AI4S Seminar by Wei Huang "Diffusion Models for Scientific Data"

Speaker: Wei Huang (Riken AIP)

Title: Diffusion Models for Scientific Data: From Continuous Manifolds to Discrete Token Ordering
Abstract:

Diffusion models are emerging as powerful generative tools for scientific data, ranging from molecular structures and biological measurements to protein and DNA sequences. This talk explores how diffusion models exploit structure in both continuous and discrete state spaces. For continuous data, I will introduce Score-induced Latent Diffusion (SiLD), which reveals a “collapse-and-refine” mechanism under the manifold hypothesis: score learning first discovers the low-dimensional data manifold and then refines the distribution within it. I will discuss its theoretical motivation and application to molecular generation. For discrete data, I will introduce DPRM, a plug-in token-ordering module for masked diffusion models inspired by the Doob (h)-transform. By combining model confidence with estimates of future task reward, DPRM adapts the generation order to global objectives without modifying the underlying diffusion model. I will discuss its applications and limitations in language and scientific sequence generation.
I plan to introduce the basic ideas of continuous and discrete diffusion models before presenting the two recent works, so that the talk will be accessible to participants from different backgrounds.