Beyond NTK with Vanilla Gradient Descent: A Mean-Field Analysis of Neural Networks with Polynomial Width, Samples, and Time

Despite recent theoretical progress on the non-convex optimization of two-layer neural networks, it is still an open question whether gradient descent on neural networks without unnatural modifications can achieve better sample complexity than kernel methods. This paper provides a clean mean-field analysis of projected gradient flow on polynomial-width two-layer neural networks. Different from prior works, our analysis does not require unnatural modifications of the optimization algorithm. We prove that with sample size $n = O(d^{3.1})$ where $d$ is the dimension of the inputs, the network trained with projected gradient flow converges in $\text{poly}(d)$ time to a non-trivial error that is not achievable by kernel methods using $n \ll d^4$ samples, hence demonstrating a clear separation between unmodified gradient descent and NTK. As a corollary, we show that projected gradient descent with a positive learning rate and a polynomial number of iterations converges to low error with the same sample complexity.

Paper

Similar papers

Peer review

Reviewer HQjX8/10 · confidence 5/52023-07-05

Summary

This paper studies the global convergence of gradient descent for training two-layer networks in learning a high dimensional quartic function. The authors show that GD converges when the sample size $n = O(d^{3.1})$ and the width of the neural network grows at most polynomially in d. The authors also showed that any kernel method with $n \ll d^4$ cannot achieve the same order of accuracy.

Strengths

First of all, this is an excellent paper on the non-asymptotic mean field analysis of GD training for two-layer neural networks and its separation from kernel method. Albeit the simplicity of the target function, the convergence analysis requests new ideas and tools. One crucial ingredient of the proof lies in showing that the projected gradient flow (population) dynamics can escape saddle points. The other key ingredient is bounding the coupling error between the empirical particle dynamics and the population dynamics. For the later, the authors tame the usual exponential growth rate in the coupling error with a delicate investigation on the relationship between the growth of the error and that of the signal.

Weaknesses

I do not have many comments on the weakness, but have few questions, which are mostly relevant to future considerations. See below.

Questions

The authors established the results for learning the quartic function. Does the proof generalize to a more general class of functions? If so, can the authors highlight the further work to be done? If not, it would be good if the authors can comment on the major challenges.

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

None

Reviewer ZhLf5/10 · confidence 4/52023-07-07

Summary

This paper studied the projected gradient flow on two-layer neural networks in the mean-field regime with polynomial width and quartic activation function. With data sampled uniformly on the sphere, the authors proved that to learn a single-index model with an even quartic link function, this neural network needs $n=O(d^{3.1})$ training samples. However, we know for any inner product kernel ridge regression requires $n\gg d^4$ to learn a quartic target function. This provides a concrete example of practical feature learning of neural networks outperforming the kernel methods and NTK regime when the target function has a low-rank structure.

Strengths

Overall, I found the paper well-written and easy to follow. The paper presents interesting insights into more accurate models of neural network training. The results obtained by the authors are as sharp as can be reasonably expected given the problem in certain cases. Unlike previous analyses of feature learning with two-stage training procedures, this work shows that neural networks can learn both feature direction and nonlinear link function when training in a more practical way with the projected gradient flow. The methods of the proof may provide further insights to obtain more general results on feature learning.

Weaknesses

1. The main limitation is that the activation function and the target link function are special, only quartic functions. It would be nice to check if the proof techniques can handle more general activation and target functions. Another thing that can be improved is the precise order of the width: how wide the neural network is sufficient for this feature learning. 2. There should be a comparison between the results of the current submission and [10]. [10] uses information exponent to determine the sample complexity for online SGD to learn a single-index target. When $\gamma_2=0$ in Assumption 3.2, [10] implies that sample size should be $n\gg d^{3}\log n$ which is similar to the results in the current submission. Further explanations are needed. 3. And it should be more convincing to visualize Phases 1-3 of the training dynamic in Section 5 if there is some simulation of synthetic data or real-world data for this finite-width training dynamic.

