Mind the spikes: Benign overfitting of kernels and neural networks in fixed dimension

The success of over-parameterized neural networks trained to near-zero training error has caused great interest in the phenomenon of benign overfitting, where estimators are statistically consistent even though they interpolate noisy training data. While benign overfitting in fixed dimension has been established for some learning methods, current literature suggests that for regression with typical kernel methods and wide neural networks, benign overfitting requires a high-dimensional setting where the dimension grows with the sample size. In this paper, we show that the smoothness of the estimators, and not the dimension, is the key: benign overfitting is possible if and only if the estimator's derivatives are large enough. We generalize existing inconsistency results to non-interpolating models and more kernels to show that benign overfitting with moderate derivatives is impossible in fixed dimension. Conversely, we show that rate-optimal benign overfitting is possible for regression with a sequence of spiky-smooth kernels with large derivatives. Using neural tangent kernels, we translate our results to wide neural networks. We prove that while infinite-width networks do not overfit benignly with the ReLU activation, this can be fixed by adding small high-frequency fluctuations to the activation function. Our experiments verify that such neural networks, while overfitting, can indeed generalize well even on low-dimensional data sets.

Paper

References (94)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Si3s8/10 · confidence 4/52023-06-29

Summary

This paper generalize existing inconsistency results to non-interpolating models and more kernels to show that benign overfitting with moderate derivatives is impossible in fixed dimension. Moreover, this paper proves that interpolation with spiky-smooth kernels can be consistent and such kernels can be induced by certain activation functions. There are also experiments supporting the authors' claims.

Strengths

The paper is well-written and easy to follow. It also provides us with novel understandings of the benign overfitting phenomenon: the smoothness of the estimators, and not the dimension, matters. In particular, by considering spiky-smooth kernels, the authors find that interpolation with kernels can be consistent in fixed dimension. This new result is very interesting. On the technical side, the authors improve previous inconsistency results for RKHS equivalent to Sobolev RKHS of smoothness $s > d/2$ on the sphere $\mathbb{S}^d$. The techniques here are solid.

Weaknesses

N/A

Questions

1. Can Theorem 5 be extended to $H^s(\Omega)$, $\Omega \subset \mathbb{R}^d$ ? If not, could the authors explain the difficulties here? 2. Is it possible to derive optimal rates for interpolation with spiky-smooth kernels?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have adequately addressed the limitations.

Reviewer GSDW6/10 · confidence 2/52023-07-01

Summary

This work studies the generalization behavior of overfitting methods in terms of the smoothness of the estimator, showing that only non-smooth estimators can interpolate benignly. They give a discussion of this result in the context of NTKs and their corresponding infinite-width architectures.

Strengths

Originality and quality: this work seems to shed some new light on a well-studied problem. I'm not very familiar with background results on asymptotic risk w.r.t. function smoothness, so it's difficult for me to assess the first several theorem statements. I'm quite familiar with the latter results on NTKs, and to my knowledge this observation about spikiness in activation functions is new, and I find it well-explained. Clarity: the paper seems pretty clear and well-written. Significance: the math here is nice, and it seems like this might lead to some more clarity regarding exotic activation functions.

Weaknesses

On significance: it's sort of unclear to me how this matters, or where we go from here. For example: the results of Mallinar et al. suggest that early stopping in nets (or including a ridge parameter in KRR, as shown by prior work too) is enough to make fitting consistent. These questions like "how do we design our model so it overfits benignly?" seem like they're actually missing some motivation -- why do we need to do that? Can't we just use a ridge parameter or optimal stopping? The results around spike-inducing kernels seem quite hacky. The fact that the spike has to get smaller with dataset size makes it clear that this is just effectively adding a ridge. (The authors acknowledge this, but it's still not a surprising conclusion in my view.) It's also not useful -- the authors do not try to train their Hermite-polynomial nets, and (based on my own attempts to do such things) I suspect this may be because they're unstable, even at moderately large width! As stated above, I'm unfamiliar with prior art around function smoothness and overfitting behavior, so there's a fair bit of uncertainty in my assessment here.

Questions

The authors argue that the key object to consider is estimator smoothness, not input dimension. Given that there exists a body of results that do find that input dimension is important, is it clear how to reconcile the author's new results with these old ones? For example, do the smoothness requirements get more relaxed as dimension increases? Note: I think an activation fn corresponding to the Gaussian NNGP kernel can be written down in closed form. (I do not know if this is true for the NTK, and would be curious to know if it can.) Note: a high-frequency sinusoidal activation function can also give you a spiky kernel! Note for future research: this smooth + spiky kernel thing seems like a hack. However, the results of Mallinar et al. suggest that if you could write down any kernel with an eigendecay like $\lambda_i \sim i^{-1} \log^{-\alpha} i$ for $\alpha > 1$, you'd get a consistent estimator without the need to scale down the kernel. It seems plausible to me that by taking a sort of exponent-equals-zero limit of Bietti and Bach's analysis, you could construct a kernel that actually gives you this decay (even on the unit circle), and it'd be consistent while being (a) a continuous function and (b) not needing to be scaled with dataset size. Could be an interesting thing to look into.

Rating

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer jpYR6/10 · confidence 3/52023-07-04

Summary

This paper extends previous results on the inconsistency of ridgeless kernel regression in fixed dimension by showing that non-interpolating estimators whose norm grows comparably to the minimum-norm interpolator are also inconsistent. On the other hand, it is shown that so-called spiky-smooth kernels whose derivatives grow with the number of samples are able to interpolate training data consistently. These results are further specialized to Neural Tangent Kernels and Neural Network Gaussian Processes, and experimental results support the theory.

