BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20250824T112345EDT-5165jzvhTz@132.216.98.100
DTSTAMP:20250824T152345Z
DESCRIPTION:Depth-Adaptive Neural Networks from the Optimal Control viewpoi
 nt\n\nAbstract:\n	In recent years\, deep learning has been connected with o
 ptimal control as a way to define a notion of a continuous underlying lear
 ning problem. In this view\, neural networks can be interpreted as a discr
 etization of a parametric Ordinary Differential Equation which\, in the li
 mit\, defines a continuous-depth neural network. The learning task then co
 nsists in finding the best ODE parameters for the problem under considerat
 ion\, and their number increases with the accuracy of the time discretizat
 ion. Although important steps have been taken to realize the advantages of
  such continuous formulations\, most current learning techniques fix a dis
 cretization (i.e.~the number of layers is fixed). In this work\, we propos
 e an iterative adaptive algorithm where we progressively refine the time d
 iscretization (i.e.~we increase the number of layers). Provided that certa
 in tolerances are met across the iterations\, we prove that the strategy c
 onverges to the underlying continuous problem. One salient advantage of su
 ch a shallow-to-deep approach is that it helps to benefit in practice from
  the higher approximation properties of deep networks by mitigating over-p
 arametrization issues. The performance of the approach is illustrated in s
 everal numerical examples.\n\nFor Zoom meeting information please contact 
 tim.hoheisel [at] mcgill.ca\n
DTSTART:20210322T200000Z
DTEND:20210322T210000Z
SUMMARY:Olga Mula (Paris Dauphine)
URL:/mathstat/channels/event/olga-mula-paris-dauphine-
 329615
END:VEVENT
END:VCALENDAR