Competitive-Collaborative GAN with Performance Guarantee

CHENHAN XIAO, Yang Weng

Primary Area: generative models

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.

Keywords: Generative adversarial network, Collaboration, Equilibrium point

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Generative Adversarial Networks (GANs) generate data based on a competition game to minimize the distribution distance between existing and new data. However, such a competition game falls short when insights about data distributions beyond their authenticity are imperative, such as in multi-modal generation and image super resolution. In recognition of the limitations inherent to the pure-competitive mechanism, we introduce CCGAN, a Collaborative-Competitive Generative Adversarial Network scheme to enable data generation with additional knowledge beyond the provided dataset distribution. For theoretically preserving the equilibrium point and numerically avoiding training collapse issue, we show the need to convert regularization term into a divergence, so that the modified GAN is well-defined in game theory. By harmonizing the competition and collaboration losses in CCGAN, we effectively reduce the degree complexity of solving the optima, facilitating the establishment of a closed-form equilibrium point. This equilibrium point serves as a guidance for training and hyper-parameter tuning, resulting in consistently high-quality generated samples. Meanwhile, the regularization breaks the mutual dependency between the generator and discriminator. This newfound independence empowers the CCGAN to explore a broader parameter space, effectively mitigating the training collapse issue. To validate the capabilities of CCGAN, we design comprehensive experiments across four publicly available datasets and systematically compare CCGAN against a range of baseline models. The experiments demonstrate the efficacy of CCGAN on generating satisfactory samples tailored to specific requirements, particularly when applied to the generation of images featuring regularly shaped objects.

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.

Submission Number: 8075