BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:PDE-based Algorithms for Convolution Neural Network - Lars Ruthott
 o (Emory University)
DTSTART:20171030T140000Z
DTEND:20171030T145000Z
UID:TALK94015@talks.cam.ac.uk
CONTACT:INI IT
DESCRIPTION:This talk presents a new framework for image classification th
 at exploits the relationship between the training of deep Convolution Neur
 al Networks (CNNs) to the problem of optimally controlling a system of non
 linear partial differential equations (PDEs).  This new interpretation lea
 ds to a variational model for CNNs\, which provides new theoretical insigh
 t into CNNs and new approaches for designing learning algorithms.   We exe
 mplify the myriad benefits of the continuous network in three ways. First\
 , we show how to scale deep CNNs across image resolutions using multigrid 
 methods. Second\, we show how to scale the depth of deep CNNS in a shallow
 -to-deep manner to gradually increase the flexibility of the classifier. T
 hird\, we analyze the stability of CNNs and present stable variants that a
 re also reversible (i.e.\, information can be propagated from input to out
 put layer and vice versa)\, which in combination allows training arbitrari
 ly deep networks with limited computational resources.   This is joint wor
 k with Eldad Haber (UBC)\, Lili Meng (UBC)\, Bo Chang (UBC)\,  Seong-Hwan 
 Jun (UBC)\,  Elliot Holtham (Xtract Technologies)
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR