Simon Kohl: Introduction to Generative Adversarial Networks
10. April 2017
Generative Adversarial Networks (GANs) currently are a much hyped deep neural network concept. Unlike other hype bubbles, there is some true merit to this one: The quality of generated image samples is at a previously unreachable level and applications that were thought to require 'creative human' input, like text-to-image or image-to-image translations, are now being solved by trainable algorithms to an acceptable degree. This presentation gives an introduction to the GAN framework, tips on how to stabilize their training and shows how GANs can be leveraged for challenging segmentation tasks.