BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:On the Bethe approximation - Adrian Weller (Columbia University)
DTSTART:20140811T100000Z
DTEND:20140811T110000Z
UID:TALK53562@talks.cam.ac.uk
CONTACT:Zoubin Ghahramani
DESCRIPTION:Belief propagation is a remarkably effective tool for inferenc
 e in graphical models\, even when applied to networks with cycles. A varia
 tional perspective shows that it may be viewed as a way to seek the minimu
 m of the Bethe free energy\, though it may converge only to a local optimu
 m or may not converge at all.\n\nWe shall cover a brief introduction to th
 ese ideas\, then go on to describe a recent algorithm we developed for any
  binary pairwise model which\, to our knowledge\, is the first to guarante
 e to return an epsilon-approximation to the _global minimum_ of the Bethe 
 free energy. The approach involves discretizing to yield a discrete optimi
 zation problem\, which may be viewed as multi-label MAP inference. If the 
 initial model is fully attractive\, this yields a fully polynomial-time ap
 proximation scheme (FPTAS).\n\nIf time\, we can also discuss work that fur
 ther explores the Bethe approximation and tries to tease apart the two way
 s it differs from exact inference: (i) the true entropy is approximated by
  the Bethe (pairwise) entropy\, and (ii) the minimization is performed ove
 r a relaxation of the marginal polytope (which enforces a globally consist
 ent probability distribution) termed the local polytope (which enforces on
 ly pairwise consistency). \n\nThis is joint work with Tony Jebara at Colum
 bia University.\n\nRelated papers:\nA. Weller and T. Jebara\, "Approximati
 ng the Bethe Partition Function" . Uncertainty in Artificial Intelligence 
 (UAI)\, 2014.\nA. Weller\, K. Tang\, D. Sontag and T. Jebara\, "Understand
 ing the Bethe Approximation: When and How can it go Wrong?" . Uncertainty 
 in Artificial Intelligence (UAI)\, 2014.\n
LOCATION:Engineering Department\, CBL Room BE-438.
END:VEVENT
END:VCALENDAR