AGE Networks and BEGAN

The adversarial game is formalized by the value function … Lacks an efficient mechanism to infer z given x Goal of extensions: Align real data distribution to generated data distribution and create a correspondence between data and latent samples … The adversarial game is formalized by the value function … Similar to EBGAN, use auto-encoder as discriminator Unlike EBGAN and other GANs, match auto-encoder loss distributions Divergence between distributions using a lower bound of Wasserstein distance Proportional Control Theory to maintain equilibrium between power of discriminator and generator Control image diversity and visual quality trade-off using diversity ratio Similar to Wasserstein GAN, define an approximate measure of convergence Simple training with fast and stable convergence No need to pretrain discriminator No need to alternate training discriminator and generator. Can update parameters of generator and encoder …