Measure-conditional Discriminator with Stationary Optimum for GANs and Statistical Distance Surrogates
This addresses training instability in GANs for generative modeling, but it appears incremental as a plug-and-play modification.
The authors tackled the instability in GAN training by proposing measure-conditional discriminators, which incorporate generated distributions as input to achieve stationary optima, resulting in improved robustness and applications like transfer learning.
We propose a simple but effective modification of the discriminators, namely measure-conditional discriminators, as a plug-and-play module for different GANs. By taking the generated distributions as part of input so that the target optimum for the discriminator is stationary, the proposed discriminator is more robust than the vanilla one. A variant of the measure-conditional discriminator can also handle multiple target distributions, or act as a surrogate model of statistical distances such as KL divergence with applications to transfer learning.