Please note: This PhD seminar will take place in DC 1304.
Haoye Lu, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Yaoliang Yu
Diffusion models need large amounts of clean data, but in settings like medical imaging, privacy-sensitive data, and scientific measurement, clean examples are rare while noisy or incomplete ones are plentiful. Can we train a generative model of the clean data when we mostly only see corrupted versions of it?
This talk traces the SFBD series, which answers yes, with convergence guarantees and strong results from very few clean samples. I’ll start with the original SFBD to build intuition: alternating a noise-adding diffusion process with a learned denoiser gradually recovers the true data distribution. This first version is tailored to Gaussian noise; a natural extension, SFBD-Omni, replaces the denoiser with a conditional diffusion model to handle nearly arbitrary corruption processes and clarifies when recovery is possible at all. Both methods, however, train a sequence of networks one after another, which leaves practical questions open, such as how long to train at each round before moving on. SFBD Flow resolves this by smoothing the alternating procedure into a single end-to-end algorithm that trains faster and performs better. I’ll close with applications in medical domains.