BEGIN:VCALENDAR
PRODID:-//Columba Systems Ltd//NONSGML CPNG/SpringViewer/ICal Output/3.3-
M3//EN
VERSION:2.0
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
DTSTAMP:20200113T124515Z
DTSTART:20200211T120000Z
DTEND:20200211T130000Z
SUMMARY:Martin Benning - Deep learning as optimal control problems
UID:{http://www.columbasystems.com/customers/uom/gpp/eventid/}j7v-k5cfwvb
s-x7kurp
DESCRIPTION:Join us for this research seminar\, part of the SQUIDS (Stati
stics\, quantification of uncertainty\, inverse problems and data scienc
e) seminar series.\n\nAbstract: We consider recent works where deep neur
al networks have been interpreted as discretisations of an optimal contr
ol problem subject to an ordinary differential equation constraint. We r
eview the first order conditions for optimality\, and the conditions ens
uring optimality after discretisation. This leads to a class of algorith
ms for solving the discrete optimal control problem which guarantee that
the corresponding discrete necessary conditions for optimality are fulf
illed. The differential equation setting lends itself to learning additi
onal parameters such as the time discretisation. We explore this extensi
on alongside natural constraints (e.g. time steps lying in a simplex) an
d compare these deep learning algorithms numerically in terms of induced
flow and generalisation ability. We conclude by addressing the interpre
tation of this extension as iterative regularisation methods for inverse
problems.\nThis is joint work with Elena Celledoni\, Matthias J. Ehrhar
dt\, Brynjulf Owren and Carola-Bibiane SchĂ¶nlieb.
STATUS:TENTATIVE
TRANSP:TRANSPARENT
CLASS:PUBLIC
LOCATION:Frank Adams 1\, Alan Turing Building\, Manchester
END:VEVENT
END:VCALENDAR