NeurIPS 2020

Top-k Training of GANs: Improving GAN Performance by Throwing Away Bad Samples

Meta Review

Three reviewers recommend weak accept, while one recommends weak reject. One the one hand, reviewers found the extreme simplicity of the proposed approach appealing and the gains over baseline GANs compelling. On the other hand, many reviewers viewed the novelty as relatively incremental. Additionally, multiple reviewers requested stronger analysis -- e.g. How does the proposed approach affect diversity? -- and comparison with related techniques. Author response largely addressed these questions and concerns. I agree that the simplicity of the proposed method is a strength, rather than a weakness, so long as sufficient analysis is provided. Given the author response, I recommend acceptance. However, I strongly encourage authors to take reviewer feedback seriously and attempt to address all their concerns in camera ready.