NeurIPS 2020

Over-parameterized Adversarial Training: An Analysis Overcoming the Curse of Dimensionality

Review 1

Summary and Contributions: 1. The paper proves that for two layer neural networks, adversarial training finds a robust model near the initialization. 2. The paper also provides a new approximation theory, which shows that the step function can be approximated well by a polynomially wide ReLU network whose weights are close to the initialization.

Strengths: 1. The question of how over-parameterized a network should be, so that it is adversarially robust, is really important. Answering this question will help in the design of robust networks. This work takes a step in that direction and improves upon some aspects of previous work. 2. The proof of the claims seem correct and are really well explained.

Weaknesses: 1. The $\gamma$-separability requirement seems odd in the sense that it also requires samples from the same class to be separated by a margin. Earlier work by Gao et al.(2019) does not seem to have this requirement. This seems artificial and probably the analysis can be tightened to fix this. Would replacing such close samples by just one 'representative' sample from the same class help? 2. The previous work by Gao et al.(2019) and this work, both only guarantee robustness against the adversary used during training. Can something be said about the robustness against the most intelligent adversary $\mathcal{A}^*$, for a network trained using a polynomial time adversary $\mathcal{A}$? This seems to be much more important question than whether a network trained using a particular adversary can defend against the same adversary. ------ Post author feedback comments ------- My concerns have been satisfactorily answered in the feedback.

Correctness: The proof overview (section 5) and proof of Theorem 5.3 (section 6) seems to be correct and are well explained.

Clarity: The paper is really well written. It was easy to understand the intuitions and the flow of the proof.

Relation to Prior Work: The relation to prior work has been discussed well and differences have been highlighted.

Reproducibility: Yes

Additional Feedback: Please see the Weakness section above.

Review 2

Summary and Contributions: This paper follows a recent line of work that analyzes the convergence properties of adversarial training. The authors show that adversarial training always converges in polynomial time on two-layer ReLU networks assuming the separability of training data. The general idea is to construct a robust net near initialization using polynomial approximation.

Strengths: This paper is of interesting topic, clear problem formulation and important results. The authors makes a detailed analysis and proof of the convergence properties of adversarial training in polynomial time on the two-layer ReLU network by using polynomial approximation, which significantly improves the previous theoretical results. The logic of the whole paper is clear and the proof part is easy to understand. The convergence property of standard adversarial training is clearly revealed. The proof in the supplementary material is detailed and the conclusions are clear. In addition, the new results of approximation theory can be further applied to the theory of over-parameterized model.

Weaknesses: I don’t have much negative observations about this paper.

Correctness: There is no obvious error as far as I can see.

Clarity: The full paper is clear and easy to follow.

Relation to Prior Work: The authors introduce adversarial example and defense, convergence of adversarial training and polynomial approximation in the related work part. I think the introduction to the prior work is sufficient.

Reproducibility: Yes

Additional Feedback: The paper could benefit from presenting the principle of adversarial attack in the related-work section. In addition, although the main contribution of this paper lies in the theoretical analysis of the convergent properties of the adversarial training, simple experiments on some small datasets may further verify the conclusions of this paper and demonstrate the correctness of the conclusions more intuitively. ========================== After rebuttal: Thanks author for putting the updated results. It does solve some concerns to me. This is a promising submission and I would maintain my original score of 7.

Review 3

Summary and Contributions: The paper studies the convergence theory of over-parameterized adversarial training. Based on Wang et al.'s work, it achieves further results: it proves the convergence to low robust training loss of two-layer ReLU activated neural network in standard adversarial training and for polynomial width and running time of the input dimension d.

Strengths: The paper gives a proof overview of the main result Theorem 4.1, so it is easy to understand. The assumption, gamma-separability is very clever. It helps the proof a lot. Also, its rationality is verfied empirically in popular dataset. Settings are more realistic. It solve the curse of dimensionality (which is a future work of Wang et al.'s work) perfectly.

Weaknesses: Although this paper has an obvious improvement on Gao et al.'s work, I have to say that it lacks novelty, and the contribution is small. Width, runing time and activation funtion are not a huge gap in Gao et al.'s work. What I really want to see is to remove projection in deep net, improve the online learning or analysis the robust generalization. However, both the results and the proof methods have little inspiration. This paper claims that '(Gao et al.) require the width of the net and the running time to be exponential in input dimension d, and they consider an activation function that is not used in practice' in abstract. This is likely to cause a serious misleading. In fact, the main results of Gao et al.'s work, Theorem 5.1 and 5.2 only requires polynomial of d with a seldom used activation function. However, it is quadratic ReLU activation, which is a common activation function, that the width and running time are exponential with. Gao et al. put the exponential case in appendix C.2 (instead of C.1, this paper makes a misleading mistake in section 1) in order not to use the Lipschitz assumption. These words may make readers think Wang et al. require both exponential width and seldom used activation. ---------------------------------------------------------- Post rebuttal: The authors' rebuttal solves some of concerns, and I would raise my score to weak accept.

Correctness: I have carefully checked the results and proofs of this paper, and ensure that it is correct.

Clarity: Its writing is fluent and there are no obscure places. The structure is clear and concise. It also shows some important proof technique in advance.

Relation to Prior Work: I am afraid not. Although authors show they have a good understanding of prior work's (especially Wang et al.'s work) contributions and defects between the lines, as said above, the expression in some places may leads a serious mistake.

Reproducibility: Yes

Additional Feedback: The assumption, gamma-seperable can be verified on a large dataset, such as Imagenet. It would be better if it could give a stronger intuitive feeling.

Review 4

Summary and Contributions: The paper presents a new theoretical result on adversarial training of two-layer ReLU networks, more concretely, it shows that for networks of polynomial width the adversarial training algorithm converges to arbitrarily small robust training loss.

Strengths: I find the result interesting and an important step to understanding adversarial training, since the conditions hold in practical scenarios (e.g., CIFAR10) and for networks of polynomial widths (as opposed to exponentially wide networks, considered before).

Weaknesses: While the main focus of paper is theoretical, I feel like it could benefit from small toy examples experimentally demonstrating the authors' claim. E.g., plot the dependence of robust training loss vs depth.

Correctness: Correct.

Clarity: The paper is well written. I suggest the authors to add some explanation of the logic behind the definition 6.1. Additionally, some visualizations (as noted before) could simplify the understanding. For instance, a visualization explaining definition 3.4 could be helpful.

Relation to Prior Work: Related work is thoroughly reviewed.

Reproducibility: Yes

Additional Feedback: Please see the previous comments.