Approximation-Generalization Trade-offs under (Approximate) Group Equivariance

The explicit incorporation of task-specific inductive biases through symmetry has emerged as a general design precept in the development of high-performance machine learning models. For example, group equivariant neural networks have demonstrated impressive performance across various domains and applications such as protein and drug design. A prevalent intuition about such models is that the integration of relevant symmetry results in enhanced generalization. Moreover, it is posited that when the data and/or the model may only exhibit $\textit{approximate}$ or $\textit{partial}$ symmetry, the optimal or best-performing model is one where the model symmetry aligns with the data symmetry. In this paper, we conduct a formal unified investigation of these intuitions. To begin, we present general quantitative bounds that demonstrate how models capturing task-specific symmetries lead to improved generalization. In fact, our results do not require the transformations to be finite or even form a group and can work with partial or approximate equivariance. Utilizing this quantification, we examine the more general question of model mis-specification i.e. when the model symmetries don't align with the data symmetries. We establish, for a given symmetry group, a quantitative comparison between the approximate/partial equivariance of the model and that of the data distribution, precisely connecting model equivariance error and data equivariance error. Our result delineates conditions under which the model equivariance error is optimal, thereby yielding the best-performing model for the given task and data. Our results are the most general results of their type in the literature.

Paper

References (68)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer PxmR8/10 · confidence 2/52023-07-02

Summary

The paper presents a theoretical investigation into the approximation and generalization errors for a model exhibiting approximate equivariance. The paper introduces the concept of "error to equivariance," a measure of the deviation of the error function from perfect equivariance. It then defines the $\epsilon$-stabilizer, a subset of a group under whose transformations the error function maintains approximate equivariance. Leveraging these definitions, the paper dives into the analysis of generalization error for an approximately equivariant model in terms of PAC-bounds and Rademacher complexity, after which, It also explores the approximation error. Ultimately, the performance error of a model is defined as the sum of its generalization error and approximation error, providing a quantitative measure of the performance of an approximate equivariant model. In the end, the performance error can quantitatively analyze the trade-off between its two terms (generalization error and approximation error).

Strengths

1. The paper theoretically studies approximate equivariance, which is a very significant problem in the equivariant learning literature as real-world problems are not always perfectly equivariant.  2. The paper provides general bounds, contributing valuable theoretical tools to the field.

Weaknesses

1. Section 4 is somewhat challenging to read, especially for readers who are not familiar with PAC-bounds and Rademacher complexity. Additional explanations or intuitive examples could improve accessibility. 2. While not a requirement for a theory paper, the absence of empirical validation or experiments leaves a gap. Experiments that align with the theoretical bounds could provide practical guidance on how the paper's findings might be applied empirically.

Questions

1. Could the authors comment on the broader implications of the proposed theory? Specifically, could the theory help model selection under approximate equivariance?  2. The paper is titled "Approximation-Generalization Trade-offs", however, the high-level message regarding this trade-off isn't clear.  3. In line 140, there is a closing parenthesis but no opening parenthesis.

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

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The paper properly discusses its limitations. An additional limitation to consider is the assumptions of the paper. The paper makes a couple of assumptions throughout the theoretical analysis, it might be good to summarize and discuss those assumptions in the limitation section.

Reviewer 5syP7/10 · confidence 3/52023-07-04

Summary

In this paper, the authors spell out explicitly the benefit of capturing task specific symmetries into machine learning models. This goal is achieved by deriving rigorously generalization error bounds that takes into account the cardinality of the set of transformations that a model is (approxaimately) equivariant to (Stab_\epsilon) as well as the group dimensionality these transformations are elements in. The results are very general and unlike previously presented results in the literature, the sets of transformation the model is equivariant to, are not required to form a group.

Strengths

The presented approach to derive the bounds is principled and very interesting.

Weaknesses

While the derived results are very general and rigorous, it is not directly obvious how these results can be applied. Although the authors give examples in the appendix, the impact of these bounds seems relatively limited. It could have been maybe useful to show an actual model performance w.r.t. the presented bound.

Questions

Is there anything that can be said about the tightness of the presented bounds?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors state that the derived bounds are not model specific and stronger versions could be the subject of future work.

