BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning and Regularizing Score-Based Diffusion Models - Ricardo B
 aptista (CALTECH (California Institute of Technology))
DTSTART:20240716T090000Z
DTEND:20240716T100000Z
UID:TALK219028@talks.cam.ac.uk
DESCRIPTION:Diffusion models have emerged as a powerful framework for gene
 rative modeling that relies on score matching to learn gradients of the da
 ta distribution's log-density. A key element for the success of diffusion 
 models is that the optimal score function is not identified when solving t
 he denoising score matching problem. In fact\, the optimal score in both u
 nconditioned and conditioned settings leads to a diffusion model that retu
 rns to the training samples and effectively memorizes the data distributio
 n. In this presentation\, we study the dynamical system associated with th
 e optimal score and describe its long-term behavior relative to the traini
 ng samples. Lastly\, we show the effect of three forms of score function r
 egularization on avoiding memorization: restricting the score's approximat
 ion space\, early stopping of the training process\, and early stopping of
  the diffusion process during sample generation. Moreover\, we establish a
  connection between early stopping of the diffusion and explicit Tikhonov 
 regularization of the score matching problem. These results are numericall
 y validated using distributions with and without densities including image
 -based problems.
LOCATION:External
END:VEVENT
END:VCALENDAR