BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//6.4.7.2//EN
TZID:Europe/Paris
X-WR-TIMEZONE:Europe/Paris
BEGIN:VEVENT
UID:0-1037@lptms.universite-paris-saclay.fr
DTSTART;TZID=Europe/Paris:20250425T110000
DTEND;TZID=Europe/Paris:20250425T120000
DTSTAMP:20250422T085625Z
URL:http://www.lptms.universite-paris-saclay.fr/seminars/mlpp-seminars-bea
 trice-achilli-bocconi/
SUMMARY:MLP@P seminars : Beatrice Achilli (Bocconi) - IPhT\, amphi Bloch\, 
 CEA-Saclay - 25 Avr 25 11:00
DESCRIPTION:Life\, Death and Miracles of Diffusion Models\nBeatrice Achilli
  (Bocconi)\n\n\nSeminar of the Series MLP@P (Machine Learning Physics @ Pl
 ateau)\, joint with LISN and IPhT.\nWhere: IPhT\, Salle Itzykson *change o
 f place!*\n\nIn recent years\, generative diffusion models have emerged as
  powerful tools in unsupervised learning\, achieving impressive results in
  image\, text\, and data generation. In this presentation I will analyze f
 rom a geometric and statistical physics perspective how these models learn
 \, generalize\, and eventually memorize data. In particular\, I will highl
 ight how the intrinsic structure of manifold data impacts the dynamics of 
 the diffusion process. Our analysis reveals the existence of distinct phas
 es during the generative process\, in particular a manifold coverage phase
  where the diffusion process fits the distribution internal to the manifol
 d\, and a consolidation phase where the score becomes orthogonal to the ma
 nifold. By mapping diffusion models driven by the empirical score function
  onto the Random Energy Model (REM)\, we are able to characterize memoriza
 tion and generalization timescales. These insights clarify the role of dat
 a structure in mitigating the curse of dimensionality and contribute to a 
 deeper understanding of how diffusion models capture complex data distribu
 tions.\n
CATEGORIES:MLP@P
LOCATION:IPhT\, amphi Bloch\, CEA-Saclay\, CEA - Saclay\, Saclay\, France
GEO:48.73668;2.180033999999978
X-APPLE-STRUCTURED-LOCATION;VALUE=URI;X-ADDRESS=CEA - Saclay\, Saclay\, Fra
 nce;X-APPLE-RADIUS=100;X-TITLE=IPhT\, amphi Bloch\, CEA-Saclay:geo:48.7366
 8,2.180033999999978
END:VEVENT
BEGIN:VTIMEZONE
TZID:Europe/Paris
X-LIC-LOCATION:Europe/Paris
BEGIN:DAYLIGHT
DTSTART:20250330T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
END:DAYLIGHT
END:VTIMEZONE
END:VCALENDAR