Provable Guarantees for Neural Networks via Gradient Feature Learning

Neural networks have achieved remarkable empirical performance, while the current theoretical analysis is not adequate for understanding their success, e.g., the Neural Tangent Kernel approach fails to capture their key feature learning ability, while recent analyses on feature learning are typically problem-specific. This work proposes a unified analysis framework for two-layer networks trained by gradient descent. The framework is centered around the principle of feature learning from gradients, and its effectiveness is demonstrated by applications in several prototypical problems, such as mixtures of Gaussians and parity functions. The framework also sheds light on interesting network learning phenomena such as feature learning beyond kernels and the lottery ticket hypothesis.

Paper

Similar papers

Peer review

Reviewer Sj277/10 · confidence 3/52023-07-01

Summary

This paper proposes a general framework for analyzing feature learning in two-layer ReLU neural networks. The idea is to consider the class of two-layer ReLU networks with “gradient features”, i.e. features aligned with the gradients of the loss induced by the distributions of the data and initial model parameters. The main result (Theorem 3.12) is that gradient descent on the two-layer ReLU network achieves generalization error close to the that of the optimal model in this class under weak assumptions. Instantiations of this result are provided for the special cases of the data being generated by a mixture of gaussians and parity functions. In these settings, the optimal loss among models in the gradient feature class is computed, with final generalization rates strictly improving over kernel methods and matching or surpassing the best known results in the feature learning literature.

Strengths

1. The paper analyzes a highly relevant topic. 2. The idea is simple, yet powerful and novel, to my knowledge. The proposed notion of gradient features can likely be used to show feature learning guarantees for models beyond two-layer ReLUs. At present, it allows for competitive feature learning guarantees with very general boundedness assumptions on the data distribution and smoothness of the loss function. 3. The two instantiations of the main result are very helpful to concretize the significance of the main result. Both are very well-studied settings, and the provided guarantees strictly improve upon kernel methods in both cases, and match the rates for learning parities albeit with more general assumptions. 4. The presentation is clear and rigorous. 5. The related works are well-covered.

Weaknesses

1. On its own, it is difficult to gauge the significance of the main result, since in general we do not know whether gradient features can lead to a good model. Indeed, it would be helpful to have more discussion of when gradient features can lead to a good model beyond the two special cases. 2. Related to above, a result showing that if the optimal model in the gradient feature class has large risk, then the ground-truth mapping from inputs to labels is not learnable by gradient descent would strengthen the paper. 3. Like other feature learning studies, the analysis leverages feature learning happening in the early steps of gradient descent, while subsequent updates are shown only to not corrupt the learned features, although it is unclear how well this aligns with practice.

Questions

Minor note: the restrictions on $\tau$ should appear in the statements of Theorems 4.4 and 4.8.

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

4 excellent

Limitations

Yes

Reviewer H9VP5/10 · confidence 3/52023-07-06

Summary

This paper proposes a general framework of feature learning for two-layer neural networks trained with gradient descent. This framework covers a variety of classification losses and data distributions. Specifically, the authors establish that the loss of neural networks trained with gradient descent is comparable to that of networks with first layers in the direction of a simplified gradient and optimal second layers. The framework is then specialized to multiple examples, including Gaussian mixture classification and learning parity functions.

Strengths

* The presented framework is rather general and unifies the approaches of several recent works, while covering distributions that were not previously covered in the literature. * The studied problem is exciting and significant for the community as the existence of such a framework is necessary given the ad-hoc nature of current feature learning analyses. * The related literature is covered extensively.

Weaknesses

* The general guarantees provided by the framework (Theorem 3.12) are limited to the performance of the optimal network with first-layer directions approximately aligned with the gradient. Such a statement does not directly imply feature learning for general problems/distributions, as a key part of feature learning seems to be showing that the gradient directions are indeed useful. To that end, the guarantees are more or less similar to prior work and suboptimal in some cases (e.g. Gaussian mixtures). * The sample complexities of Theorems 4.4 and 4.8 are only stated as polynomials in problem parameters, thus while outperforming kernel methods, provide limited insight in comparison with prior works. Furthermore, the $\tilde{O}(d^{1.5})$ sample complexity to learn an XOR label from a mixture of 4 Gaussians in Section 4.1.1 seems to be suboptimal, as [92] shows a sample complexity of $O(d)$ for learning the same problem. * The presentation of the paper can be improved, with specific examples given below.

Questions

* Is it possible to characterize the reason behind the suboptimality in the $\tilde{\mathcal{O}}(d^{1.5})$ sample complexity of this work in comparison with the $O(d)$ sample complexity of [92] to learn a mixture of 4 Gaussians with XOR-like structure? * A number of figures can be added to improve the presentation of the paper, e.g. for the definition of gradient features and introducing the Gaussian mixtures and the input distribution for learning parity functions. * Including the choices of step size in Theorem 3.12 can provide additional insights. Specifically, it seems like the theorem is still operating in the regime of a few gradient steps for the first layer, similar to prior work. A more explicit discussion of the training of the first layer would clarify the similarities and distinctions with prior work.

