The Exact Sample Complexity Gain from Invariances for Kernel Regression

In practice, encoding invariances into models helps sample complexity. In this work, we tighten and generalize theoretical results on how invariances improve sample complexity. In particular, we provide minimax optimal rates for kernel ridge regression on any manifold, with a target function that is invariant to an arbitrary group action on the manifold. Our results hold for (almost) any group action, even groups of positive dimension. For a finite group, the gain increases the"effective"number of samples by the group size. For groups of positive dimension, the gain is observed by a reduction in the manifold's dimension, in addition to a factor proportional to the volume of the quotient space. Our proof takes the viewpoint of differential geometry, in contrast to the more common strategy of using invariant polynomials. Hence, this new geometric viewpoint on learning with invariances may be of independent interest.

Paper

Similar papers

Peer review

Reviewer DGX17/10 · confidence 3/52023-07-05

Summary

This paper analyzes the impact on sample complexity of encoding invariances into kernel functions in the context of kernel ridge regression. A kernel function is invariant to the actions of a group if its output does not change when its inputs are acted on by members of the group. The paper shows that for finite groups, the sample efficiency is effectively multiplied by the size of the group. For groups of positive dimension, it shows the sample efficiency is improved by (1) reducing the effective dimension of the input manifold, and (2) reducing the volume of the manifold.

Strengths

- Significance: This paper analyzes how invariance improves sample complexity for kernel ridge regression in a much more general setting than prior work. For example, it allows the input space to be any compact manifold, and the invariance to be represented by any smooth compact Lie group (prior work, for example [5], had assumed the input manifold was a sphere and that the invariance was to a finite group). - Quality: Although I have not checked the proofs, the quality of the work appears high. The generalization bound in Theorem 3.1 is proved to be minimax optimal in Theorem 3.3, an interesting/important result. - Clarity: The paper is generally well written and clear. See “Weaknesses” section for places for improvement. - Originality: It seems the approach of using differential geometry to analyze these generalization bounds is original, and that the authors developed some tools of independent interest in this area. Overall, I believe this paper makes an important theoretical contribution to a problem space that has been of significant interest to the machine learning community: namely, how to make models invariant to certain input transformation (either via data augmentation, or by changing model architecture), and understanding how encoding these invariances impacts the models’ generalization performance and sample complexity.

Weaknesses

Perhaps the greatest weakness of this paper is at times seeming too “mathy” (https://arxiv.org/pdf/1807.03341.pdf), without sufficient grounding in concrete examples or empirical results. For example: - It would have been useful to go over more examples of invariances whose groups have positive dimension, and understand more concretely how these invariances affect the generalization bound. - It could be useful, for equations (5) and (7), to give examples of how this quantity changes for different finite vs. infinite (positive dimension) groups. One example I had in mind, for $x \in R^2$, was rotational invariance, where the rotations could be in a finite group (e.g., rotate by multiples of $\alpha$ degrees, for any $\alpha$ that evenly divides 360) or an infinite group (e.g., rotate by $\beta$ degrees for any real-value $\beta \in [0, 360)$). How would these situations compare, in terms of equations (5) and (7), as the value of $\alpha$ approaches 0? - It would have been helpful to understand which of the examples given would not have been possible to analyze with prior results. - It would have been useful to run experiments (even toy/synthetic experiments if necessary) to demonstrate that the bounds from the paper are actually predictive of model performance. NIT: $\omega_d$ is defined in Theorem 3.4, but used in prior theorems.

Questions

See the weaknesses section.

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

4 excellent

Limitations

Perhaps a broader discussion of limitations would be useful. - What deep learning scenarios would the results from this paper apply to or not apply to? - What important open problems remain to be solved? - How empirically predictive are these theoretical results, in different settings? In what cases would these results definitely not be empirically predictive?

Reviewer doFr7/10 · confidence 3/52023-07-06

Summary

The paper analyzes the generalization error of kernel ridge regression on manifolds with a kernel invariant to a group action. Compared with previous work, the results hold more generally for groups of positive dimension.

