Agnostic Multi-Group Active Learning

Inspired by the problem of improving classification accuracy on rare or hard subsets of a population, there has been recent interest in models of learning where the goal is to generalize to a collection of distributions, each representing a ``group''. We consider a variant of this problem from the perspective of active learning, where the learner is endowed with the power to decide which examples are labeled from each distribution in the collection, and the goal is to minimize the number of label queries while maintaining PAC-learning guarantees. Our main challenge is that standard active learning techniques such as disagreement-based active learning do not directly apply to the multi-group learning objective. We modify existing algorithms to provide a consistent active learning algorithm for an agnostic formulation of multi-group learning, which given a collection of $G$ distributions and a hypothesis class $\mathcal{H}$ with VC-dimension $d$, outputs an $ε$-optimal hypothesis using $\tilde{O}\left( (ν^2/ε^2+1) G d θ_{\mathcal{G}}^2 \log^2(1/ε) + G\log(1/ε)/ε^2 \right)$ label queries, where $θ_{\mathcal{G}}$ is the worst-case disagreement coefficient over the collection. Roughly speaking, this guarantee improves upon the label complexity of standard multi-group learning in regimes where disagreement-based active learning algorithms may be expected to succeed, and the number of groups is not too large. We also consider the special case where each distribution in the collection is individually realizable with respect to $\mathcal{H}$, and demonstrate $\tilde{O}\left( G d θ_{\mathcal{G}} \log(1/ε) \right)$ label queries are sufficient for learning in this case. We further give an approximation result for the full agnostic case inspired by the group realizable strategy.

Paper

References (27)

Scroll for more · 15 remaining

Similar papers

Peer review

Reviewer T3gQ4/10 · confidence 3/52023-06-22

Summary

This paper studies active learning in a multi-group setting, in which the learner performs adaptive sampling over groups of distributions for solving a binary classification task. The performance of the learner is measured by its maximal loss over the groups. The main goal is to extend the disagreement-based active learning approaches to the multi-group setting. To achieve this, the paper proposes a novel worst-case (over groups) loss estimator for the learner. The theoretical analyses on the sample complexity of the algorithms using this new loss estimator are provided.

Strengths

- The paper considers the multi-group extension of the active learning problem, which is likely a hard problem of high interest for the active learning community. The setting considered in the problem can also be seen as a variant of the Group-DRO problem in which the set of hypotheses and the loss function are not assumed to be convex. - The paper shows rigorously why the traditional disagree-based active learning approaches in the multi-group setting, and proposes a novel multi-group loss estimator with reasonable theoretical guarantees. - For the most parts, the writing is clear and easy to follow.

Weaknesses

Major: 1. The main algorithm, Algorithm 1, seems computationally intractable. A thorough computational complexity analysis is lacking, especially in the most important steps in Algorithm 1 on line 4 (computing $R_i$), line 21 (computing $\hat{h}_i$) and line 22. These steps involve filtering two sets ($\mathcal{X}$ and $\mathcal{H}_i$) that are possibly uncountably infinite, however no explanation is given on how these steps are done. This reduces the impact of the proposed approach to practical scenarios. 2. It is not clear why the sample complexity bound given in Theorem 1 is an improvement over passive and collaborative learning. The bound in Theorem 1 still depends on $d/\epsilon^2$, which is confusing since the text wrote ``the interaction between $d$ and $1/\epsilon^2$ is removed". In fact, the additional multiplicative factor $G\theta_g^2\log(1/\epsilon)\nu$ can be larger than 1, which makes the bound in Theorem 1 worse than the collaborative learning bound shown in Table 1. 3. A lower bound is lacking, which makes it impossible to judge how tight the upper bounds are. 4. Empirical evaluation is lacking. Minor: 1. Throughout the paper, the references to various technical lemmas are missing. In Section 5.2, the paper wrote "it is a lemma of ours that...", with no references. Similarly, the proof of Lemma 1 in the Appendix (line 433) mentions "a lemma of Vapnik [28]", which is meaningless because Vapnik [28] is a 700-page book. 2. One the same line 433 - 434, the $1/m$ term in the uniform convergence bound looks redundant. Mathematically it might be correct, but a tighter (and simpler) bound that only depends on $\sqrt{\frac{\ln(8/\delta) + d\ln(2em/d)}{m}}$ (up to constants) are well-known and can be found in any standard machine learning texts (e.g. see [1], section 28.1, page 393). 3. This factor of $1/m$ is missing in the second equation on line 440, the first term. 4. In the main text: "with the ability decide" -> with the ability to decide; "not enough infer" -> not enough to infer; "would chose" -> would choose. 5. Typo in the equations in Example 1: it should be $\ell(h, i, A)$, not $\ell(h, i, S)$. [1] Shai Shalev-Shwartz and Shai Ben-David. 2014. Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press, USA.

