BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning and inference in probabilistic submodular models - Andrea
 s Krause (ETH Zürich)
DTSTART:20180116T114500Z
DTEND:20180116T123000Z
UID:TALK97636@talks.cam.ac.uk
CONTACT:INI IT
DESCRIPTION:I will present our work on inference and learning in discrete 
 probabilistic models defined through submodular functions.  These generali
 ze pairwise graphical models and determinantal point processes\, express n
 atural notions such as attractiveness and repulsion and allow to capture r
 ichly parameterized\, long-range\, high-order dependencies. The key idea i
 s to use sub- and supergradients of submodular functions\, and exploit the
 ir combinatorial structure to efficiently optimize variational upper and l
 ower bounds on the partition function.  This approach allows to perform ef
 ficient approximate inference in any probabilistic model that factorizes i
 nto log-submodular and log-supermodular potentials of arbitrary order.  Ou
 r approximation is exact at the mode for log-supermodular distributions\, 
 and we provide bounds on the approximation quality of the log-partition fu
 nction with respect to the curvature of the function.  I will also discuss
  how to learn log-supermodular distributions via bi-level optimisation. In
  particular\, we show how to compute gradients of the variational posterio
 r\, which allows integrating the models into modern deep architectures.  T
 his talk is primarily based on joint work with Josip Djolonga 
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR