B1201
Title: Fast inference with submodular functions
Authors: Stefanie Jegelka - MIT (United States) [presenting]
Abstract: Submodularity - also known as a discrete analogue of convexity - has been an important concept in areas like graph theory, game theory, and combinatorial optimization. Submodular functions also arise in various forms in statistics in machine learning. They may, for example, define probabilities in discrete graphical models and point processes, or structured norms for M-estimation. Submodularity will be introduced and recent work will be highlighted on exploiting properties of submodular functions and related polyhedra to solve inference problems, focusing in particular on the case of having a sum of submodular functions.