Reviewer R2KT6/10 · confidence 3/52023-07-24

Summary

This paper provides theoretical investigation on the intuition that, machine learning models (e.g. neural networks) whose symmetry aligns with the data symmetry achieve best performance and generalization. The paper formulate the relationship between model equivariance error and data equivariance error by providing quantitative bounds, which are applicable in the presence of approximate or partial data/model equivariance.

Strengths

1. To my best knowledge, the paper provides a systematic and comprehensive review of the historical development of neural networks with group symmetries (**since this is not exactly my area of research, please correct me if I'm wrong**). 2. The paper provides detailed and systematic analyses of the model generalization bounds as well as approximation bounds under the more general case of partial or approximate equivariance. According to the authors, as well as to my best knowledge, this is the first theoretical work that addresses this problem. Therefore its contribution and novelty is significant. 3. Although not being able to check every technical details and derivations, I find the paper well-structured and easy to follow.

Weaknesses

**Major issues:** 1. Derivation of the approximation error bounds: - First of all, in line 291, I don't think it's legal to write $\min A - \min B$ as $\min(A-B)$, since the second term has a minus sign. So I think the correct way to write is that $\text{AppGap}(\mathcal{F}, \mathcal{D}) := \underset{f\in\mathcal{F}}{\min}\mathbb{E}[\mathbb{E}_g[f(g\cdot Z)]] - \underset{F\in\mathcal{M'}}{\min}\mathbb{E}[F(Z)] \ge \underset{F\in\mathcal{M'}}{\min}\mathbb{E}[\mathbb{E}_g[F(g\cdot Z)]] - \underset{F\in\mathcal{M'}}{\min}\mathbb{E}[F(Z)] $ - The expression above provides a **lower bound** to the approximation gap, so how does the inequality become equality in Line 304? If it is derived by assuming $\mathcal{F} = (\mathcal{M}^{(G)})'$ (assumption 3 in line 301=303), then line 304 should be written as $\text{AppGap}((\mathcal{M}^{(G)})',\mathcal{D}) = \underset{F\in(\mathcal{M}^{(G)})'}{\min}\mathbb{E}_Z[\mathbb{E}_g[F(g\cdot Z)]]$ - Moreover, if the upperbound in Prop 5.1. is derived based on the equation above, I think it's necessary to prove that $(\mathcal{M}^{(G)})'\subseteq \mathcal{C}_{\epsilon,\lambda}$ , for any $\epsilon\ge 0, \lambda\in(0, 1]$, which I don't see in the proof. 2. Interpretation of the "generalization-approximation trade-off"/"optimal equivariance" is not clear: In line 360-362, it is claimed that the optimal $\lambda^*$ "validates the intuition stated at the onset that for the best model its symmetries must align with that of the data". However, based on my understanding, the symmetries of the model **only fully align with those of the data** when $\lambda^*=1$, given $\epsilon=0$. When $0<\lambda^*<1$, I think it means the model only **partially preserves the symmetries of the data**. Do you agree? If this is the case, how do you justify the intuition that "the integration of relevant symmetry results in enhanced generalization."? 3. Similar to reference 9 in the paper, I think one key point I would expect to see in this paper is proving that, models with aligned group symmetries of data exhibit higher sample efficiency/complexity, which can generalize to scenarios of partial/approximate symmetry. I see no discussion or proofs regarding this intuition in the paper. **I will be happy to raise my score given satisfactory explanation or revision.** **Minor Issues:** 1. This paper targets specifically on the theory of machine learning models with partial/approximate group symmetries, which might be of interests only to a very specific subgroup of the whole machine learning community. 2. There are some presentation issues need to be improved: - All hyperlinks do not work when clicked. Some numberings are also confusing, e.g. what is Thm A.4 in line 242, 246, 268? - Some of the statements would benefit from more citations/examples, e.g. "It has often been noted that this may impose an overly stringent constraint as real-world data seldom exhibits perfect symmetry” (line 36 - 37) 3. Typos - may have not -> may not have (line 312) - $\text{Stab}{\epsilon}(\mathcal{F'}) \Longrightarrow \text{Stab}{\epsilon}(\mathcal{F})$ (line 309)

