BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:AI-for-Physics needs Physics-for-AI
DTSTART:20260407T103000Z
DTEND:20260407T113000Z
DTSTAMP:20260503T180500Z
UID:indico-event-9315@scitalks.tifr.res.in
DESCRIPTION:Speakers: Anindita Maiti (Perimeter Institute for Theoretical 
 Physics)\n\nNext-generation discoveries in quantum computing and particle 
 physics require large-scale data generation and classification strategies\
 , with theoretical support for its precision and accuracy. While such need
 s are empirically addressed by state-of-the-art Artificial Intelligence (A
 I) and Machine Learning (ML)\, theoretical support for robust data generat
 ion via reliable algorithms\, that meet scientific uncertainty quantificat
 ion benchmarks\, are still majorly lacking. The performance of AI/ML hinge
 s on an ability to classify and discard data features as per a hierarchy o
 f their relevance\, leading to emergent behaviours e.g. neural scaling law
 s\, phase transitions\, and universality – mirroring phenomena in theore
 tical physics. I will present a few directions where tools of theoretical 
 physics\, such as statistical physics\, field theory\, and renormalization
  (RG) etc.\, – leveraged to study AI/ML at a microscopic level – have 
 predicted macroscopic model behaviour e.g. out-of-distribution generalizat
 ion\, uncertainty quantification\, and neural scaling laws\, beyond the sc
 ope of traditional statistical means. Going beyond improved AI/ML explaina
 bility and robustness\, these works enable interpretation of field theorie
 s in terms of Neural Network statistics\, quantify model error propagation
  as RG flows\, and provide theoretical benchmarking via neural scaling law
 s for AI/ML performance on data from Rydberg atom arrays.\n\nhttps://scita
 lks.tifr.res.in/event/9315/
LOCATION:AG69 and zoom
URL:https://scitalks.tifr.res.in/event/9315/
END:VEVENT
END:VCALENDAR