Annealed Generative Adversarial Networks
This addresses a fundamental problem in generative modeling for researchers and practitioners by providing a more stable training method, though it appears incremental as it builds on existing GAN frameworks.
The paper tackles the instability and mode collapse issues in GAN training by introducing an annealing framework that gradually transitions from a uniform to the data distribution, resulting in stable training without mode collapse on synthetic data, MNIST, and CelebA.
We introduce a novel framework for adversarial training where the target distribution is annealed between the uniform distribution and the data distribution. We posited a conjecture that learning under continuous annealing in the nonparametric regime is stable irrespective of the divergence measures in the objective function and proposed an algorithm, dubbed ß-GAN, in corollary. In this framework, the fact that the initial support of the generative network is the whole ambient space combined with annealing are key to balancing the minimax game. In our experiments on synthetic data, MNIST, and CelebA, ß-GAN with a fixed annealing schedule was stable and did not suffer from mode collapse.