Chapter 6: Generative Models and Adversarial Attacks

Being able to differentiate between two or more classes is certainly impressive, and a healthy sign that deep neural networks do, in fact, learn.

But if traditional classification is impressive, then producing new content is staggering! That definitely requires a superior understanding of the domain. So, are there neural networks capable of such a feat? You bet there are!

In this chapter, we'll study one of the most captivating and promising types of neural networks: Generative Adversarial Networks (GANs). As the term implies, these networks are actually a system comprised of two sub-networks: the generator and the discriminator. The job of the generator is to produce images so good that they could come from the original distribution (but actually don't; they're generated from scratch), thereby fooling the discriminator, whose task is to discern between real and fake images.

GANs are the tip of the spear...