Strengths

* The paper is generally well-written with understandable and insightful theorem statements and proof ideas, even for non-experts. * Modifying the kernel to achieve consistency in fixed dimension via inspirations from high-dimensional benign overfitting seems novel and interesting. * The assumptions seem to be transparent and clearly stated in the main text.

Weaknesses

* Having explicit rate estimates in addition to asymptotic consistency guarantees could provide further insight into how the spiky-smooth kernels/activations must be designed to achieve optimal performance. * While mathematically exciting and valuable, the results might have limited applicability in modern settings with typically high-dimensional datasets. * The results concerning the NTK can be applied to networks in the kernel regime (a.k.a. "lazy training"), which can only provide a partial picture as there is increasing evidence that in certain settings, neural networks can generalize better in the "feature learning" regime, see e.g. [1]. [1] Malach et al. "Quantifying the benefit of using differentiable learning over tangent kernels." ICML 2021.

Questions

* I believe it can be helpful to add a discussion on the possibility of initializing the weights of a neural network such that training remains in the NTK regime, while the initial predictor is 0, which is an assumption throughout the work. * In Equation (2) of Theorem 1, is there a hidden dependence on $\sigma$ in $c$? In particular, is this intuition incorrect that in the noiseless setting of $\sigma = 0$, consistent estimation ($c = 0$) must be possible? * In line 232, it is mentioned that Section 5 achieves consistency by violating Assumption (N). To my understanding, it seems like Section 5 takes advantage of having the norm (of any estimator) depend on the number of samples. On the other hand, the specific statement of Assumption (N) seems to still hold in Theorem 7, as it is stated for the minimum-norm interpolant (hence $C_\text{norm} = 1$). Minor questions/comments: * In line 88, it is mentioned that training infinitely wide NNs with gradient flow corresponds to learning with the NTK. Perhaps it is worth mentioning that this is only true under a certain scaling that leads to the "lazy training" regime [2], and other smaller initialization scalings that lead to the "mean field/feature learning" regime can have better generalization performance [1]. * In line 95, rotation-invariant kernels are suddenly discussed without prior context. It could be useful to mention their importance given that under standard initializations the NTK is rotation-invariant. [1] Malach et al. "Quantifying the benefit of using differentiable learning over tangent kernels." ICML 2021. [2] Chizat et al. "On Lazy Training in Differentiable Programming." NeurIPS 2019.

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

Please see "Weaknesses".

Reviewer P7716/10 · confidence 4/52023-07-07

Summary

In this paper, the authors studied the problem of benign overfitting for kernels and wide neural networks (in kernel regime) in fixed dimension. The authors showed that benign overfitting is possible if and only if the learner model has large derivatives. This implies that benign overfitting is not possible for those models with small derivatives (including ReLU NTK). On the other hand, they showed that for certain spiky-smooth activations/kernels, benign overfitting is achievable. Experiments are provided in the paper to verify the results.

Strengths

1. The paper is clearly written and easy-to-follow. The proof sketch is given to help the readers to understand the proof easier. 2. Understanding benign overfitting phenomenon is an important problem for deep learning. The current paper focuses on the kernel regression setting and connects to neural networks via the neural tangent kernel (NTK) theory 3. The idea of introducing spiky terms in NTK seem to be interesting. The observation in the proof that such spiky term in NTK behaves like regularization term so that the whole solution is now approximating kernel ridge regression also seems to be interesting.

Weaknesses

1. The current paper focuses on kernel regression and neural networks in the kernel regime. It would be more interesting to go beyond the kernel regime to see if similar results would also hold (e.g., in feature learning regime).

Questions

1. I was wondering what the dependency on dimension $d$ would be in the results such as Theorem 1.

Rating

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The limitation is discussed in the paper. This is a theoretical work and therefore does not seem to have negative societal impact.

Reviewer m3mo7/10 · confidence 3/52023-07-09

Summary

This paper explains that benign overfitting when the dimensionality of data is fixed is possible if one looks for estimators differently from conventional minimal norm. If one allows the estimator to be spiky, benign overfitting can still be possible. These results are extended via NTK to two-layer infinite-width neural networks. The authors showed that a nearly imperceptible sinusoid added to a RELU nonlinearity allows the trained neural network to overfit noisy data without compromising generalization.

Strengths

(Please note that I could not have proofread the submission as well as I would have liked, due to various external obligations.) This paper extends the results on consistency / inconsistency for infinite-dimensional data to the finite dimensional data case. Although the main points - allowing the estimator to go out of its "smooth" way to fit noisy data - could have been argued to be expected, actually seeing this established is still a contribution of remark. The other main point - replacing commonly used non-linearity with one that contains tiny level of sinusoidal perturbation - is not expected by me, that's a very nice way of understanding the non-linearity for neural networks.

Weaknesses

(Please note that I could not have proofread the submission as well as I would have liked, due to various external obligations.) This paper promotes benign overfitting, but it still leaves open whether not allowing overfitting is better or worse than going into benign overfitting, and within the space of benign overfitting how to reach good results, for a given (n,d). In Section 6.2 I felt that the noise standard deviation of 0.25 and the neural tangent kernel of gamma = 1/5000, these two quantities are somehow linked. It could turn out that claiming that such a network can be trained to overfit without regularization might be a dangerous overstatement.

Questions

N/A

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

3 good

Contribution

3 good

Limitations

N/A

Reviewer GSDW2023-08-12

Response to response

I thank the authors for their detailed and thoughtful response. I'm more convinced that these results will prove interesting and clarifying for the community, and I've thus increased my score slightly.

Reviewer P7712023-08-13

Thanks for the response. I will keep my score.

Reviewer jpYR2023-08-18

Thank you for your thorough response. I will maintain my positive evaluation of the work.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC