BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Talk of Professor Dan Alistarh (ISTA)
DTSTART:20240830T110000
DTEND:20240830T120000
DTSTAMP:20260504T041925Z
UID:6eae8cdbb789cee12a351ae30681c26ca05455909eb3c09eadccc923
CATEGORIES:Conferences - Seminars
DESCRIPTION:Professor Dan Alistarh\n\n\nTitle: Accurate Model Compression 
 at GPT Scale\n\n \n\nAbstract: A key barrier to the wide deployment of hi
 ghly-accurate machine learning models\, whether for language or vision\, i
 s their high computational and memory overhead. Although we possess the ma
 thematical tools for highly-accurate compression of such models\, these th
 eoretically-elegant techniques require second-order information of the mod
 el’s loss function\, which is hard to even approximate efficiently at th
 e scale of billion-parameter models. In this talk\, I will describe our wo
 rk on bridging this computational divide\, which enables the accurate seco
 nd-order pruning and quantization of models at truly massive scale. Compre
 ssed using our techniques\, models with billions and even trillions of par
 ameters can be executed efficiently on a few GPUs\, with significant speed
 ups\, and negligible accuracy loss. Models created using our techniques ha
 ve been downloaded millions of times from open-source repositories such as
  HuggingFace. \n\n \n\nBio: Dan Alistarh is a Professor at ISTA. Previou
 sly\, he was affiliated with ETH Zurich\, MIT\, and Microsoft Research\, h
 aving received his PhD from the EPFL under the guidance of Rachid Guerraou
 i. His research is on algorithms for efficient machine learning and high-p
 erformance computing\, with a focus on scalable DNN inference and training
 \, for which he was awarded ERC Starting and Proof-of-Concept Grants. In h
 is spare time\, he works with the ML research team at Neural Magic\, a sta
 rtup based in Boston\, on making compression faster\, more accurate and ac
 cessible to practitioners.  \n\n
LOCATION:MED 0 1418 https://plan.epfl.ch/?room==MED%200%201418
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR