BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Isaac Newton Institute Seminar Series
SUMMARY:Learned forward operators: Variational regularizat
ion for black-box models - Jonas Adler (KTH - Roya
l Institute of Technology )
DTSTART;TZID=Europe/London:20171031T154000
DTEND;TZID=Europe/London:20171031T160000
UID:TALK94123AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/94123
DESCRIPTION:In inverse problems\, correct modelling of the for
ward model is typically one of the most important
components to obtain good reconstruction quality.
Still\, most work is done on highly simplified for
ward models. For example\, in Computed Tomography
(CT)\, the true forward model\, given by the solut
ion operator for the radiative transport equation\
, is typically approximated by the ray-transform.
The primary reason for this gross simplification i
s that the higher quality forward models are both
computationally costly\, and typically do not have
an adjoint of the derivative of the forward opera
tor that can be feasibly evaluated. The communit
y is not un-aware of this miss-match\, but the wor
k has been focused on &ldquo\;the model is right\,
lets fix the data&rdquo\;. We instead propose goi
ng the other way around by using machine learning
in order to learn a mapping from the simplified mo
del to the complicated model using deep neural net
works. Hence instead of learning how to correct co
mplicated data so that it matches a simplified for
ward model\, we accept that the data is always rig
ht and instead correct the forward model. We the
n use this learned forward operator\, which is giv
en as a composition of a simplified forward operat
or and a convolutional neural network\, as a forwa
rd operator in a classical variational regularizat
ion scheme. We give a theoretical argument as to w
hy correcting the forward model is more stable tha
n correcting the data and provide numerical exampl
es in Cone Beam CT reconstruction.
LOCATION:Seminar Room 1\, Newton Institute
CONTACT:INI IT
END:VEVENT
END:VCALENDAR