Swap Agnostic Learning, or Characterizing Omniprediction via Multicalibration

We introduce and study Swap Agnostic Learning. The problem can be phrased as a game between a predictor and an adversary: first, the predictor selects a hypothesis $h$; then, the adversary plays in response, and for each level set of the predictor $\{x \in \mathcal{X} : h(x) = v\}$ selects a (different) loss-minimizing hypothesis $c_v \in \mathcal{C}$; the predictor wins if $h$ competes with the adaptive adversary's loss. Despite the strength of the adversary, we demonstrate the feasibility Swap Agnostic Learning for any convex loss. Somewhat surprisingly, the result follows through an investigation into the connections between Omniprediction and Multicalibration. Omniprediction is a new notion of optimality for predictors that strengthtens classical notions such as agnostic learning. It asks for loss minimization guarantees (relative to a hypothesis class) that apply not just for a specific loss function, but for any loss belonging to a rich family of losses. A recent line of work shows that omniprediction is implied by multicalibration and related multi-group fairness notions. This unexpected connection raises the question: is multi-group fairness necessary for omniprediction? Our work gives the first affirmative answer to this question. We establish an equivalence between swap variants of omniprediction and multicalibration and swap agnostic learning. Further, swap multicalibration is essentially equivalent to the standard notion of multicalibration, so existing learning algorithms can be used to achieve any of the three notions. Building on this characterization, we paint a complete picture of the relationship between different variants of multi-group fairness, omniprediction, and Outcome Indistinguishability. This inquiry reveals a unified notion of OI that captures all existing notions of omniprediction and multicalibration.

Paper

References (32)

Scroll for more · 20 remaining

Similar papers

Peer review

Reviewer Xn6q7/10 · confidence 4/52023-06-28

Summary

This paper introduces and studies 3 learning theoretic problems, namely "swap" versions of agnostic learning, omniprediction, and multicalibration. These swap versions are inspired by the notion of "swap regret" in online learning. They show that in some settings, all three of these learning-theoretic problems are equivalent (up to polynomial factors). Then they also introduce another notion called Swap Loss Outcome Indistinguishability and study relationships between it and the others.

Strengths

- The paper is clearly written, and all the definitions of the different learning problems make sense. The proofs are easy to follow/digest. - The results seem novel, and provide a relatively complete picture of the relationships between the different learning problems that are introduced.

Weaknesses

- The main question which is not really addressed is why one should care about these learning problems? Agnostic learning and multicalibration are important problems because they have tremendous practical significance. Maybe the authors can touch upon this in the rebuttal. At first glance (since I do not have much background in the field of multicalibration), it seems like the authors introduced three "brand new" learning problems and proved relationships between them.

Questions

1. Could the authors address why one should study these learning problems (swap versions)? 2. While the paper was motivated by studying omniprediction and multicalibration, the results herein focus on the swap version. What are the high level implications for the relationships between omniprediction and multicalibration? 3. Can one do an online2batch reduction for the swap versions of online learning and agnostic learning? 4. How do the results depend on the size of the support of $\tilde{p}$? Do they become less interesting if $\tilde{p}$ is assumed to be a continuous-valued predictor (as is done in many practical scenarios)? 5. Could the authors state some examples of "nice" loss functions? Are all the loss functions commonly used in learning theory, e.g, square loss, hinge loss, exp loss, log loss, etc all "nice"? Minor typos/clarifications: - line 63: "defintions" - I may have missed this, but what is meant by the notation $\ell_2$ and $\ell_4$? In some cases, it seems like it means square loss, but in others it is used to denote a loss function that is picked for a value $v$. - What is meant by "sparse"? Typically it means a bound on $\ell_0$ (or the number of nonzero weights), but it seems like you are using it to mean $\ell_1$ bound. - line 297: "ot" - line 188: do you mean $\lVert \mathcal{C} \rVert_\infty \le 1$?

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

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Yes, n/a.

Reviewer if878/10 · confidence 4/52023-07-05

Summary

The authors propose a stronger variant of agnostic learning called *swap* agnostic learning, and show that this notion of learning is equivalent to “swap” variants of multicalibration and omniprediction. Multicalibration means that the predictor is simultaneously calibrated on a given collection of subpopulations $\mathcal{C}$. Omniprediction means that a *single* predictor’s output values can be post-processed so that its performance is competitive with the best from a given hypothesis class $\mathcal{C}$ w.r.t. multiple loss functions. Here, the post-processing must depend only on the given loss function and *not* the covariates. Multicalibration is related to algorithmic fairness since it means that “chance” predictions of a given (potentially malicious) predictor “mean the same thing” across many subpopulations. Omniprediction is useful for training machine learning (ML) models since it means that one can train a *single* predictor that performs well w.r.t. multiple loss functions. Thus, an omnipredictor obviates the need to train separate models for each loss function. Previous work has shown that multicalibration is sufficient for omniprediction, thereby showing that satisfying fairness constraints may lead to practical benefits in machine learning. The question of whether multicalibration is *necessary* for omniprediction is still open. This work takes a step towards this question by showing that their *swap* analogues are equivalent to each other.

