Consequently, there exists a nearly trivial discriminator that can perfectly distinguish real data samples from synthetic ones.

Training of Generative Adversarial Networks (GANs) is notoriously fragile, requiring to maintain a careful balance between the generator and the discriminator in order to perform well. To mitigate this issue we introduce a new regularization technique - progressive augmentation of GANs (PA-GAN). The key idea is to gradually increase the task difficulty of the discriminator by progressively augmenting its input or feature space, thus enabling continuous learning of the generator. We show that the proposed progressive augmentation preserves the original GAN objective, does not compromise the discriminator's optimality and encourages a healthy competition between the generator and discriminator, leading to the better-performing generator. We experimentally demonstrate the effectiveness of PA-GAN across different architectures and on multiple benchmarks for the image synthesis task, on average achieving 3 point improvement of the FID score.

Neural Information Processing Systems http://nips.cc/

This paper introduces a novel regularization method (e.g. Instead of weakening or regularizing the discriminator, the idea is to augment the data samples or features with random bits to increase the discrimination task difficulty. In this way, it could prevent the discriminator from being overconfident and maintain a healthy competition, which would enable the generator to be continuously optimized. The augmentation could be progressively levelled up during the training by evaluating the kernel inception distance between synthetic samples and training data samples. The proposed method has been demonstrated on different datasets and compared with other regularization techniques.

The paper proposes progressive augmentation for GANs and shows that it leads to stable training and improves FID consistently. The author response addressed some of the initial concerns, and all the reviewers lean towards accepting the paper. It's nice to see that the proposed technique appears to be complementary to other regularization schemes, so it has the potential to be more widely useful for other machine learning problems (the authors themselves mention this as one of the future directions). I encourage the authors to incorporate reviewer feedback into the final version.

Training of Generative Adversarial Networks (GANs) is notoriously fragile, requiring to maintain a careful balance between the generator and the discriminator in order to perform well. To mitigate this issue we introduce a new regularization technique - progressive augmentation of GANs (PA-GAN). The key idea is to gradually increase the task difficulty of the discriminator by progressively augmenting its input or feature space, thus enabling continuous learning of the generator. We show that the proposed progressive augmentation preserves the original GAN objective, does not compromise the discriminator's optimality and encourages a healthy competition between the generator and discriminator, leading to the better-performing generator. We experimentally demonstrate the effectiveness of PA-GAN across different architectures and on multiple benchmarks for the image synthesis task, on average achieving 3 point improvement of the FID score.

Training of Generative Adversarial Networks (GANs) is notoriously fragile, requiring to maintain a careful balance between the generator and the discriminator in order to perform well. To mitigate this issue we introduce a new regularization technique - progressive augmentation of GANs (PA-GAN). The key idea is to gradually increase the task difficulty of the discriminator by progressively augmenting its input or feature space, thus enabling continuous learning of the generator. We show that the proposed progressive augmentation preserves the original GAN objective, does not compromise the discriminator's optimality and encourages a healthy competition between the generator and discriminator, leading to the better-performing generator. We experimentally demonstrate the effectiveness of PA-GAN across different architectures and on multiple benchmarks for the image synthesis task, on average achieving 3 point improvement of the FID score.