BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CAIL//Events//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:CAIL Events
BEGIN:VEVENT
UID:2025-11-07-florentin-guth@cail.columbia.edu
DTSTAMP:20260823T060831Z
DTSTART:20251107T160000Z
DTEND:20251107T170000Z
SUMMARY:ML Seminar: Florentin Guth - Learning Normalized Probability Model
 s with Dual Score Matching
LOCATION:School of Social Work\, Room C03
DESCRIPTION:Learning probability models from data is at the heart of many 
 learning tasks. We introduce a new framework for learning normalized energ
 y (log probability) models inspired from diffusion generative models. The 
 energy model is fitted to data by two "score matching" objectives: the fir
 st constrains the gradient of the energy (the "score"\, as in diffusion mo
 dels)\, while the second constrains its time derivative along the diffusio
 n. We validate the approach on both synthetic and natural image data: in p
 articular\, we show that the estimated log probabilities do not depend on 
 the specific images used during training. Finally\, we demonstrate that bo
 th image probability and local dimensionality vary significantly with imag
 e content\, challenging simple interpretations of the manifold hypothesis.
 \n\nhttps://cail.columbia.edu/events/2025-11-07-florentin-guth
URL:https://cail.columbia.edu/events/2025-11-07-florentin-guth
END:VEVENT
END:VCALENDAR