[論文レビュー] InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

Xi Chen, Yan Duan+4 more|Ghent University Academic Bibliography (Ghent University)|Jun 12, 2016

Generative Adversarial Networks and Image Synthesis参考文献 4被引用数 1,246

ひとこと要約

InfoGAN は GAN に情報理論的正規化を追加し、潜在コードのごく一部と生成画像との相互情報を最大化することで、MNIST、SVHN、CelebA、3D データセット全体で、監視なしで分解可能で解釈可能な表現を学習できる。

ABSTRACT

This paper describes InfoGAN, an information-theoretic extension to the Generative Adversarial Network that is able to learn disentangled representations in a completely unsupervised manner. InfoGAN is a generative adversarial network that also maximizes the mutual information between a small subset of the latent variables and the observation. We derive a lower bound to the mutual information objective that can be optimized efficiently, and show that our training procedure can be interpreted as a variation of the Wake-Sleep algorithm. Specifically, InfoGAN successfully disentangles writing styles from digit shapes on the MNIST dataset, pose from lighting of 3D rendered images, and background digits from the central digit on the SVHN dataset. It also discovers visual concepts that include hair styles, presence/absence of eyeglasses, and emotions on the CelebA face dataset. Experiments show that InfoGAN learns interpretable representations that are competitive with representations learned by existing fully supervised methods.

研究の動機と目的

• Motivate unsupervised learning of meaningful, disentangled representations for complex visual data.
• Improve GANs by encouraging the generator to use latent codes to encode semantic factors of variation.
• Demonstrate that mutual information regularization yields interpretable factors without labeled supervision.

提案手法

• Decompose GAN input into incompressible noise z and structured latent code c.
• Maximize a variational lower bound of the mutual information I(c; G(z,c)) via an auxiliary distribution Q(c|x).
• Formulate a minimax objective VInfoGAN(D, G, Q) = V(D, G) − λ LI(G, Q).
• Parameterize Q as a neural network shared with the discriminator, enabling end-to-end training.
• Use softmax for discrete latent codes and diagonal Gaussian for continuous codes in Q.
• Train with DC-GAN stabilization techniques and Adam optimization.

実験結果

リサーチクエスチョン

• RQ1 Can mutual information regularization induce interpretable and disentangled latent factors in an unsupervised GAN framework?
• RQ2 What kinds of semantic factors (e.g., digit type, pose, lighting, hair, emotion) can InfoGAN discover across diverse datasets without labels?
• RQ3 How does InfoGAN performance compare to supervised or semi-supervised approaches in learning useful representations?

主な発見

• InfoGAN successfully learns disentangled representations on MNIST, SVHN, CelebA, and 3D face/chair datasets without supervision.
• Discrete latent codes capture category-level variation (e.g., digit type on MNIST) and serve as interpretable classifiers.
• Continuous latent codes capture smooth variations (e.g., rotation, width, azimuth, lighting) affecting generated images realistically.
• InfoGAN discovers semantic concepts such as hair styles, presence of eyeglasses, and emotions in CelebA.
• The learned representations are competitive with representations learned by supervised methods for downstream tasks.