• Calendar

November 16, Tuesday
12:00 – 14:00

Supervised Learning problems often involve inference of complex structured labels such as image segmentations of grid shape graphs. To achieve high accuracy in these tasks, one is often interested in introducing dependencies between label parts. However this usually results in inference problems that are NP hard. A natural approach is to use tractable approximation of the inference problems. In this talk I will present our recent work on approximate inference, using duality to extend the belief propagation algorithms to convex programs. Specifically, we 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. Also, I will present how duality and local inference can be applied to approximate current learning frameworks of conditional random fields (CRFs) and structured support vector machines (SVMs), and show highly scalable message-passing algorithms for these approximations. I will also present how these approximations can be applied to learn image segmentations. Based on joint work with: Amnon Shashua, Raquel urtasun, Ross Girshik