Hello everyone! Mark Goldstein is visiting us to give a talk about his research on diffusion models this Friday 11 AM — 12 PM at Friend Center 008. Details below.
Abstract:
This talk will consist of two parts. First, we will discuss continuous diffusion models and the role of noising processes and base distributions, which govern the sequence of densities that a diffusion model must traverse to generate data. These details in turn govern problem difficulty. We will discuss choices that can improve models and reduce the amount of optimization required, in the context of natural images, video generation, Navier-Stokes equations, and protein design. We will then see model likelihood bounds that unify several variants of diffusion models, flow matching, etc...
Second, we will discuss discrete diffusion models, which have been around in varying forms for a few years but have recently been formalized as proper non-autoregressive generative models for discrete data like text, images with integer pixels, atomic spins, and amino acid sequences. For discrete diffusion models, we will discuss limitations in current practice (sub-optimal factorization choices) and possible routes and challenges involved in improving the models.
Bio:
Mark Goldstein is a PhD candidate in Computer Science at New York University working with Rajesh Ranganath, and a current Student Researcher at Google Deepmind. Mark's work focuses on deep generative models for high-dimensional generation problems, with applications in health and science. Core to this work is rethinking fundamental choices in diffusion model training, and investigating how they affect performance and efficiency. Mark has studied these models in the context of video generation, Navier-Stokes equations, and medical imaging data. Previously, Mark completed a Bachelor's of Music Composition at New England Conservatory of Music.
Best,