Strengths

- **Contribution.** The paper makes a significant and timely contribution to the fast-growing literature on multicalibration by showing that seemingly distinct notions of success in learning and algorithmic fairness are in fact equivalent. Key to their result is their definition of swap agnostic learning, a stronger learning requirement than (vanilla) agnostic learning, and the swap variants of multicalibration and omniprediction. The swap analogues of each notion enable tight connections among themselves. - **Exposition.** The paper is overall well-written. Ideas are presented clearly and concisely. The authors also do a good job situating their results in relation to previous work.

Weaknesses

**Motivation behind swap multicalibration and swap omniprediction.** As a stand alone concept, i.e., for purposes other than establishing equivalence between the notions, the motivation behind the definition of swap multicalibration and swap omniprediction is unclear to me. I expand on this point in the Questions section.

Questions

- Are there provable separations between swap and vanilla versions of each notion (multicalibration, omniprediction)? What separations are known and which ones are unknown? - Multicalibration can be interpreted as “calibration w.r.t. any subpopulation in $\mathcal{C}$”. Does swap multicalibration have such simple interpretation? How would one motivate swap multicalibration as a stand alone concept? Ditto for swap omniprediction. - Calibration has a long history dating back to early research on weather forecasts and their assessments. The natural setting of this early literature is sequential prediction (i.e., inputs are *adversarial* sequences consisting only of labels $y$ rather than i.i.d. samples $(x,y)$ with covariates $x$ drawn from an unknown, but fixed, distribution). Even in the sequential setting, multicalibration has a natural interpretation: calibration w.r.t. multiple subsequence selection rules (see e.g., [Daw85]). How should one think of infinitary analogues of swap mulitcalibration? **Minor editorial comments** - In the abstract, it’s not clear what $p$ refers to when the authors say “the predictor wins if $p$ competes with the adaptive adversary’s loss”. Did the authors mean $h$? - The projection operator $\Pi_\ell$ in Definition 3.1 is not defined anywhere in the paper. - In Theorem 4.2, the word “incomparable” is too vague and uninformative, even for an informal theorem statement. I would suggest the authors provide a more concrete statement. **References** - [Daw85]: Dawid, A. P. Calibration-based empirical probability. *Annals of Statistics*. 1985

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Yes.

Reviewer HdZq5/10 · confidence 2/52023-07-07

Summary

The authors study swap agnostic learning, or a game between a predictor and an adversary where the predictor has to come up with a family of hypotheses to compete with a strong adversary's loss. They prove a surprising feasibility result for swap agnostic learning for any convex loss by proving that swap agnostic learning is equivalent to swap variants of omniprediction and multicalibration. Furthermore, they connect these ideas even further by relating them under the umbrella of outcome indistinguishability.

Strengths

+ The equivalence result seems like it is a novel and useful finding -- however, I was not familiar with the notions of omniprediction and multicalibration. + The technical contributions of this paper are insightful and nontrivial. + This paper connects results from different areas of ML, which is always nice to see.

Weaknesses

- I found the paper to be rather inaccessible in its lack of intuition about results and its overeagerness to use jargon. I acknowledge that this is not an area I am an expert in, but I found it rather hard to follow.

Questions

None

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

3 good

Limitations

Yes

Reviewer XpDf8/10 · confidence 3/52023-07-09

Summary

The paper introduces the notion of swap agnostic learning, and proves the feasibility of swap agnostic learning for any convex loss using only weak agnostic learners. To prove this result, the authors establish the equivalence between swap agnostic learning and the swap variants of multi-calibration and omniprediction, and show that the existing multicalibration algorithm of [20] already guarantees the stronger notion of swap multicalibration. They also show that the notion of swap loss outcome indistringuishability captures all existing notions of omniprediction and multicalibration.

Strengths

- The paper explores the important connection between the group-fairness notion of multi-calibration and the loss minimization notions of agnostic learning and omniprediction, and establishes many interesting equivalences under swap variants. This equivalence result provides new insights into what multi-calibration can offer and deeps the understanding of the relevant notions. - The paper is well-written and clearly explains the problem and its contributions.

Weaknesses

The paper would benefit from additional discussions on the differences between the 'swap' variations and their standard counterparts. For instance, clarifying how 'swap multicalibration' diverges from the conventional concept introduced in [20] would be beneficial. Further, it would be insightful if the paper could elaborate on why equivalences under the 'swap' notions are easier to establish. What are the challenges involved in extending these to the standard definitions? This additional discussion could give readers a more thorough understanding of the subject.

Questions

- In line 5 of the abstract, what does $p$ refer to? Is this supposed to be $h$? - In Definition 2.5, multi-calibration is defined in terms of the expectation of the absolute difference over $v\sim\cD_{\tilde{p}}$, which contrasts with the definition involving $\max_v$ as in Definition 2.6 from reference [20]. How critical is this choice of definition for your results? If multi-calibration were defined in terms of the maximum rather than the expectation, would this make swap multi-calibration and multi-calibration equivalent? Could this be why existing algorithms for achieving multi-calibration, such as those in references [20,15], are able to guarantee swap multi-calibration without modification?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes.

