The paper makes two incremental contributions in using online cluster assignments in self-supervised learning and using multiple crops in different resolutions for data augmentation. When these contributions are combined, decent gains in classification accuracy are obtained. The reviewers raise many issues with the current manuscript, including the discussion of momentum encoder, the discussion of existing clustering-based approaches, and the potential misuse of the term clustering. I ask the authors to incorporate all of these comments in the final version, but I believe the contributions even though incremental in nature, can benefit the fast growing field of self-supervised learning.

Quality: This paper suffers from a few critical issues. Clarity: The experiment setting ups can be described with more details. Sec 3.2 and 3.4 is missing important information such as the datasets used for conducting the experiments. Significance: Although the quality of the proposed model remains unclear because of the previously mentioned critical issues, it's a significant work because it's the first GAN-based model for spectrogram-to-waveform conversion which seems to be working at some degree. It's significantly over-claimed: 1) claiming state-of-the-art for spectrogram-to-waveform conversion (line 6) with MOS 3.09 is surprising; many previous works are at a much higher level (e.g.

The paper describes a successful approach for non-autoregressive spectrogram inversion based on Generative Adversarial Networks. The reviewers noted that even though the results are not at the level of state-of-the-art, the paper addresses a difficult and timely problem, with a convincing experimental validation and ablation study. The rebuttal addressed the main concerns of the reviewers; the authors should nonetheless make sure to address other concerns in the camera-ready version.

As the topic of the this work is very interesting, I'm not frustrated by the presentation, which has a large space for improvement. First, the motivation of the work is not so clear. Why is that an important concern to be able to poison a G-SSL model? The three references [1-3] provided seem at least a decade ago. Are they widely applied in modern security/privacy sensitive applications?

I recommend acceptance of this article subject to some corrections. The novelty of considering data poisoning in GSSL can be interesting for the NeurIPS community. The results presented here are founded by theoretical results and experiments are conducted. It opens new lines of research and can trigger new results. Because the paper has several weak points, we insist on the importance to handle the following comments. However, the overall structure of the presentation is acceptable.

Additional Feedback: The main comment I have regarding the paper is that the authors do not provide adequate justification as to why the advanced data augmentation work compared to the simple ones and when to apply them. This same intuition can be applied for other semi-supervised methods like nearest neighbor and label propagation. These methods will assign the same labels to unlabeled data examples within its component in a graph. This is intuitive but does not explain why the noise from the advanced data augmentation methods are better for semi-supervised learning or provide guarantees for when they work. I acknowledge that I read the rebuttal and thank the authors for providing explanations to the questions and concerns I had.

Summary and Contributions: This paper proposes a contrastive learning algorithm to learn graph representations in an unsupervised manner. It is an extension of SimCLR [1] applied to learn graph representations that can be used for different graph classification tasks, either in semi-supervised learning, unsupervised learning or transfer learning scenarios. To do so, the authors propose several graph augmentation techniques that are needed for the contrastive learning algorithm, and analyse its effects on different types of datasets. The four different types of data augmentation techniques explored in the paper are: node dropping, edge perturbation, attribute masking and subgraph. In their empirical study, the authors explore the effect of these data augmentation techniques in different kinds of graph structure data like social networks and biochemical molecules, showing that different techniques work better on each domain, depending on the nature of the structure represented by the graph. This pre-training technique shows promising results across different datasets and tasks.

The main contribution of this paper is in setting up a 3 player game for semi-supervised learning where the generator tries to maximize the margin of the examples it generates in competition with a classifier the traditional GAN approach of fooling a discriminator. This idea is novel to my knowledge. One small reservation I have with this method is that as the quality of the GAN and generated images increases the margin maximization for the classifier for generated examples becomes counter productive (as acknowledged by the authors) which requires careful early stopping. But this is standard practice with GANs and it should not be held against this paper. The paper is generally of high quality and significance but these could be improved by a broader treatment of related works.

The paper formulates semi-supervised learning as a 3 player game among a generator, a classifier, and a discriminator. The generator and discriminator compete to train realistic examples, as in usual GANs, and the key new idea is that the classifier tries to maximize the margin of real examples and minimize the margin of fake examples. The method both improves predictive performance and greatly reduces training time. The reviewers agree that it is a significant contribution.

Weaknesses: The hypothesis as it stands now is somewhat under-substantiated. Concretely: From a numerical analysis point of view, truncation error order and long-time convergence of the discrete sequence from numerical differencing are separate concepts. RK methods have higher order convergence on fixed time intervals, but its domain of absolute stability is not fundamentally different from that of forward Euler. All explicit methods suffer from limited stability, especially for stiff or conservative systems. Figure 1 shows this effect, but if one takes smaller step sizes the Euler method will converge.