InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

InfoGAN is an information-theoretic extension to the generative adversarial network that is able to learn disentangled representations in a completely unsupervised manner. In addition to the usual adversarial game, it maximizes the mutual information between a small subset of the latent variables and the observation. The authors derive a lower bound to this mutual information objective that can be optimized efficiently, and show that the resulting training procedure can be interpreted as a variation of the Wake-Sleep algorithm.

Experimentally, InfoGAN successfully disentangles interpretable factors of variation: writing styles from digit shapes on MNIST, pose from lighting in 3D rendered images, and background digits from the central digit on SVHN. It also discovers visual concepts such as hair styles, the presence or absence of eyeglasses, and emotions on the CelebA face dataset, learning interpretable representations that are competitive with those learned by existing fully supervised methods.