BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Effective Theories of Learning
DTSTART:20260507T113000Z
DTEND:20260507T123000Z
DTSTAMP:20260511T004900Z
UID:indico-event-1670960@indico.cern.ch
DESCRIPTION:Speakers: Noam Itzhak Levi\, Noam Levi (Tel Aviv University)\n
 \nModern AI systems are a transformative technological advancement\, but m
 any of their most important behaviors still lack simple organizing princip
 les. In this talk\, I will argue that physics offers a useful language for
  building effective theories of learning: identifying the right observable
 s\, isolating minimal solvable models\, and understanding which features o
 f data and learning dynamics are universal.I will first discuss the struct
 ure of natural data. I will describe how tools from statistical physics an
 d random matrix theory reveal universal structure in complex datasets. I w
 ill then show how diffusion models can be used as probes of hierarchical c
 ompositional structure: by partially noising and denoising data\, one can 
 expose latent features at different depths\, observe a semantic phase-tran
 sition\, and begin to reconstruct the organization of the data itself. I w
 ill next turn to scaling laws and reasoning. I will describe simple models
  in which test-time inference scaling can be understood by the lens of the
  effective difficulty of the questions beind asked.Finally\, I will discus
 s how these ideas connect to AI for physics\, namely\, ensuring that AI mo
 dels are truly learning underlying physical laws rather than pattern match
 ing to heuristics.\n\nhttps://indico.cern.ch/event/1670960/\n\nZoom: https
 ://cern.zoom.us/j/65562333965?pwd=VUcvclhVMnBkSXFlcTlCRVltd1VRUT09
LOCATION:4/2-037 - TH meeting room (CERN)
URL:https://indico.cern.ch/event/1670960/
END:VEVENT
END:VCALENDAR