Online Adaptative Curriculum Learning for GANs
share
Generative Adversarial Networks (GANs) can successfully learn a probability distribution and produce realistic samples. However, open questions such as sufficient convergence conditions and mode collapse still persist. In this paper, we build on existing work in the area by proposing a novel framework for training the generator against an ensemble of discriminator networks, which can be seen as a one-student/multiple-teachers setting. We formalize this problem within the non-stationary Multi-Armed Bandit (MAB) framework, where we evaluate the capability of a bandit algorithm to select discriminators for providing the generator with feedback during learning. To this end, we propose a reward function which reflects the amount of knowledge learned by the generator and dynamically selects the optimal discriminator network. Finally, we connect our algorithm to stochastic optimization methods and show that existing methods using multiple discriminators in literature can be recovered from our parametric model. Experimental results based on the Fréchet Inception Distance (FID) demonstrates faster convergence than existing baselines and show that our method learns a curriculum.
READ FULL TEXT