Semi-supervised Conditional GANs

Generative adversarial networks (GAN's) [2] are a recent popular technique for learning generative models for high-dimensional unstructured data (typically images). GAN's employ two networks - a generator G that is tasked with producing samples from the data distribution, and a discriminator D that aims to distinguish real samples from the samples produced by G. The two networks alternatively try to best each other, ultimately resulting in the generator G converging to the true data distribution. While most of the research on GAN's is focused on the unsupervised setting, where the data is comprised of unlabeled images, there has been research on conditional GAN's [1] where the goal is to learn a conditional model of the data, i.e. to build a conditional model that can generate images given a particular attribute setting. In one approach [1], both the generator and discriminator are fed attributes as side information so as to enable the generator to generate images conditioned on attributes. In an alternative approach proposed in [5], the authors build auxiliary classifier GAN's (AC-GAN's) where side information is reconstructed by the discriminator instead. Irrespective of the specific approach, this line of research focuses on the supervised setting where it is assumed that all the images have attribute tags. Given that labels are expensive, it is of interest to explore semi-supervised settings where only a small fraction of the images have attribute tags, while a majority of the images are unlabeled. There has been some work on using GAN's in the semi-supervised setting.