Strengths

The paper tackles an important issue, is well-written, with well-chosen and appropriate examples, and appears technically sound. It also offers a differential geometric perspective on learning theoretic issues which might have its own interest.

Weaknesses

Though already significant in its present form, the paper would gain in interest if it could provide more "agnostic" generalization error bounds in various ways, or at least comment on how (and if) this would be possible. - Assumptions on f*: While interesting per se, the results are inapplicable to practical cases without knowledge of the target function f* (and the noise level sigma^2). In particular, Theorem 3.1 makes strong assumptions on the target function f* and holds only for the optimal value of the regularization parameter. It would be much stronger if it could be reformulated to hold more broadly for all regularization parameters with, e.g., the norm of \hat{f} instead of that of f*, and with f* only involved through the empirical risk (like more "traditional" error bounds based for instance on the Rademacher complexity). In other words, if invariance to a (finite) group G intuitively implies that "each data point is worth |G| data points", does it also imply something like "the complexity of a model class taking into account the invariance is the standard complexity of the model class divided by (the square-root of) |G|" (where the complexity can refer for instance to the Rademacher complexity)? - Assumptions on G: the results depend on knowledge of the group G. What if G is only approximately known? For instance, what if the kernel K is invariant to G' while f* is invariant to G with G' slightly different from G (or if G' contains only a subset of the transformations in G)? - How much of the result relies on the uniform distribution of x on the manifold? Could you comment on what terms would be affected by a change of measure in Theorem 3.1 and if the gain as G grows remains of the same order in this case?

Questions

No question

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

No limitation identified.

Reviewer xLSq8/10 · confidence 3/52023-07-10

Summary

In this work the authors theoretically study how encoding invariances into models improves sample complexity. They approached the problem from differential geometric viewpoint, rather than common strategy of using invariant polynomials. Since the problem is algorithm and model dependent, the authors considered kernel-based algorithms, as neural networks in certain regimes behave like so. The obtained results generalize and greatly expand previous state-of-the-arts. Hence, the results provide a reduction in sample complexity that was not possible with previous assumptions. The paper also shows how these results transfer to popular invariant models, such as DeepSets, GNNs, PointNet, and SignNet.

Strengths

- The paper discuss one of the major challenges in machine learning, i.e., sample complexity. - This work provides theoretical results on how incorporating invariance in model can improve sample complexity. Their results are more general and can achieve better bound. - Well written.

Weaknesses

- None

Questions

- None

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

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

- Theoretically intense, so for some readers, might be hard to follow.

Reviewer jgVU5/10 · confidence 4/52023-07-30

Summary

This article investigates the sample complexity gained from encoding invariances into learning models. The article focuses on the study of kernel ridge regression on compact manifolds for functions that are invariants to a group action on the manifold. The main result of this article (Theorem 3.1) gives an upper bound of the excess population risk for functions living in the intersection of Sobolev spaces and the set of G-invariant square-integrable functions on the manifold. The analysis shows the importance of two terms: the volume of the quotient space, and the effective dimension of the quotient space. This result quantifies the exact sample complexity gain from invariances for kernel regression on compact manifolds for an arbitrary Lie group action. Moreover, the authors prove that the KRR estimator is minimax optimality. The proofs of the results make use of new results in the field of differential geometry.

Strengths

- Although heavily technical, the article is well written and the proven results were extensively discussed. - The newly proved results may be beneficial for the community of kernel methods on manifolds.

Weaknesses

To be honest, I didn't have enough time to check the proofs thoroughly. Given that the contribution is purely theoretical, I believe it will be better suited for a math-oriented journal, where a rigorous scrutiny of the proofs can be ensured.

Questions

-

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

-

Reviewer DGX12023-08-17

Thank you for the thoughtful responses, I look forward to seeing the next draft!

Reviewer xLSq2023-08-18

Response to rebuttal

I want to thank the authors for the rebuttal. I still think the paper addresses an important problem and the solution proposed is theoretically sound. I appreciate the authors intent to add figures in the main paper to help explaining the problem and contribution better. I will like to keep my rating as strong accept.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC