BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Creativity in diffusion models: Insights from statistical physics 
 and compositional grammars (Part 2) - Alessandro Favero (University of Cam
 bridge)
DTSTART:20251104T140000Z
DTEND:20251104T150000Z
UID:TALK240529@talks.cam.ac.uk
CONTACT:Sven Krippendorf
DESCRIPTION:How do generative AI systems\, such as diffusion models\, lear
 n to create new data? I will argue that natural data – such as images or
  text – can be described as hierarchical compositions of features\, whic
 h generative models learn to recombine in novel ways. To analyze this mech
 anism\, I will introduce ensembles of synthetic grammars as models of stru
 ctured data. Within this framework\, I will present a theory of compositio
 n that predicts a phase transition in the generative dynamics of diffusion
  models\, confirmed in modern architectures trained on real-world datasets
 . In the second part of the talk\, I will discuss how such grammars can be
  learned: what statistical correlations a learner can exploit to infer gra
 mmatical structure and the sample complexity required to do so. This analy
 sis shows that diffusion models progressively build internal representatio
 ns corresponding to increasingly abstract latent variables\, a procedure r
 eminiscent of the renormalization group in physics.
LOCATION:DAMTP\, MR11
END:VEVENT
END:VCALENDAR