Valentin De Bortoli

Papers

Total Citations

H-Index

About

Valentin De Bortoli is a leading researcher at the forefront of generative modelling, with a particular focus on extending diffusion models to complex, non-Euclidean spaces. His foundational work in "Riemannian Score-Based Generative Modelling" (2022, 26 citations) established a powerful framework for generating data on curved manifolds, bridging a critical gap between theoretical machine learning and applications in the natural sciences. He further advanced this frontier with "Diffusion Models for Constrained Domains" (2023, 5 citations), which systematically addresses the challenge of generating samples subject to complex geometric or physical constraints. To ensure the practical utility of these models, De Bortoli also developed "Metropolis Sampling for Constrained Diffusion Models" (2023, 3 citations), introducing a novel correction mechanism that guarantees samples respect the target distribution on constrained spaces. His work is pivotal for enabling generative AI in fields like molecular dynamics, structural biology, and robotics, where data naturally resides on manifolds. By tackling the core mathematical and algorithmic hurdles of constrained generation, De Bortoli is shaping the next generation of generative models for scientific discovery.