Generative Deep Learning

Chapter 4. Generative Adversarial Networks

On Monday, December 5, 2016, at 2:30 p.m., Ian Goodfellow of Google Brain presented a tutorial entitled “Generative Adversarial Networks” to the delegates of the Neural Information Processing Systems (NIPS) conference in Barcelona.¹ The ideas presented in the tutorial are now regarded as one of the key turning points for generative modeling and have spawned a wide variety of variations on his core idea that have pushed the field to even greater heights.

This chapter will first lay out the theoretical underpinning of generative adversarial networks (GANs). You will then learn how to use the Python library Keras to start building your own GANs.

First though, we shall take a trip into the wilderness to meet Gene…

Ganimals

One afternoon, while walking through the local jungle, Gene sees a woman thumbing through a set of black and white photographs, looking worried. He goes over to ask if he can help.

The woman introduces herself as Di, a keen explorer, and explains that she is hunting for the elusive ganimal, a mythical creature that is said to roam around the jungle. Since the creature is nocturnal, she only has a collection of nighttime photos of the beast that she once found lying on the floor of the jungle, dropped by another ganimal enthusiast. Some of these photos are shown in Figure 4-1. Di makes money by selling the images to collectors but is starting to worry, as she hasn’t actually ever seen the creatures and is concerned that ...