The Bid Picture: Auction-Inspired Multi-player Generative Adversarial Networks Training

TL;DR

The paper tackles mode collapse in GANs by extending the traditional two-player game into a multi-player setting with auction-inspired valuation among multiple GANs. It introduces an auction-based valuation process and an auxiliary training stage guided by the best-performing discriminator, operating over N GAN pairs $(G_i,D_i)$ with a cross-GAN score $S(i)$. The training alternates between independent updates and auxiliary updates, selecting and aligning participants through $L_{aux}$ to discourage failure modes. Experiments on a synthetic eight-mode 2D Gaussian mixture show improved mode coverage and stability for both vanilla GANs and WGANs, though at the cost of extra computation. This approach offers a scalable mechanism to promote diversity without major architectural changes and points to avenues for refining bid mechanisms and extending to other GAN variants.

Abstract

This article proposes auction-inspired multi-player generative adversarial networks training, which mitigates the mode collapse problem of GANs. Mode collapse occurs when an over-fitted generator generates a limited range of samples, often concentrating on a small subset of the data distribution. Despite the restricted diversity of generated samples, the discriminator can still be deceived into distinguishing these samples as real samples from the actual distribution. In the absence of external standards, a model cannot recognize its failure during the training phase. We extend the two-player game of generative adversarial networks to the multi-player game. During the training, the values of each model are determined by the bids submitted by other players in an auction-like process.

Figures (2)

• Figure 1: Example auction framework
• Figure 2: Example single auction phase