Beyond the previously discussed tasks of classification, segmentation, and detection, there is one major problem type that we have not yet covered at the seminars. This is image generation, and while it is not usually classified as an analysis task, it actually requires the most in-depth modeling of image data. In order to create realistic images, the networks must be able to encode not only the contents of individual samples, but the distribution of information over a sample set. At this tutorial, we are going to showcase some common methods for image generation, including variational autoencoders and the extremely popular generative adversarial networks.

Goals:

• Learn about autoencoders and their variational counterparts.
• Learn about classical and conditional GANs.
• Explore latent space embeddings and how they can be used to generate images with a specific class.