Stackelberg GAN: Towards Provable Minimax Equilibrium via Multi-Generator Architectures

Abstract

We study the problem of alleviating the instability issue in the GAN training procedure via new architecture design. The discrepancy between the minimax and maximin objective values could serve as a proxy for the difficulties that the alternating gradient descent encounters in the optimization of GANs. In this work, we give new results on the benefits of multi-generator architecture of GANs. We show that the minimax gap shrinks to $\epsilon$ as the number of generators increases with rate $\tilde{O}(1/\epsilon )$. This improves over the best-known result of $\tilde{O}(1/{\epsilon }^2)$. At the core of our techniques is a novel application of Shapley-Folkman lemma to the generic minimax problem, where in the literature the technique was only known to work when the objective function is restricted to the Lagrangian function of a constraint optimization problem. Our proposed Stackelberg GAN performs well experimentally in both synthetic and real-world datasets, improving Fréchet Inception Distance by $14.61\%$ over the previous multi-generator GANs on the benchmark datasets.

arXiv comment27 pages, 13 figures, 6 tables

References in corpus (11)

• Auto-Encoding Variational Bayes
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• NIPS 2016 Tutorial: Generative Adversarial Networks
• Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples
• BEGAN: Boundary Equilibrium Generative Adversarial Networks
• Adversarially Learned Inference
• A Downsampled Variant of ImageNet as an Alternative to the CIFAR datasets
• Generating Adversarial Examples with Adversarial Networks
• Dual Discriminator Generative Adversarial Nets
• MAGAN: Margin Adaptation for Generative Adversarial Networks
• Deep Neural Networks with Multi-Branch Architectures Are Less Non-Convex

Cited by in corpus (4)

• Theoretically Principled Trade-off between Robustness and Accuracy
• Distributed Traffic Synthesis and Classification in Edge Networks: A Federated Self-supervised Learning Approach
• Design and Interpretation of Universal Adversarial Patches in Face Detection
• Game of GANs: Game-Theoretical Models for Generative Adversarial Networks