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

Wei Huang from RIKEN AIP will present recent work on diffusion models for scientific data at this AI4S seminar. The talk covers how diffusion models exploit structure in both continuous and discrete state spaces, starting with foundational ideas before diving into two recent research contributions. For continuous data, the talk introduces Score-induced Latent Diffusion (SiLD), which reveals a "collapse-and-refine" mechanism under the manifold hypothesis, where score learning first discovers the low-dimensional data manifold before refining the distribution within it, with applications to molecular generation. For discrete data, the talk introduces DPRM, a plug-in token-ordering module for masked diffusion models inspired by the Doob h-transform, which combines model confidence with estimates of future task reward to adapt generation order without modifying the underlying diffusion model. The seminar is designed to be accessible to participants from different backgrounds, with basic concepts introduced before the two recent works are presented.