Questions

1. Can the authors comment on the computational complexity issue above regarding the computation of $R_i$ and $\mathcal{H}_{i+1}$? 2. How exactly is the bound in Theorem 1 better than the existing bounds in related settings? 3. How can we compare the results in the paper with the sample complexity lower bound for adaptive sampling strategies in the Group-DRO setting [1]? I am aware that the two settings have different assumptions on the convexity and compactness of $\mathcal{H}$ and $\ell$, so my question is restricted to compact convex $\mathcal{H}$ and convex $\ell$ only. In other words, would the lower bound in [1] apply if $\mathcal{H}$ is compact convex and $\ell$ is convex in your active learning setting? [1] Optimal algorithms for group distributionally robust optimization and beyond. Soma et al., 2022.

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

I suggest the authors to include more detailed discussion of the limitations of the proposed algorithms in the main text. These include any computational as well as theoretical issues raised above.

Reviewer 7xXY5/10 · confidence 4/52023-07-05

Summary

This paper looks at disagreement-based active learning (DBAL) where the goal is to minimize the maximum error over a (known) set of G distributions ("groups"). One could learn each group individually, but the optimum for each group may be very bad on other groups. By combining the search, a decent sample complexity can be obtained.

Strengths

It's an interesting problem, fairly well motivated and it's clearly explained (Example 1) why standard DBAL won't quite cut it. You definitely need to do something clever to handle having multiple groups, and the bound they get is competent. I enjoyed reading the paper.

Weaknesses

The main problem with Theorem 1 is that the result has theta^2 where standard DBAL gets theta (here, theta is the disagreement coefficient). It really seems like theta should be possible here. In the easier regime of getting error O(eta) [Theorem 3], the dependence is theta as one would expect. But the bound there is d G theta log^2(1/eps), and it seems like the log^2 should be log(1/eps): Standard DBAL only needs log(1/eps). Something like: (1) get loss (eta + eps) within each class, using d G theta log (1/eps) samples. This gives hypotheses h_1, ..., h_G such that each h_g has error at most (eta + eps) on class g. (2) output the hypothesis h^ that minimizes max_g Pr_{x ~ g} [ h^ != h_g]. Since the optimum h* has error eta on each class relative to the truth on class g, it has error at most (2 eta + eps) relative to h_g so h^ has error at most (2 eta + eps) relative to h_g on g so h^ has error at most (4 eta + 2 eps) relative to h* on g. I guess this is a 4 eta approximation, not a 3 eta approximation, but that doesn't seem very significant (also unsure if it's necessary).

Questions

Is there any reason to expect the theta^2 or log^2 dependences in Theorem 1/3 to be right? Is it possible to have experimental evidence that this stuff ever works, even synthetically?

Rating

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

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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes.

Reviewer rZgu7/10 · confidence 3/52023-07-07

Summary

The paper considers an active binary classification setting where the feature space is partitioned into several groups. The performance of the learner is given as their worst case performance across all groups. The effect of considering such a multi-group setting is that, it is no longer efficient to sample exclusively in the disagreement region of said class. The authors propose an algorithm which iteratively eliminates hypothesis from the hypothesis class, while keeping track of the performance of hypothesis outside the disagreement area. The trick they use, is to split the loss of a classifier between the loss on the disagreement region and its complement. The later can be estimated by the loss of an arbitrary hypothesis from all remaining hypothesis. This removes dependence on the VC dimension from the estimation of the second term. They give PAC guarantees for their algorithm with an upper bound on expected sampling time. The authors also consider the problem of group realizable learning. In this setting a perfect classifier is assumed to exist for each group. Here, they propose an algorithm which first runs a sub routine to find an $\epsilon$ good classifier on each group. They then use said classifiers to label points in each group and finally do ERM with respect to the multi group objective. They provide PAC guarantees for the algorithm with an upper bound on the sampling time.