Questions

1. More regarding the interpretation of the "generalization-approximation trade-off"/"optimal equivariance": If we examine the contribution of $\lambda$ to the generalization and approximation bounds given by Theorem 4.4 and 5.2 respectively, assuming perfect symmetry of data ($\epsilon=0$), we see that the former decreases when $\lambda$ increases ($\frac{1}{\lambda^{1/2}}$ or $\frac{1}{\lambda^{1/d_0}}$ scaling), while the latter increases when when $\lambda$ increases ($\lambda^{1/d_G}$ scaling), which gives what the authors called "generalization-approximation trade-off". To me, this seems to show that minimizing the generalization error effectively aligns the model and data symmetries, whereas minimizing the approximation error functions in the opposite direction, resulting in a partially aligned model symmetry. Do you agree with this analysis? Any further explanation/intuition? 2. Can you explain what you mean by "oscillations", which is mentioned repeatedly? Example contexts like: - "since $\text{Stab}_{\epsilon}$ includes more elements of G as ε increases, and we can expect that actions of elements generating higher equivariance error in general have higher oscillations". (line 274-275) - line 320-322

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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer y3sQ7/10 · confidence 4/52023-07-25

Summary

In their theoretic work, the authors presents quantitative bounds on the generalization error and approximation error. They extend the results to approximate symmetries, which refers to functions that are close, in the supremum norm, to a group action. Their first results shows that when working with only the representatives of $G$-Orbits, and the models are equivariant (or almost-equivariant) The generalization error can be nearly-recovered. Their second results show that their exists a class almost-equivaraint estimators, such that its approximation error can be bounded (and the bound depends mostly on the properties of G, and somewhat on the pertubations of the optimal estimator).

Strengths

This paper is a novel and well presented work on approximate symmetry. The bounds presented are a significant contribution to the field, further deepening its theoretical foundations.

Weaknesses

- The discussion section is lacking. It focuses on combining previous results and quantifying the lambda parameters. It is lacking interpretation of the results, and addressing your hypothesis. Under which conditions equivariance error is optimal? What is the interplay between data equivaraince to model equivariance? How does it compare to other bounds in the literature? - The limitation of this paper is not addressed properly. The authors only mentions that "the results are not model specific", which is not an actual limitation. It would be beneficial to discuss the actual limitation of the work. Can the bounds be computationally intractable? for example, how can one compute $ddim(G)$? What is the asymptotic growth of the presented bounds? - The paper is lacking intuition into the $\lambda$ parameter. Explanations are postponed into the discussion, and even there, the interplay between $\lambda$ and the error is drawn, but there is no attempt to demystify this parameter. I would appreciate further elaboration about it, and about the meaning of $Dens(\epsilon)$. Is this the same as 'approximate orbits'$^{-1}$?

Questions

- Theorem 5.2 holds under the condition that there \textbf{exists} $\mathcal{F} \in \mathcal{C}_{\varepsilon, \lambda}$ such that the theorem holds. I am concerned that ideally, we would like it to apply to \textbf{any} such $\mathcal{F}$ . It would seem to me that it would align better with our original motivation as we would like to prove that any such partial symmetry family (with limited perturbation) can generalize well. I note that theorem 6.1 is under the same assumption so my question additionally applies to it - I would appreciate further elaboration specifically on the interplay between model equavarince against data equavarince. How does the presented bound connect between them? can you show how further restriction of the models, choosing $\mathcal{F}' \subset \mathcal{F}$ would affect the generalization and approximation bound.

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

2 fair

Contribution

3 good

Limitations

.

Reviewer W1Je4/10 · confidence 2/52023-07-26

Summary

The submission discusses the role of task-specific symmetries in enhancing machine learning models. It presents a quantitative analysis substantiating the common intuition that task-specific symmetries lead to improved generalization. The authors delve into the interplay between model and data symmetry, particularly in the case of partial or approximate symmetries, establishing a defined link between model equivariance error and data equivariance error. They outline the conditions that lead to the optimal model equivariance error. However, the results are general and not model-specific, with the authors suggesting future research to formulate more specific versions of their theorem tailored to specific groups and neural network architectures.