Reviewer 7A6v7/10 · confidence 4/52023-07-09

Summary

This paper studies algorithmic fairness from a learning theoretic lens, specifically extensions of the recent notion omniprediction by Gopalan et al and its connections to to agnostic learning - multicalibration is sufficient to obtain omniprediction and guarantee agnostic learning for all convex loss functions simultaneously. The main contributions are in introducing a new learning paradigm called Swap Agnostic Learning (SAL), and subsequently new notions of "swap" omniprediction and multicalibration in a similar vein, and proving that they are all equivalent. The authors show that the standard multicalibration algorithm of Herbert-Johnson et al. in fact guarantees swap multicalibration algorithm, and consequently can also be used to obtain a SAL guarantee. The contributions all hinge on the introduction of the new learning task - swap agnostic learning. SAL requires, informally, that a predictor plays a hypothesis h such that on any level set of their predictor - i.e. set of all x's such that h(x) = v for some fixed v - they have no incentive to change their predictions to any other hypothesis from the hypothesis class on that set. This contrasts the standard agnostic learning which says that the predictor is competitive with the same, single error-minimizing hypothesis in expectation over all the level sets. The swap notion of omniprediction is similar. A standard omnipredictor for a class of losses L and hypothesis C is a predictor that can be post-processed (in a way that depends on the chosen loss from L) such that it is competitive with the best hypothesis for each loss in L simultaneously. The swap variant requires that the omnipredictor be competitive (up to a level-set dependent post-processing) the best hypothesis on each level set for any assignment of loss functions to level sets. Section 3 proceeds with the proof of equivalence between these three notions and a brief explanation of how the standard McBoost multiwcalibration algorithm in fact satisfies swap multicalibration, and consequently can be used to achieve the other two swap notions.

Strengths

This paper makes a significant contribution in the area of omniprediction and multicalibration, which are currently popular topics in the study of algorithmic fairness. The newly introduced notion of SAL is interesting and provides a complete picture of the connections between the swap variants of mutlicalibration, omniprediction, and agnostic learning. Though the content is technically and notationally dense, the presentation is clear.

Weaknesses

A minor suggestion is that if possible, adding a bit of exposition to the swap omniprediction definition to guide the reader through all the moving parts would be helpful. It feels a little trickier to parse than the swap multi calibration and swap agnostic learning definitions. Also, per the below question, maybe adding some exposition describing why these swap variants are natural / intuitive could be helpful. Simiarly, in Section 3.1, if possible, proof sketches/intuition would be useful as well. There are maybe one or two parts where it is done but the section feels largely like a proof with limited expositional guidance.

Questions

I understand the technical conditions of swap AL, MC, omniprediction and how they are different from the standard notions, but I am slightly confused by the motivation. The original notions seem natural in some sense, is there a reason why these additional technical conditions for the swap variants are intuitive or useful - e.g. giving adversary the additional control down to the granularity of level sets? If I understand correctly, the analogous versions of connections established in this work for all of the "non-swap" notions of these learning tasks are already shown, so again I am slightly confused what the advantage of these swap variants are.

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

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

-

Reviewer 7mxb5/10 · confidence 2/52023-07-26

Summary

This paper provides a new learning framework referred to as Swap Agnostic Learning, a variant of agnostic learning aiming to a stronger goal. It is shown that this framework is equivalent ot swap variants of the notions of Omniprediction and Multicalibration. An algorithm for Swap Agnostic Learning is provided, which is shown to work well for any convex losses.

Strengths

- A new problem setup is introduced along with its interpretation, which would be of interest for readers in the NearIPS community. - The relationship with various existing studies is well explained.

Weaknesses

There are unclear points in the algorithm procedure and analysis. For example, the paper states that MCBoost from [20] is used, but it is not clear what the corresponding algorithm in the paper [20] is. Algorithm 2 given in Appendix B.1 looks very different from algorithms in [20]. In Algorithm 2, there is no explanation of how to set the parameter $\eta$. I cannot find a proof that the iterations in Algorithm 2 terminate. In the algorithm section, the fact that the MCBoost algorithm, originally designed to guarantee multicalibration, actually also guarantees swap multicalibration, would be a very important and nontrivial point, in my understanding. I believe it would be better if a more detailed explanation of this fact and the intuition were added to the text.

Questions

I would appreciate an answer to the concerns in "Weaknesses".

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

I have no concerns about the limitations and potential negative societal impact.

Reviewer if872023-08-14

Thank you for your response! P.S. This is probably irrelevant to the current rebuttal/discussion phase, but it would be great if the authors could provide a reference for the recent arXiv paper on the infinite sequence setting.

Reviewer 7mxb2023-08-16

Thank you for the response. My question in the review has been resolved. I have no additional questions at this time.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC