Generative Adversarial Networks (or GANs) are a landmark architecture in the field of machine learning, being the first successful architecture for generating realistic images. We’ll begin with an introduction to Convolutional Neural Networks, followed by a description of the two sub-network setup – the generator network and the discriminator network. The generator network’s objective is to generate images that are realistic, and the discriminator network’s objective is to distinguish a generated image from a real image. The two networks are playing a two-player minimax game.

This will be followed by a description of the various possible degenerate solutions that the GAN model may converge to and methods to overcome them. For example, one degenerate solution occurs when the generator networks keeps hopping between generating a small number of different images. Each time the discriminator network figures out how to distinguish the images, the generator network moves onto the next image.

In the last part of the talk, we’ll take a step back and look at the increasing modularity that we are moving towards in Deep Learning. This is the phenomenon of using quite large sub-networks in deep learning from one application to another. In the case of GANs, the discriminator network is a standard Convolutional Neural Network. We’ll discuss other such examples from recent research literature, and also discuss some possible proposals. We’ll also discuss how this kind of a system is much more suited for generative tasks as compared to a static function which scores the goodness of a generated sample.