Graphical Models and Approximations
Justin Domke (NICTA)
NICTA SML SEMINARDATE: 2012-09-20
TIME: 11:15:00 - 12:00:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
The first part of this talk will describe "marginal based" learning of graphical model parameters. The idea is to fit a graphical model in terms of an application-specific loss function, taking into account both inference approximations (made for the sake of tractability) and modeling approximations (made out of ignorance). Results will be given on computer vision labeling/segmentation problems. The second part of the talk will be more unusual. I will discuss a set of problems of interest, briefly describing the motivation, possible strategies of attack, and the main apparent technical challenges. Rather than describing past work, the goal is to spark possible collaborations
