A Pedagogical Introduction to Score Models


Eric J. Ma


April 21, 2022

One of the coolest new developments in generative modelling is the use of score models to generate samples that look like an existing collection of samples. Pioneering this work is Yang Song, who did this work while a PhD student at Stanford University. The way I first uncovered his work is through Twitter, where I was brought to his excellently written blog post on the topic.

In this collection of notebooks, I would like to explore the fundamental ideas that underlie his work. Along the way, we will work towards a pedagogical implementation of score models as generative models. By the end of this journey, we should have a much better understanding of score models and their core concepts, and should also have a framework for writing the code necessary to implement score models in JAX and Python.