This is an informal whiteboard/blackboard presentation on generative adversarial networks by a non-expert. I will explain integral probability metrics and their use on training deep generative models, widely known as GAN (generative adversarial networks). My plan is to focus on two well-known metrics for probability distributions, Wasserstein distance and Maximum Mean Discrepancy. I will explain how their dual characterisations have been used in the adversarial training of deep generative models, and I will make superficial/speculative comparisons between these metrics. My talk will be based on the following papers and the slides:

References: Wasserstein GAN https://arxiv.org/abs/1701.07875

Generative Models and Model Criticism via Optimized Maximum Mean Discrepancy https://arxiv.org/abs/1611.04488

Maximum Mean Discrepancy http://alex.smola.org/teaching/iconip2006/iconip_3.pdf