Questions

1. [Ba et al. 2022] also study the feature learning and single-index model when $n\asymp d \asymp m$ with two-stage training processes. And [Arnaboldi et al. 2023] and [Berthier et al. 2023] also studied the gradient flow dynamics of two-layer neural networks in high-dimension to a single-index model. Some comparisons with the current paper could be made. 2. In line 88, you mentioned that a similar coupling error was shown in [49] but [49] is for random feature regression with proportional limit. Can you explain more about it? 3. Line 147 typo. 4. In Theorem 3.4, the sample complexity is $n\ge d^\mu$. Is there an explicit formula or upper bound for the power $\mu$? 5. In Theorem 3.5, you considered the limitation of the inner-product kernel with data points drawn from the unit sphere. Does the data distribution have to be uniform unit sphere distribution? Can Theorem 3.5 be extended to a general rotational invariant kernel or Euclidean distance kernel? 6. Line 233 double "Appendix" 7. Line 274 $w$ should be $w_t$? 8. In the analysis, Phases 1 and 2 are controlled by the quadratic components in the activation and target functions, and the dynamics are analogous to a power method update. Is there any relation between this and the PCA warmup proposed by [Chen and Meka 2020] before SGD training? 9. In Assumption 3.2, we need $\gamma_4\ge 1.1\gamma_2^2$. Why do we need this assumption and coefficient $1.1$? What if $\gamma_2=\gamma_4=1$? =============================================================================================== - Chen, S. and Meka, R., 2020, July. Learning polynomials in few relevant dimensions. In Conference on Learning Theory (pp. 1161-1227). PMLR. - Ba, J., Erdogdu, M.A., Suzuki, T., Wang, Z., Wu, D. and Yang, G., 2022. High-dimensional asymptotics of feature learning: How one gradient step improves the representation. Advances in Neural Information Processing Systems, 35, pp.37932-37946. - Arnaboldi, L., Stephan, L., Krzakala, F. and Loureiro, B., 2023. From high-dimensional & mean-field dynamics to dimensionless ODEs: A unifying approach to SGD in two-layers networks. arXiv preprint arXiv:2302.05882. - Berthier, R., Montanari, A. and Zhou, K., 2023. Learning time-scales in two-layers neural networks. arXiv preprint arXiv:2303.00055.

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

3 good

Contribution

3 good

Limitations

I think the authors have adequately addressed the potential negative social impact of their work. The conclusion lays out some suggested next steps and in doing so highlights certain current limitations.

Authorsrebuttal2023-08-19

Please let us know if you have any questions!

Dear reviewer, since the end of the discussion period is approaching, we just wanted to check if we have addressed your questions, or if you have any additional questions. Thank you very much!

Reviewer c2Ri7/10 · confidence 4/52023-07-08

Summary

This paper studies the statistical efficiency of the projected gradient dynamics on the sphere for (polynomial-width) two-layer neural networks under the mean-field regime. In particular, this work proves the sample complexity of $O(d^{3.1})$ for learning the single-index model with an unknown quartic link function.

Strengths

The results can be of interest to the readers and are technically sound. The remarkable contributions compared to most related works are that this work - studies not unnatural optimization dynamics but the standard projected gradient flow on the sphere, - shows the separation between two-layer neural networks in the mean field regime and kernel methods by providing the sample complexities. Technically, the improvement in particle complexity is also worth noting. In a naive way, the approximation accuracy to the distribution by using finite particles exponentially deteriorates in time, but this study prevents such deterioration by making use of the problem and model structures. Hence, I think this work certainly contributes to the context.

Weaknesses

- This theory deals with not standard activation such as ReLU, sigmoid, and tanh but a special type of activation function (i.e., fourth-order polynomial). - In my understanding, Abbe et al. (2023) [2] also studies the standard optimization dynamics and shows the superiority of the mean-field regime against linear models. A detailed discussion of the relationship to [2] would help clarify the position of the work.

Questions

Could you specify the dependence of $m$ stated in the next sentence of Lemma 5.1: ``network width $m$ is sufficiently large polynomial of $d$?'' What does the highest degree of this polynomial depend on?

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

The limitation of the paper is well addressed.

Reviewer aN4G5/10 · confidence 3/52023-07-11

Summary

Analyzed the (projected) gradient flow dynamics of a two-layer neural network in the mean-field regime in learning a specific degree-4 single-index target function. The main contribution is a polynomial-time convergence guarantee and a sample complexity that outperforms kernel methods. This differs from the naive mean-field analysis where the network width can grow exponentially.

Strengths

Most existing mean-field analyses of two-layer neural network focus on the optimization aspect, with the goal of showing global convergence of the training loss, but the generalization properties and sample complexity in learning certain class of target functions are largely unknown. The only exception that I know of is (Abbe et al. 2022). So this submission definitely tackles a challenging and interesting problem. On the technical side, the analysis differs from many previous works on learning single-index model with two-layer neural network, where the link function mismatch is typically handled by training the second-layer parameters.

