Reviews: Dual Discriminator Generative Adversarial Nets

This paper presents a variant of generative adversarial networks (GANs) that utilizes two discriminators, one tries to assign high scores for data, and the other tries to assign high scores for the samples, both discriminating data from samples, and the generator tries to fool both discriminators. It has been shown in section 3 that the proposed approach effectively optimizes the sum of KL and reverse KL between generator distribution and data distribution in the idealized non-parametric setup, therefore encouraging more mode coverage than other GAN variants. The paper is quite well written and the formulation and analysis seems sound and straightforward. The proposed approach is evaluated on toy 2D points dataset as well as more realistic MNIST, CIFAR-10, STL and ImageNet datasets. I have one concern about the new formulation, as shown in Proposition 1, the optimal discriminators have the form of density ratios.