NeurIPS 2020

Instance Selection for GANs

Meta Review

Reviewers were almost unanimous in voting to accept this paper, and I think overturning the reviewer decisions should be done very cautiously, so I will recommend acceptance here. However, I have a serious problem with this paper that I really hope that authors will address: Your results are interesting, but you are wrong about why they are interesting! The reviewers have discussed this at length, but just to summarize, you have not shown a technique for actually making GANs better at doing the thing that people want them to do, you have instead shown why the way the people evaluate GANs is all wrong. I think it would be a real disservice to the community (and cause a bunch of unnecessary confusion) to not change the narrative in the paper to reflect this. If you don't make this change, I will be very sad, and I bet the reviewers will be sad also, and besides, from a selfish perspective, I think the paper will be better received if you make this change anyway.