E-processes automate GAN training switching, beating fixed update ratios

Training generative adversarial networks (GANs) typically requires manually setting how often to alternate between discriminator and generator updates—a fixed ratio that rarely adapts to the training dynamics. A new arXiv paper (2608.10096) from Hyunjoo Kim and colleagues reframes this as a sequential hypothesis testing problem. They introduce two e-processes: one tests whether the discriminator-induced separation between empirical data and generator distribution stays below a target level during discriminator updates; the other tests the reverse null during generator updates. Using conditional e-values, these processes accumulate evidence to decide when to switch training phases, with provable anytime-valid Type I error control—even as models update adaptively and switching is data-dependent.

The method is evaluated on multimodal synthetic distributions and image benchmark datasets under several widely used GAN objectives. Results show it matches or outperforms the best fixed-ratio baselines, suggesting that adaptive switching can replace manual tuning without sacrificing performance. The theoretical guarantee means practitioners can rely on statistically sound stopping rules, eliminating trial-and-error grid searches over update schedules. While the paper is focused on GANs, the e-process framework could extend to other stochastic min-max optimization problems, including adversarial robustness training and reinforcement learning with opponent modeling. This work offers a principled alternative to heuristic switching criteria and opens the door to more autonomous, statistically grounded training pipelines.