Although I’ve worked with machine learning for over six years—mostly in variational inference, energy-based models, and normalizing flows — I had never implemented a generative adversarial network (or GAN) from scratch until recently. My antagonism towards GANs stemmed from (possibly outdated) misunderstandings — many of which were perpetuated due to a variety of different interpretations from many different papers, blogposts and articles.

Neural networks are all the rage in nowadays. These brain-inspired architecture that power many of todays applications are capable of recognizing images and playing video games at superhuman levels, as well as generate art and perform robotics tasks.