NeurIPS 2020

Your GAN is Secretly an Energy-based Model and You Should Use Discriminator Driven Latent Sampling


Meta Review

Two reviewers strongly recommend acceptance, while two recommend reject. R1 and R4 argue that the proposed connection between energy networks and GANs is new and will be of general interest to the community, and the results on CIFAR-10 are sufficient to demonstrate the utility of the proposed approach. R2 and R3 argue that the results are unconvincing because important evaluations are missing (FID score) on CelebA. I take this point, and encourage authors to add these results. Further, R3 points out that Figure 5 doesn't actually demonstrate MCMC mixing as claimed. However, overall I agree with R1 and R4 that the positives outweigh the negatives. I recommend acceptance, but strongly encourage authors to take reviewer suggestions seriously and implement changes for camera ready. The wording around MCMC mixing and Figure 5 should be clarified in camera ready to avoid confusion.