BrGANs: Stabilizing GANs' Training Process with Brownian Motion Control

Tianjiao Luo, Ziyu Zhu, Gabriele Oliaro, Jun Zhu, Zhidong Deng

Keywords: GAN, stability, control theory, Brownian motion

TL;DR: We propose a higher order Brownian Motion Controller (BMC) for BrGANs to stablize GANs' training process

Abstract: The training process of generative adversarial networks (GANs) is unstable and does not converge globally. In this paper, we propose a universal higher order noise based control called Brownian Motion Control (BMC) that is invariant to GANs frameworks so that the training process of GANs is exponential stable almost surely. Specifically, starting with the prototypical case of Dirac-GANs, we design a BMC and propose Dirac-BrGANs that retrieve exactly the same but reachable optimal equilibrium regardless of GANs' framework. The optimal equilibrium of our Dirac-BrGANs' training system is globally unique and always exists. Furthermore, the training process of Dirac-BrGANs achieve exponentially stability almost surely for any arbitrary initial value. Then we extend our BMC to normal GANs' settings and propose BrGANs. We provide numerical experiments showing that our BrGANs effectively stabilizes GANs's training process and obtains state-of-the art performance compared to other stabilizing methods.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Generative models