BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Quantization for matrix multiplication and LLMs
DTSTART:20260220T151500
DTEND:20260220T161500
DTSTAMP:20260416T181853Z
UID:a57b402afd3a1e9128ebb1c838be7fd733c3d313dfa8c9b5257da532
CATEGORIES:Conferences - Seminars
DESCRIPTION:Yury Polyanskiy\, MIT\nStatistical science teaches that parame
 ters can only be learned to within a certain precision due to stochasticit
 y of the measurement operations. Similarly\, modern AI research shows that
  LLM's parameters may be perturbed without significantly affecting the end
 -to-end performance. Given that the primitive operation that LLM rely on i
 s matrix multiplication\, a natural basic question is to understand the tr
 adeoffs and algorithms required for low-precision approximations for this 
 operation. In this talk we will discuss information-theoretic bounds\, met
 ric entropy and practical algorithms (NestQuant\, WaterSIC) that emerged a
 s by-products of the theory.
LOCATION:CM 1 517 https://plan.epfl.ch/?room==CM%201%20517
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR