A pioneer of generative models who proposed diffusion probabilistic models in 2015, inspired by non-equilibrium thermodynamics physics. His idea of slowly turning noise into data became the foundation of modern neural networks that generate images, sound, and video, including DALL‑E 2 and Stable Diffusion.
Biography
Studied physics and computer science at the University of Michigan, earned a PhD at Stanford, specializing in machine learning and neuroscience. Worked at Google Brain, then moved into industry, continuing to develop generative AI.
Key discoveries
Diffusion probabilistic models (DPM) — a class of generative algorithms
based on a reversible chain of adding and removing noise. Application of stochastic differential equations to data synthesis. A bridge between statistical physics and modern machine learning.
💡 Sohl-Dickstein was initially inspired by thermodynamics and diffusion processes to create a model that step by step turns noise into meaningful data, which now underlies many generative systems.
"He saw diffusion models as a link between statistical physics and modern AI."