Strengths

The paper provides a rigorous, quantitative exploration of task-specific symmetries in machine learning models.

Weaknesses

The presentation of the paper's findings could be clearer, particularly with regard to illustrating their novelty and applicability, and the frequent use of different notations without a consolidated reference adds unnecessary complexity and potential confusion for readers.

Questions

The paper's main contributions could be made more explicit. As it stands, the authors appear to enumerate their findings without adequately illustrating their novelty, importance, or applicability. Furthermore, the usage of numerous different notations creates some confusion; a table of notation might ameliorate this issue. While the paper seems to address a fundamental problem in machine learning theory, its presentation could be improved, and the clarity of its contributions could be enhanced. Questions: 1. The majority of results in the paper are upper bounds, but there is no discussion regarding their tightness. Could the authors elaborate on this aspect? 2. In Lines 351-352, the authors mention that "the above bounds are not informative." If the bounds lack informativeness, could the authors clarify the significance or utility of these results?

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

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

1 poor

Contribution

1 poor

Limitations

N/A

Reviewer CLQc7/10 · confidence 4/52023-07-31

Summary

There are two main results in this paper. The first one (Theorem 4.4) obtains a PAC-generalization bound for a hypothesis space which is approximately equivariant to transformations by elements g in G. This bound shows that in addition to the classical complexity term that depends upon the size of the hypothesis space, there are two additional terms that arise, one due to the approximate invariance (where one restricts to only the hypotheses that are equivariant to G) and another due to partial equivariance (where one relaxes the error to equivariance constraint). The proof of this result is classical (e.g., it is close to the Baxter’s proofs in “A Model of Inductive Bias Learning“). The second result (Theorem 5.2) is about studying the interplay between equivariance imposed in the model and that possessed in the data. The key idea here is to construct a set of hypothesis spaces that capture a large subset of the "symmetries" in the data G, depending upon a parameter lambda; larger the lambda, smaller the set of hypotheses. This allows an upper bound on the approximation error in terms of lambda and the dimension of G.

Strengths

This is a thorough and pertinent paper that asks the question: how does the equivariance baked into the model relate to the inherent symmetries present in the data. This is a rare perspective. The proofs are somewhat straightforward but they are well-formulated and rigorous.

Weaknesses

The writing in Section 5 can be improved a lot. I thought that the discussion in Theorem 6.1 could be done in a different way. Specific values of lambda and epsilon entail a choice of hypothesis space. The constant M in Theorem 6.1 therefore is smaller if lambda is larger (?). Minor I suspect the \in S on line 251 is a typo. tilde cal F on line 289 and 298 should be just cal F I think. cal F’ on line 309 should be cal F. There is clearly a typo on Line 310. C_G should be defined before Line 327. Theorem 6.1 should have n^{-1/2} for the first term.

Questions

It would be useful to demonstrate a meaningful experiment (even if it is with synthetic data) where the three terms in Theorem 6.1 trade off against each other.

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

4 excellent

Limitations

N/A

Reviewer R2KT2023-08-10

Reply

I would like to thank the authors for providing satisfactory rebuttal comments. I think this work has significant contribution to the ML community and would like to recommend acceptance. However, more insight on the main result (especially the "generalization-approximation trade-off"/"optimal equivariance") and experimental evidence would make the paper stronger.

Authorsrebuttal2023-08-10

Thank you for the quick ack

Dear Reviewer, Thanks for again taking the time to look at our rebuttal. We are glad that we were able to address some of the question that you had raised. - We will certainly add more discussion around the results, discuss limitations, and add a sub-section explicating further on the "trade-off"/"optimal equivariance." If accepted, we can use the extra page to include these additions. - For the experiments, we will think how to structure them using the paper of Finzi et al. (as we have mentioned in a few comments here) and if accepted, we can include these in the appendix in the camera ready. As a side note: We were primarily thinking of scenarios where the interpretation is somewhat interesting/has not been studied before -- for this we identified a particular problem in NLP that we think is a great illustration of the benefit accorded by partial equivariance -- but were thinking of it in terms of future/ongoing work.

