[論文レビュー] VEEGAN: Reducing Mode Collapse in GANs using Implicit Variational Learning

Akash Srivastava, Lazar Valkov+3 more|arXiv (Cornell University)|May 22, 2017

Generative Adversarial Networks and Image Synthesis参考文献 20被引用数 273

ひとこと要約

VEEGAN はデータを再びガウスノイズへマッピングする再構成器ネットワークを導入し、生成器と共に暗黙の変分目的で訓練することでモード崩壊を緩和し、より高品質なサンプルを生成する。

ABSTRACT

Deep generative models provide powerful tools for distributions over complicated manifolds, such as those of natural images. But many of these methods, including generative adversarial networks (GANs), can be difficult to train, in part because they are prone to mode collapse, which means that they characterize only a few modes of the true distribution. To address this, we introduce VEEGAN, which features a reconstructor network, reversing the action of the generator by mapping from data to noise. Our training objective retains the original asymptotic consistency guarantee of GANs, and can be interpreted as a novel autoencoder loss over the noise. In sharp contrast to a traditional autoencoder over data points, VEEGAN does not require specifying a loss function over the data, but rather only over the representations, which are standard normal by assumption. On an extensive set of synthetic and real world image datasets, VEEGAN indeed resists mode collapsing to a far greater extent than other recent GAN variants, and produces more realistic samples.

研究の動機と目的

• Motivate and address mode collapse in GANs where generators miss data distribution modes.
• Propose a reconstructor network that maps real data to Gaussian noise and approximately inverts the generator.
• Develop an implicit variational objective that combines a reconstruction loss on latent representations with a KL-like term.
• Show that optimizing this objective encourages the generator to cover the full data distribution without requiring explicit data-space reconstruction losses.

提案手法

• Introduce a reconstructor network F_theta that maps data X to latent noise Z and approximately inverts the generator G_gamma.
• Formulate an implicit variational objective that combines an autoencoder-like loss on latent representations with a cross-entropy term ensuring F_theta(X) matches the prior Z~p0(z).
• Derive a computable bound using a variational distribution q_gamma(x|z) to handle implicit distributions.
• Use a learned discriminator D_omega to estimate a density-ratio term needed for the KL-like objective in the presence of implicit models.
• Optimize the joint objective with respect to gamma (generator) and theta (reconstructor) using stochastic gradient descent, together with discriminator updates (as in GANs).
• Explain differences to BiGAN/ALI, InfoGAN, and adversarial autoencoders, highlighting the noise-space autoencoding and the data-to-noise mapping distinction.

実験結果

リサーチクエスチョン

• RQ1 Does adding a reconstructor that maps data to Gaussian noise help detect and mitigate mode collapse in GANs?
• RQ2 Can an implicit variational objective, coupled with a noise-space autoencoder, provide strong learning signals even when the discriminator is non-informative?
• RQ3 How does VEEGAN compare to existing GAN variants (e.g., ALI, Unrolled GAN, InfoGAN) in terms of mode coverage and sample quality across synthetic and real image datasets?
• RQ4 What are the practical training considerations and benefits of using a noise-based autoencoder over a data-space autoencoder in GAN training?

主な発見

• VEEGAN reduces mode collapse more effectively than several state-of-the-art GAN variants on synthetic and real image datasets.
• The approach yields more diverse and realistic samples, with better coverage of data modes.
• Using a noise-space autoencoder (autoencoding latent z) provides stable training signals without requiring a data-space reconstruction loss.
• The method remains effective with default hyperparameters and does not rely on extensive tuning of regularization weights.
• VEEGAN demonstrates improved mode capture and sample fidelity on stacked MNIST and CIFAR-10 datasets compared to baselines like GAN, ALI, and Unrolled GAN.