Summary: Authors present a new stage-wise learning paradigm for training generative adversarial networks (GANs) called RankGAN that progressively strengthen the discriminator (and thereby, the generators) with each stage. To train the multiple stages of RankGAN, the authors propose a margin-based loss for the discriminator and extend it to a margin-based ranking loss.
+The novelty of the work is in the RankGAN framework which consists of a discriminator that ranks the quality of the generated images from several stages of generators. The ranker guides the generators to learn the subtle nuances in the training data and progressively improve with each stage.
+ a margin-based loss function for training the discriminator in a GAN;
+ a self-improving training paradigm where GANs at later stages improve upon their earlier versions using a maximum-margin ranking loss; and
+ a new way of measuring GAN quality based on image completion tasks.
+ The writing seems technically correct and adequate clarity
– The experiment section demonsrates perforamnce comparison between RankGAN with WGAN only. But, to further establish the efficacy of the method, a comparison with more state-of-the-art should have been performed
– Some minor mistakes in sentence structures found. Abstract is missing word and punctuation. Authors should address this.
The proposed stage wise RankGAN method and the Margin Loss and Rank Margin Loss are novel concept, experiment on CelebA database shows good performance for both face completion and face generation, which indicates the effectiveness of the proposed method. Additional comparison with the state-of-the-art should have included in the Tables to further strengthen their claim.