Reviewer PxmR2023-08-11

Thank you for the rebuttal. Most of my concerns are resolved, and I will keep my recommendation as strong accept.

Authorsrebuttal2023-08-12

Thanks!

Dear Reviewer, Thanks for taking the time to read our rebuttal and acknowledging it! We are thankful for your feedback. Please let us know if we can provide any further clarifications or if you have more suggestions vis-à-vis the presentation of our results.

Reviewer W1Je2023-08-17

Thanks for the answers. I still don't see any numerical/theoretical evidence/explanation for the tightness of the upper bounds. An upper bound is only useful when it is tight. Therefore, I still doubt the usefulness of these upper bounds. As the authors said, the bound is useless when $\lambda\to\infty$. Does this imply that the upper bound is pretty loose when $\lambda$ is very small? This also related to my concern that the upper bounds in this paper can be very loose and useless in general.

Authorsrebuttal2023-08-18

Response

Dear Reviewer, Thank you for going over our rebuttal and for taking the time to respond. We would like to take the chance to further clarify why tightness is difficult in such settings (and perhaps not even desirable). To do so, we will first try to emphasize our main contribution and motivation. As we mentioned in some responses, we conceived of the work here as really being the first to address the approximation and generalization tradeoffs in very general equivariant settings. We think of our work's main contribution as establishing intuition and fleshing out some of the important principles at work. As we have noted, existing works deal with generalization in very restricted settings (group --> exact equivariance --> more restricted settings there too etc). Since we can not expect that most real-life situations would be exactly equivariant, we wanted to understand how this fits when we try to work out theoretical estimates. We instead try to cover a full range of generality, and make very weak assumptions vis-à-vis equivariance. With the above motivation, we work with standard PAC theory, using which we prove upper bounds that are distribution independent. As we mentioned to another reviewer, proving tightness would involve building pathological data distributions. Using these then, we would be able to show that the bounds in the worst case are sharp. However, this is not the focus in PAC theory. Building such distributions would require quite technical, abstract, and complicated constructions, which would also be unrealistic. It is not clear if such constructions could shed light on this phenomenon in such a manner that it is useful for applications. Exploring stronger results for specific groups is currently our focus, however. Such results would be sharper. We could also consider making our results data-dependent, although we haven't staked out this direction properly yet. For such stronger results, we are also looking at empirically-driven work as well. However, such work, we believe is out of the scope of this current paper. We appreciate you taking the time.

Reviewer 5syP2023-08-17

I appreciate the effort put in the rebuttal and the clarifications. I do see the value of the presented bound and appreciate the mentioned challenges in deriving tightness results which is something that you might consider adding as limitation in the introduction or concluding remarks or explicit limitations/assumption paragraph or section as suggested by CLQc. Implementing a modified versions of the experiments of reference [1] would be great but I wouldn't call it "required". My initial rating was that the paper is technically solid and I would keep the same rating.

Authorsrebuttal2023-08-18

Response

Dear Reviewer, Thank you for taking the time to go over our rebuttal(s). We are glad that we were able to convey at least some of the value of the presented bounds. We would also like to note that an alternative to some of the directions we mentioned in the response could also possibly include something inspired by https://proceedings.neurips.cc/paper_files/paper/2022/file/5a829e299ebc1c1615ddb09e98fb6ce8-Paper-Conference.pdf this is however, currently not on our radar. Having said that, we will certainly add what you mention to the limitations, which we are planning to add as a separate explicit subsection. Lastly, for the experiments, we have been planning some, and thinking how to fit them in the narrative of theorem 6.1. If we add experiments to the appendix based on [1], we will point to them below theorem 6.1. Although we do think that pursuing empirical work would be more appropriate for stronger results (and as we mentioned in one of the rebuttals, we are pursuing this direction in an NLP context).

Reviewer y3sQ2023-08-22

Thank you for the detailed response and for the promise to add more context and discussion. About the passing from $\mathcal{F}$ to $\mathcal{F}'$, I did meant 'making "less partial" the partial equivariance constraint', and I am happy to hear you plain to explain more in the appendix. My questions have been addressed and I retain my positive opinion of the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC