Lab Project: Intro to Statistical Distances for Generative Modeling
- 1 min
Earlier this year, I had a chance to take part in a unique project in our group at mackelab, where the entire lab took part. While we work in vastly diverse domains, we all typically encountered the much-maligned “table of metrics” in machine learning papers, measuring the goodness-of-fit of generative models to some tasks. We wanted to collectively dig deeper into these metrics - how are they calculated? What are their biases? Which ones should I use to train my generative model? So the lab cleared our schedule for a collective few days of hacking, covering theoretical and computational aspects of commonly used distances! We learned a lot from this exercise, and produced what is a hopefully useful guide to practitioners like us!