Strengths

The paper is well written and easy to read.The motivating example of section 4.2 gives good intuition as to why it is not sufficient to sample on in the disagreement region, which is in itself an interesting phenomenon. The mechanism of algorithm 1 is well described and the reader is given good intuition as to how it's sampling strategy leads to a reduction in VC dimension.

Weaknesses

The authors compare the result of theorem 1 to a lower bound from passive learning. It would be of more interest to consider lower bounds in the active setting, even in the form of conjectures. Discussion on the necessity of the $\log(1/\epsilon)$ term is also missing.

Questions

Have the authors considered a setting where, instead of two oracles, the learner can only receive limited samples from the Rademacher random variable with parameter $\mathbb{P}_{D_g}(Y=1 | x)$, for an $x$ of there choosing, where the partition of the feature space across the groups is known. In particular this should change the nature of group realizable learning.

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

The authors could address the lack of lower bounds for the active setting in more detail. Also Theorem 3 is presented in isolation, it would be nice to have some speculation on whether this result is optimal and also to view it in context of previous literature.

Reviewer eXQR6/10 · confidence 3/52023-07-07

Summary

This paper studies multi-group learning from the active learning perspective. The authors first discuss why classical disagreement-based active learning methods cannot work and provide an example. Then, they propose a modified algorithm and provide theoretical guarantees. The label complexity bound is smaller in order than the lower bound in collaborative learning under certain condition. For the special case of group realizable setting, they also establish improved theoretical results.

Strengths

1. This paper is well-written and easy to follow. 2. This paper clearly states why the classical disagreement-based active learning methods cannot work in the multi-group learning setting. The example in Section 4.2 is interesting. 3. The authors compare their label complexity bound with the lower bound in collaborative learning setting and show that their bound is smaller in order under certain condition. 4. Improved theoretical results are achieved for the group realizable setting.

Weaknesses

1. This paper only discusses the drawback of disagreement-based methods in multi-group learning. How about uncertainty-based methods? What if the learner just label the sample from the most uncertain group? 2. From line 22 in algorithm 1, the proposed method is computationally inefficient and cannot be implemented in practice. Minor: The caption of table 1 states that $\widetilde{\mathcal{O}}$ hides $\log(1/\epsilon)$, it should be $\log(\log(1/\epsilon))$?

Questions

1. How about uncertainty-based methods? What if the learner just label the sample from the most uncertain group? 2. Is the proposed algorithm computationally efficient? Are there any insights for the computational complexity?

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

3 good

Contribution

3 good

Limitations

N/A

Reviewer rZgu2023-08-17

Thank you for the detailed answer to my question, that makes sense! My score remains unchanged.

Reviewer T3gQ2023-08-18

I have read the authors' responses. In Section 3.1, the paper states that the strong assumption $\nu = 0$ is not used, and this seems to be an advantage of the paper over existing works. However, the rest of paper (and also in the authors' rebuttals) indicates that the bound in Theorem 1 (the main theorem of the paper) is either significant and/or comparable to existing bounds only when $\nu$ is small and close to zero. This seems to be a major contradiction that has not been properly addressed in the rebuttals, hence my score remains unchanged.

Authorsrebuttal2023-08-19

The lack of assumption on $\nu$ presents the challenge of creating a statistically consistent algorithm in arbitrary noise regimes, which this paper solves. It is known in the active learning literature that active algorithms only improve over passive approaches in low noise regimes: in the case of the 0-1 loss, one has a label complexity lower bound of $\Omega(d\nu^2/\epsilon^2)$ [12]. Thus, any active scheme one designs will have this term of $d \nu^2/\epsilon^2$ in the upper bound, and improvement over passive will rely on $\nu$ being small. Conceptually, active learning algorithms thus give the following advantage over passive: if one believes that $\nu$ is small, but cannot be sure, running the active scheme will definitely get you an optimal classifier in the limit - no matter what the noise - but you will only enjoy a label complexity speed up if you're right about the noise rate $\nu$ being small. Our work presents an algorithm with such guarantees.   The two faces of active learning. Dasgupta, 2012. [12]

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC