Section 6.6 in the appendix does not have any text below it

Questions

- Terminology: robust and agnostic learnability are used interchangeably (e.g., l.32), which is in contrast to other set ups (e.g. binary classification), for which the two terms have vastly different meanings. The set up described on l.32-37 (and Defs 1.2 and 1.3) sounds more like the agnostic setting in spirit, where we want to compare the algorithm's output with the best possible one from the class $\mathcal{C}$. In my mind, robustness implies the presence of an adversary which modifies data (at training or testing time), so more in line with definition 1.4. I think using robust and agnostic interchangeably is a bit confusing and I would appreciate some clarification on the matter (here and on a future version of the paper), perhaps a distinction like suggested above. General comments: - Perhaps provide references for poisoning attacks on l.28, as it is particularly related to "adaptative robustness" (though this is in the context of classification).

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes.

Reviewer Xv717/10 · confidence 5/52023-07-06

Summary

This paper study the connection between learnability and robust (agnostic) learnability for general classes of distributions in the framework of distribution learning. Differ from the PAC learnablity, they show that in the distribution learning setting the learnability does not always implies agnostic learnablity. They also explore related implications in the context of compression schemes and differentially private learnability.

Strengths

* This paper is well organized and easy to follow, I enjoy reading it. * The result that learnability does not imply robust learnability is somewhat surpring. * The proofs are solid, by my judgement.

Weaknesses

N/A

Questions

* I am a little confused about the definitions of adaptivity. In Definition 1.4 (1) and (3), what distribution does the new $\eta m$ examples follow? If they follow $p,$ these definitions seem to make no sence. If these examples are drawn from some unkown distriburtion, how should I understand the definition of **adaptive robust** learnablity? * I would be very appreciated it if the authors can give some intuition about the contradiction between their findings and the results presented in [1]. Reference: [1] Max Hopkins et al. Realizable learning is all you need. Colt 2022.

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer 1ajk2023-08-11

Response

Thanks for the response. *Update on the $\alpha$ term*: nice! If you end up updating the constant in the final version, could you include a discussion in line with the one in your response? *Robust vs agnostic terminology*: thank you for the clarification! I am looking forward to reading the above updates in the final version of your work.

Reviewer 2DNr2023-08-14

Rebuttal response

Thank you for your rebuttal response. It's a shame that one cannot upload a new PDF, because I would quite like to read another draft of this paper, given that the updated bits all center around the presentation. However, in the spirit of looking for reasons to accept rather than reasons to reject, I'll trust that the authors will follow up on the motivation, definitions, and writing style in the camera-ready version of their paper. In particular, I think that adding motivation and historical context for this problem would be helpful, particularly for someone like your reviewer, who (as you can tell based on my review) is only loosely familiar with the problem being studied here (hence, `Confidence=3`). Given that these changes will -- in my view -- improve the paper, it's only fair that I increase my score.

Reviewer Xv712023-08-17

Response

Thanks for the response.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC