[Paper Review] InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

Xi Chen, Yan Duan+4 more|arXiv (Cornell University)|Jun 11, 2016

Generative Adversarial Networks and Image Synthesis23 references2,416 citations

TL;DR

InfoGAN extends GAN with an information-theoretic regularization to maximize mutual information between a subset of latent codes and generated images, enabling unsupervised learning of interpretable and disentangled representations.

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.

Motivation & Objective

• Motivate unsupervised learning of disentangled representations without labels.
• Develop an information-theoretic extension to GANs to learn meaningful latent factors.
• Demonstrate that the method discovers semantic concepts across MNIST, SVHN, CelebA, and 3D datasets.
• Provide a scalable, trainable objective that induces latent codes to meaningfully control generated outputs.

Proposed method

• Decompose GAN input into incompressible noise z and latent code c to guide generation G(z, c).
• Introduce a mutual information term I(c; G(z, c)) regularized by a hyperparameter λ in the minimax objective.
• Derive a variational lower bound LI(G, Q) using an auxiliary distribution Q(c|x) to approximate P(c|x).
• Train D, G, and Q end-to-end with a simple reparameterization trick to maximize V(D, G) − λLI(G, Q).
• Parameterize Q as a neural network sharing layers with the discriminator D, enabling negligible additional cost.
• Use softmax for discrete codes and diagonal Gaussian for continuous codes within Q.

Experimental results

Research questions

• RQ1 Can information-theoretic regularization induce interpretable latent factors in an unsupervised GAN framework?
• RQ2 Do latent codes c correspond to semantically meaningful variations (e.g., digit shape, pose, lighting, hairstyle) across diverse datasets without supervision?
• RQ3 How does InfoGAN's learned representation compare to supervised methods in terms of disentanglement and usefulness for downstream tasks?

Key findings

• InfoGAN quickly maximizes LI(G, Q) to the entropy H(c) on MNIST, indicating the bound is tight and maximal mutual information is achieved.
• On MNIST, a single discrete code captures digit type while continuous codes model rotation and width, with meaningful, generalizable variations.
• On 3D faces and chairs, InfoGAN learns continuous factors like azimuth, elevation, and lighting, and continuous pose or width variations without supervision.
• On SVHN, InfoGAN learns factors such as lighting and central digit context, despite noisy, cluttered images.
• On CelebA, InfoGAN discovers azimuth, presence/absence of glasses, hairstyle, and emotion without labels, demonstrating high semantic disentanglement.