Seminario Machine Learning
A Contractive Geometry of Diffusion Sampling
Ponente: Ann Dooms (Vrije Universiteit Brussel)Fecha: miércoles 13 de mayo de 2026 - 16:00Lugar: Aula Naranja, ICMAT
Resumen:
Diffusion samplers are often judged by their final samples, but their failures are created step by step. This talk presents a local geometric view of deterministic DDIM sampling through the Jacobian of each reverse step. The main message is that every step contains three interacting forces turning sampler behaviour into a diagnostic problem: which timesteps, directions, and image blocks amplify perturbations, and which errors survive to the final image? The resulting theory gives contraction criteria, chain-level error weights, and blockwise tools for identifying fragile parts of a frozen sampler.
