BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20250918T093416EDT-7055oO33xj@132.216.98.100
DTSTAMP:20250918T133416Z
DESCRIPTION:Title: Tensor Programs and µP\n\nAbstract: We will discuss the 
 limiting behaviour of large neural networks as the layer width goes to inf
 inity. One of the factors that most affects limiting behaviour is the spec
 ific parametrization used\; apart from training stability\, this will dete
 rmine whether or not the neural network can learn features. We show a tech
 nique for mechanically deriving the 'best' parametrization\, known as µP. 
 As an additional empirical benefit\, we demonstrate that under µP\, hyperp
 arameters transfer directly across different sizes of models\, allowing fo
 r running all experiments at a small scale.\n
DTSTART:20230927T170000Z
DTEND:20230927T180000Z
LOCATION:Room 1214\, Burnside Hall\, CA\, QC\, Montreal\, H3A 0B9\, 805 rue
  Sherbrooke Ouest
SUMMARY:Andrew Mackenzie (9IÖÆ×÷³§Ãâ·Ñ)
URL:/mathstat/channels/event/andrew-mackenzie-mcgill-u
 niversity-351320
END:VEVENT
END:VCALENDAR