InfoGAN—Unsupervised Attribute Extraction

We have learned how to use auxiliary information, such as the labels of data, to improve the image quality generated by GANs. However, it is not always possible to prepare accurate labels of training samples beforehand. Sometimes, it is even difficult for us to accurately describe the labels of extremely complex data. In this section, we will introduce another excellent model from the GAN family, InfoGANChen, Xi, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. "Infogan: Interpretable representation learning by information maximizing generative adversarial nets." Advances in neural information processing systems 29 (2016)., which is capable of extracting data attributes during training in an unsupervised manner. InfoGAN showed that GANs could not only learn to generate realistic samples but also learn semantic features, which are essential to sample generation.

Similar to CGANs, InfoGAN also replaces the original distribution of data with conditional distribution (with auxiliary information as conditions). The main difference is that InfoGAN does not need to feed label and attribute information into the discriminator network. Instead, it uses another classifier, $Q$, to measure how auxiliary features are learned. The objective function of InfoGAN is as follows.

We may notice that it adds another objective, $-\lambda L_I$, at the end of the formula:

In this formula, $x^* = G(z, \gamma)$ is the generated sample. $z$ is the latent vector, and $\gamma$ represent auxiliary information. $P(\gamma)$ describes the actual distribution of $\gamma$, which is rather hard to find. Therefore, we use the posterior distribution, $Q(\gamma|x)$, to estimate $P(\gamma)$ ...