BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:AI Center Seminar - AI Fundamentals series - Dr. Pierre Marion
DTSTART:20260914T140000
DTEND:20260914T150000
DTSTAMP:20260922T044154Z
UID:79da6ed7a540dce301bbfa07a251771f109033b5421158c956759b92
CATEGORIES:Conferences - Seminars
DESCRIPTION:Pierre Marion\nThe talk is jointly organized by the EPFL AI 
 Center and the EPFL Statistical Physics of Computation Lab (SPOC) as part
  of the AI Fundamentals seminar series.\n\nHost: Prof. Lenka Zdeborova \n
 \nTitle \nEdge Flow: A Tractable and Predictive Continuous-Time Model for
  Gradient Descent at the Edge of Stability\n\nAbstract\nGradient descent i
 n deep learning may operate at the edge of stability (EoS)\, a regime in w
 hich the largest eigenvalue of the loss Hessian hovers near the stabilit
 y threshold 2/eta\, where eta is the learning rate. Classical analysis too
 ls such as gradient flow and the descent lemma do not apply here\, motivat
 ing the search for a continuous-time model valid at EoS. We propose Edge F
 low\, a system of three coupled ordinary differential equations that provi
 des a tractable\, faithful\, and predictive model of gradient descent dyna
 mics at EoS. Discretizing Edge Flow only requires two gradient evaluations
  and one Hessian-vector product at each iteration. We demonstrate empirica
 lly that Edge Flow tracks the dynamics of gradient descent at least as fai
 thfully as previously proposed continuous-time EoS models\, while in addit
 ion resolving the oscillation of the sharpness at the onset of EoS\, and t
 hat it provides a principled framework for understanding and mitigating in
 stabilities in this regime. Based on https://arxiv.org/abs/2606.18080.\n\
 nBio\nPierre Marion is a research faculty at INRIA within the SIERRA team\
 , which is a joint team between CNRS\, Ecole Normale Supérieure and INRIA
 . He was previously in 2024-25 a postdoctoral researcher at EPFL under the
  supervision of Lénaïc Chizat\, and before that a PhD student at Sorbonn
 e Université under the supervision of Gérard Biau and Jean-Philippe Vert
 . His research interests regard the theory of deep learning with an emphas
 is on generative models and optimization dynamics of neural networks\, as 
 well as AI for maths and reasoning.
LOCATION:ELE 117 https://plan.epfl.ch/?room==ELE%20117 https://epfl.zoom.u
 s/j/69068699604?pwd=NRPQlqXrIcVtpeNaoapJgz1VzLG3Su.1
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR