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.

It is a good paper that should definitely be accepted. The presented approach has a clear theoretical motivation and is supported by a thorough and convincing experimental evaluation. It is important that the approach does not use any domain-specific knowledge and effectively comes at zero additional computational cost. This makes it easily applicable to a wide range of generative tasks. I have several questions/comments: 1) It seems to me that the proposed approach in the end amounts to training a GAN with an additional network (or an additional branch on top of the discriminator) trained to predict part of the latent code from the generated image.

Research by Hubel and Wiesel [2,3] analyzed the striate cortex of cats and monkeys, revealing two key findings that would come to heavily influence Fukushima's work [1]. The next significant implementation of a convolution neural network was LeNet-5 proposed in 1999 by Le Cun et al. in their work "Object Recognition with Gradient Based Learning'' [4]. Their proposed network, LeNet-5 performed well on the MNIST data set and was shown to do better than state of the art (at the time) SVMs and K-nearest neighbor based approaches. Their final implementation outperformed other state of the art image classification algorithms with error rates which were 10% lower than its competitors on the ImageNet dataset. This application of a discrete convolution precisely represents local receptive fields observed by Hubel and Wiesel [2,3] and implemented in early CNNs by Fukushima and Le Cun [1,4]. Research by Hubel and Wiesel [2,3] analyzed the striate cortex of cats and monkeys, revealing two key findings that would come to heavily influence Fukushima's work [1]. The next significant implementation of a convolution neural network was LeNet-5 proposed in 1999 by Le Cun et al. in their work "Object Recognition with Gradient Based Learning'' [4].

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.

Taken from the InfoGan paper. Let's start off by developing the generator model as a deep convolutional neural network (e.g. a DCGAN). The model could take the noise vector (z) and control vector (c) as separate inputs and concatenate them before using them as the basis for generating the image. Alternately, the vectors can be concatenated beforehand and provided to a single input layer in the model. The approaches are equivalent and we will use the latter in this case to keep the model simple.