Can deep learning beat the atmosphere at its own game?

Abstract: Atmospheric seeing and image noise are the arch-nemeses of the ground-based astronomer. Particularly within the cosmology community, much ink is being spilled over how to overcome these challenges in the absence of a Large Synoptic Space Telescope. Schawinski et al. (https://arxiv.org/abs/1702.00403) have attempted to recover near-space-quality images from noisy ground-based data using a deep learning technique which Yann LeCun--a godfather of the field--called "the most interesting idea in the last 10 years of machine learning": generative adversarial networks (GANs). We will discuss what GANs are and how they work, as well as their (spoiler alert: significant) limitations in astrophysical contexts. The original paper on GANs (https://papers.nips.cc/paper/5423-generative-adversarial-nets.pdf) may be useful, although I hope to inspire you to read it rather than expecting anyone to read it ahead of time...