BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/1700
DTSTAMP:20260310T053738Z
SUMMARY:Theoretical Physics for Robust\, Interpretable AI
DESCRIPTION:Speaker: Anindita Maiti (Perimeter Institute)\n\nAbstract: \n\n
 Despite rapid progress in state-of-the-art AI models for theoretical physi
 cs\, most such methods remain black boxes: lacking guarantees of robust an
 d reliable predictions that meet uncertainty quantification benchmarks ess
 ential in scientific domains. To address this gap\, I will present a few d
 irections that improve robustness\, mechanistic interpretability\, and unc
 ertainty quantification of complex learning and sample generation abilitie
 s\, by combining quantum and statistical field theories with computational
  statistics. First\, I will present the simplest model capable of in-conte
 xt learning\, an ability that underpins Large Language Model (LLM) success
 \, particularly for quantum. Leveraging Replica Mean Field Theory and Rand
 om Matrix Theory\, the performance of a simplified LLM is exactly derived 
 in the joint asymptotic limit of a large number of training samples\, toke
 n dimensions\, sample length\, and task diversity: exhibiting a phase tran
 sition in learning abilities. Next\, I will introduce Neural Network Field
  Theory Correspondence\, a paradigm which generates field theory samples w
 ithout any training algorithms\, while guaranteeing low uncertainty bounds
  at scale. This explainable + interpretable alternative to Monte Carlo sam
 pling facilitates a bidirectional mapping between field theory actions and
  their dual Neural Network architectures. Lastly\, I will present a framew
 ork for systematic coarsegraining of data features irrelevant to learning 
 objectives. Building on the Renormalization Group (RG)\, this scheme ensur
 es that perturbations to model predictions caused by such coarsegraining a
 re bound within scientific uncertainty measures\, while capturing nontrivi
 al corrections elusive to the state-of-the-art spectral bias method. Altog
 ether\, these Physics-of-AI approaches advance Scientific AI reliability i
 n a first-principles manner\, while bridging AI with fundamental physics.\
 n \n
URL:https://www.tcs.tifr.res.in/web/events/1700
DTSTART;TZID=Asia/Kolkata:20260311T150000
DTEND;TZID=Asia/Kolkata:20260311T160000
LOCATION:AG-69 and also via Zoom
END:VEVENT
END:VCALENDAR