Abstract:

Supervised Learning problems often involve inference of complex
structured labels such as image segmentations on grids. To achieve
high accuracy in these tasks, one is often interested in introducing
dependencies between local parts. However, this usually results in
inference problems that are NP hard. A natural approach is to rely on
tractable approximations to these inference problems.
In this talk I will present our recent work on approximate inference,
that uses duality to extend belief propagation algorithms to convex
programs. Specifically, I'll show how convex belief propagation
algorithms solve convex relaxations of the partition function, also
referred as the free energy, as well as linear programming relaxations
of integer linear programs. Importantly duality and local inference
can be applied to approximate current learning algorithms such as
conditional random fields (CRFs) and structured support vector
machines (SVMs). This results in highly scalable message-passing. I
will demonstrate these approximations for the task of image
segmentation.
If time permits I will also show how message-passing can be used to
solve graph based non-convex problems that involve continuous and
discrete labels, e.g., pose estimation, joint segmentation and stereo
reconstruction. 

Bio:
Tamir Hazan is a research assistant prof. in TTI Chicago. Tamir received his
B.Sc. in Computer Science and Humanities, and M.Sc and PhD degree in
Computer Science from the Hebrew University. Tamir's research areas include
machine learning and computer vision, recently focusing on graphical models,
primal-dual inference algorithms and beliefs propagation. He is also
interested in mixture models with various divergence measures, tensor
factorization and model selection, and support vector machines.