BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.2.3.1//EN
TZID:Europe/Paris
X-WR-TIMEZONE:Europe/Paris
BEGIN:VEVENT
UID:65@osups.universite-paris-saclay.fr
DTSTART;TZID=Europe/Paris:20260319T110000
DTEND;TZID=Europe/Paris:20260319T120000
DTSTAMP:20260303T114633Z
URL:https://osups.universite-paris-saclay.fr/calendrier/diffusion-models-c
 urse-and-blessing-for-generalization/
SUMMARY:Diffusion Models: Curse and Blessing for Generalization
DESCRIPTION:Par Tony Bonnaire (IAS)\n\nAbstract : \n\nDiffusion models hav
 e emerged as powerful generative tools\, capable of producing highly reali
 stic images\, videos\, and sounds by learning a stochastic mapping startin
 g from a simple Gaussian distribution. Despite their empirical success\, t
 he underlying reasons for their effectiveness remain poorly understood. In
  this talk\, we will give a brief primer on the formalism of diffusion mod
 els and then dive into the analysis of a well-defined high-dimensional dat
 a distribution-a mixture of two Gaussians-under the assumption of an optim
 ally-trained model. In particular\, this reveals the existence of a 'memor
 ization' transition where the trajectories are inevitably attracted to one
  of the training points and reproduce it exactly. Interestingly\, we will 
 provide an explanation on why diffusion models can still generalize despit
 e this curse thanks to an implicit regularization mechanism in their train
 ing dynamics. This creates a generalization window that grows linearly wit
 h the training set size\, enabling diffusion models to generalize. These f
 indings are supported by both analytical results on simplified models and 
 large-scale numerical experiments on realistic models and datasets. If tim
 e permits\, we will discuss ongoing applications of such generative models
  in physics\, especially in disordered systems and cosmology.\n\nLieu : IA
 S\, bât. 121\, salle 4-5 (4ème étage)
CATEGORIES:Séminaires
LOCATION:IAS\, Bât. 121 rue Jean Teillac\, Orsay\, 91400\, France
X-APPLE-STRUCTURED-LOCATION;VALUE=URI;X-ADDRESS=Bât. 121 rue Jean Teillac\
 , Orsay\, 91400\, France;X-APPLE-RADIUS=100;X-TITLE=IAS:geo:0,0
END:VEVENT
BEGIN:VTIMEZONE
TZID:Europe/Paris
X-LIC-LOCATION:Europe/Paris
BEGIN:STANDARD
DTSTART:20251026T020000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
END:STANDARD
END:VTIMEZONE
END:VCALENDAR