2019
informatika
Domain Partitioning Networks
Témavezető:
Dr. Horváth András, Philip H.s. Torr
Dr. Horváth András, Philip H.s. Torr
Összefoglaló
Standard adversarial training involves two agents, namely a generator and a discriminator, playing a mini-max game. However, even if the players converge to an equilibrium, the generator may only recover a part of the target data distribution, in a situation commonly referred to as mode collapse. In this work, we present Triple-GMAN, a new approach to deal with mode collapse in generative adversarial learning. First, we use multiple discriminators each enforcing the generator to cover an additional part of the target distribution. To ensure these parts don’t overlap and collapse into the same mode, we add a classifier as a third agent in the game that decides for each sample which discriminator the generator is trained against. Through experiments on toy examples and real images, we show the merits of our method and its superiority with respect to competing methods.
Even though the architecture contains a non-differentiable sampling operation, the agents can be trained with the classic stochastic gradient update techniques.
Even though the architecture contains a non-differentiable sampling operation, the agents can be trained with the classic stochastic gradient update techniques.
Dr. Horváth András
horvath.andras@itk.ppke.hu
Philip H.s. Torr