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

2 fair

Contribution

3 good

Limitations

The authors have mostly discussed the limitations of their work. It would help the readers better understand the limitations if there is also a discussion on how realistic the training assumption for the first layer is (i.e. only a few steps or many steps), and also the potential suboptimality in the rate of learning mixtures of Gaussians in comparison with prior work.

Reviewer 3J3m7/10 · confidence 3/52023-07-07

Summary

This paper introduces a general framework for studying feature learning in two-layer NNs. This framework covers feature learning in different examples such as linear classification, mixture of Gaussians and parity functions (it also gives some intuition about the learned features). The neural network under study has one hidden layer and is initialized in a symmetric manner. The training algorithm used is mini-batch GD with fresh samples. In the first iteration, features are learned in some hidden neurons. During the next iterations both layers get updated, however the change in the hidden layer is controlled.

Strengths

- The paper does not freeze the first layer's weights, instead it trains all parameters together and controls the change in the first layer's weights with small learning rate. - The literature review is quite extensive; the paper can generally help people interested in feature learning to become familiar with this direction of research. The paper's framework has also been applied to several different learning problems. - I personally liked the connection to the lottery ticket hypothesis. I'd suggest moving (more of) it to the main.

Weaknesses

- There are a few limitations which are quite common in the current literature of deep learning theory and have been discussed in the paper: e.g., the feature learning is actually done in the first iteration of gradient descent. See below for some questions regarding the limitations of the framework. - Personally I wonder if there are new insights given by this framework or not? (Also see below)

Questions

- Q1. Is it possible to extend the results to the continuous setting and $\ell_2$ loss function (at the beginning $\ell_2$ also acts like $-y\hat{y}$) - Q2. What are the general insights given by the framework (other than the unification)? - Q3. Is it possible to use other activation functions? - Q4. Why in Theorems 33 and 46 the algorithm that trains both layers is not used? Suggestions: - It would be really great if some intuition about the gradient feature learning framework can be provided (maybe even in the abstract). - The parity formulation in section 4.2 was difficult to understand for me; maybe some changes would help.

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

This is a theoretical work and there is not negative societal impact. I think the limitations have been discussed adequately (nonetheless see above).

Reviewer zxbj6/10 · confidence 3/52023-07-10

Summary

The paper defines the concept of "gradient features" which capture the features that the network can learn after one step of gradient descent. The paper then instantiates this framework to prove optimization and generalization guarantees for various statistical learning problems.

Strengths

- The paper develops a general framework which formalizes the "one step" feature learning trick which has recently become popular. This provides an easy to use framework which makes it easy to derive sharp sample complexity guarantees for a number of well defined statistical learning problems. - The paper instantiates this framework in a number of settings (mixtures of Gaussians, parity, multi-index models) and re-derives a number of sample complexity results.

Weaknesses

- It appears that this framework can only handle features that can be learned directly at initialization (i.e. through the "one-step trick"). In particular, it seems unable to handle multi-step feature learning (e.g. the merged staircase property in [1]). [1] 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"

Questions

- What is the role of the $s$ in the gradient feature? It seems to encode the sign of $b$ but I don't immediately see why this is important. - For simple $k$-parity, my understanding is that the learned gradient feature $D_1$ is $1_{A}$, the indicator function for the subset $A$. This would therefore compute parity by computing the parity of $\sum_{i \in A} x_i$. Is this correct?

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

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

The authors adequately addressed the limitations of this paper.

Reviewer H9VP2023-08-13

Thank you for your detailed response and for the nice figure! I believe that this work presents an effective framework to unify recent approaches, however, I'm keeping my original score as I'm not entirely certain how much of the high-level intuitions obtained from this work are significantly novel in comparison to the existing literature.

Authorsrebuttal2023-08-13

Thank you

We appreciate that the reviewer believes our work presents an effective framework to unify recent approaches. We will update our work with new bound for the 4-mixture-of-Gaussians, more figures, and more insights. Thank you for your time and valuable suggestions!

Reviewer zxbj2023-08-15

Thank you for the clarifications. I have decided to keep my score.

Reviewer 3J3m2023-08-16

Thanks for the response. I think the paper can benefit from including some of these discussions. I will keep my score.

Authorsrebuttal2023-08-16

Thank you

Thank you for your positive and helpful feedback! We will make sure to include these discussions in the revision.

Reviewer Sj272023-08-18

Thank you to the authors for your response. I have decided to maintain my score, and encourage the authors to add the failure cases to the final version, perhaps in the appendix.

Authorsrebuttal2023-08-18

Thank you

Thank you for your comments and suggestions! We will add the failure cases to our revision. Thank you for your time!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC