Generative Adversarial Networks: Introduction and Survey of Some

Guys, we are back with our Coffee & Data Science meetups! Our next meeting will be about Generative Adversarial Networks. Generative models are models that can create audio signals, images, and other complex data from scratch in a highly controllable way, enabling a wide range of applications. For instance, imagine a tool that can turn crude sketches into photorealistic imagery or automatically do complex transformations of images such as turning day scenes into night scenes and vice versa. Generative Adversarial Networks (GANs) are currently one of the most promising families of generative models and have shown particular success generating images. In a GAN, two models are simultaneously trained with opposed goals. A "discriminator" is alternately fed real data and fake data cooked up by a "generator", and the discriminator is trained to tell real data from fake data while the generator is trained to fool the discriminator. If all goes will, the generator will learn to produce very convincing fake data in its effort to fool the discriminator, which itself becomes increasingly sophisticated as training goes on. In practice, GAN training often goes awry. But in the past few months there have been several exciting developments in our understanding of GANs that have dramatically improved these models and made the training of them more predictable. The lecture will explain in detail how GANs work and cover some of these recent developments, particularly a GAN variant known as the Wasserstein GAN. About Speaker: Grant Reaber is an independent researcher whose primary interest is generative models of audio. He studied mathematical logic, statistics, and machine learning at Carnegie Mellon University and holds a PhD in philosophy from the University of Aberdeen.