Generative diffusion models: Key ideas, current practice and challenges

The current landscape of generative models is rich and complex. In this talk, we will first provide some background for the generative modelling task, and compare different generative modelling techniques through the lenses of Optimal Transport. Then we will focus on a special class of generative models, called diffusion processes, and in particular score-based models. Our goal is to provide an overview of the theory behind score-based models, but also to demystify them and suggest a practical interpretation of these methods. We will conclude the talk with challenging topics and methodological advances that are required to improve the practical application of such exciting new methods.