BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:EPFL CIS – RIKEN AIP Seminar Series by Prof. Nicolas Flammarion\
 , Tenure-track assistant professor in computer science at EPFL
DTSTART:20211027T100000
DTEND:20211027T110000
DTSTAMP:20260506T084508Z
UID:8822bb9435bd9499d06aa130a75e1f41939f442e64094b612c4b364b
CATEGORIES:Conferences - Seminars
DESCRIPTION:Prof. Nicolas Flammarion\nGet your Zoom link:https://c5dc59ed9
 78213830355fc8978.doorkeeper.jp/events/128209\n\nTitle: Implicit Bias of S
 GD for Diagonal Linear Networks: a Provable Benefit of Stochasticity\n\nAb
 stract: Understanding the implicit bias of training algorithms is of cruci
 al importance in order to explain the success of overparametrised neural n
 etworks. In this talk\, we study the dynamics of stochastic gradient desce
 nt over diagonal linear networks through its continuous time version\, nam
 ely stochastic gradient flow. We explicitly characterise the solution chos
 en by the stochastic flow and prove that it always enjoys better generalis
 ation properties than that of gradient flow. Quite surprisingly\, we show 
 that the convergence speed of the training loss controls the magnitude of 
 the biasing effect: the slower the convergence\, the better the bias.\nOur
  findings highlight the fact that structured noise can induce better gener
 alisation and they help to explain the greater performances observed in pr
 actice of stochastic gradient descent over gradient descent.\n\nBio: Nicol
 as Flammarion is a tenure-track assistant professor in computer science at
  EPFL. Prior to that\, he was a postdoctoral fellow at UC Berkeley\, hoste
 d by Michael I. Jordan. He received his PhD in 2017 from Ecole Normale Sup
 erieure in Paris\, where he was advised by Alexandre d’Aspremont and Fra
 ncis Bach.\nHis research focuses primarily on learning problems at the int
 erface of machine learning\, statistics and optimization.\n 
LOCATION:By Zoom https://c5dc59ed978213830355fc8978.doorkeeper.jp/events/1
 28209
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR