Reviews: Robustness of conditional GANs to noisy labels

Conditional generative adversarial network (GAN) learn conditional distribution from a joint probability distribution but corrupted labels hinder this learning process. The paper highlights a unique algorithm to learn this conditional distribution with corrupted labels. The paper introduces two architectures i) RCGAN which relies on availability of matrix C which contains information regarding the errors, and ii) RCGAN-U which does not contain any information about the matrix C. The paper claims that there is no significant loss in performance with regards to knowledge about the matrix C. Even though the problem is unique in nature the paper contains details of some of the related work and references to techniques utilized in the paper such as projection discriminator. I believe the in-depth analysis of the assumptions with theorems and proofs solidify the claims made in the paper although the math requires a more careful check.