Weaknesses

I have the following concerns. 1. The problem setting is restrictive and convoluted, which is rather underwhelming given the promise of a *clean mean-field analysis* with no unnatural modifications. (i) The training algorithm is the unmodified projected gradient flow, but time discretization is not studied. (ii) The target function is assumed to be an even degree-4 function which is quite restrictive. Moreover, the trained neural network also uses a somewhat unnatural degree-4 activation function, and the Legendre coefficients are restricted by Assumption 3.2. The fact that this submission is lengthy and technically demanding despite these strong assumptions raises the question of whether the same analysis can be extended to more general problem settings. 2. The comparison against prior works is not sufficient, and consequently, the significance of the results cannot be easily evaluated. * In the context of learning single-index model, the sample complexity of gradient descent is decided by the information exponent from (Ben Arous et al. 2021). This definition is originally defined in the analysis of online SGD, but the authors should discuss whether the same mechanism is present in the ERM setting. Specifically, my current reading is that this submission assumes an information exponent of 2, due to the factor of $\sigma_2^2\gamma_2$ in the denominator of the stopping time $T_*$ in Theorem 3.3. Please clarify if this is the case. (Ben Arous et al. 2021) *Online stochastic gradient descent on non-convex losses from high-dimensional inference* * Related to the previous point, it is also well-known that trained two-layer neural network can outperform kernel models when the link function is a high-degree polynomial with low information exponent. For example, the staircase functions in (Abbe et al. 2022) can be learned in linear sample complexity. While these prior results typically employs a layer-wise training procedure, the similarity in the main message may undermine the significance of the theoretical results in this submission. Therefore, I feel that the authors should use more space to highlight the technical differences and challenges (no bias units and retraining of second layer, etc.). (Abbe et al. 2022) *The merged-staircase property: a necessary and nearly sufficient condition for sgd learning of sparse functions on two-layer neural networks* * The simplification of population dynamics into some low-dimensional object via symmetry has appeared in many prior works, such as the dimension-free PDE in (Abbe et al. 2022) and (Hajjar and Chizat 2022). The authors need to explain the difference and new ingredients in Section 4.1. (Hajjar and Chizat 2022) *Symmetries in the dynamics of wide two-layer neural networks* * The claim that prior mean-field analysis only proved "generic exponential growth rate in the coupling error" is not entirely accurate. For the noisy gradient descent setting, recent papers have provided uniform-in-time propagation of chaos estimates, for example see (Chen et al. 2022) (Suzuki et al. 2023). These results hold for finite-width network and discrete-time algorithm, but the current submission only handles the finite-width error. It would be a good idea to comment on the possibility/difficulty of handling the discrete projected gradient descent algorithm. (Chen et al. 2022) *Uniform-in-time propagation of chaos for mean field Langevin dynamics* (Suzuki et al. 2023) *Convergence of mean-field Langevin dynamics: Time and space discretization, stochastic gradient, and variance reduction*

Questions

I would be happy to update my evaluation if the authors could address the concerns and questions in the Weaknesses section. #####################**Post-rebuttal Update**##################### The authors addressed some of my concerns; I have therefore increased my score to 5. In the revised manuscript, please include a detailed comparison against relevant prior works, as done in the rebuttal.

Rating

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

See weaknesses.

Reviewer c2Ri2023-08-18

reply to authors

I appreciate the author's response. The authors have adequately addressed my concerns. I will keep the evaluation.

Authorsrebuttal2023-08-18

Minor Clarification

We wish to add a minor clarification to the "Precise Width" section of our rebuttal - this is not related to any of the other points that we make in the rebuttal. We just wanted to clarify that the width $m$ does not depend on the Legendre coefficients of the activations - we incorrectly said in the rebuttal that it has some factors depending on the Legendre coefficients. So the width $m$, up to $(\log d)^{O(1)}$ factors, is $d^{3 + \gamma}$ where $\gamma > 0$ is an arbitrarily small constant. We note that this change does not affect our submitted manuscript - in our actual submitted manuscript, we correctly calculated that the width only depends on $d$ and does not depend on the Legendre coefficients of the activation.

Authorsrebuttal2023-08-18

Minor Clarification on Precise Width

We wish to add a minor clarification to the "Precise Width" section of our rebuttal - this is not related to any of the other points that we make in the rebuttal. We just wanted to clarify that the width $m$ does not depend on the Legendre coefficients of the activations - we incorrectly said in the rebuttal that it has some factors depending on the Legendre coefficients. So the width $m$, up to $(\log d)^{O(1)}$ factors, is $d^{3 + \gamma}$ where $\gamma > 0$ is an arbitrarily small constant. We note that this change does not affect our submitted manuscript - in our actual submitted manuscript, we correctly calculated that the width only depends on $d$ and does not depend on the Legendre coefficients of the activation.

Reviewer HQjX2023-08-19

Keep my rating unchanged

I have read the rebuttal. Thank the authors for addressing my comments. I keep my rating unchanged.

Reviewer ZhLf2023-08-21

Thanks for the rebuttal

Thanks for the authors' detailed response. I thank the authors for their clarifications in the rebuttal and appreciate the theoretical contributions of this work. But considering the model assumptions and the lack of a tight sample complexity for GD, I will keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC