Tutte seminar - Nick Harvey

Learning Submodular Functions

Abstract:

A central topic in computational learning theory is learning Boolean functions defined on the Boolean cube. We consider a less-studied problem: learning real-valued function on the Boolean cube. Without imposing any structure on these functions, learning is impossible, so we consider the class of submodular functions. This is a very rich class of functions, in many ways similar to the class of convex functions. 

We show that it is possible to "approximately" learn submodular functions, and we prove nearly matching upper- and lower-bounds on the approximation ratios that can be achieved. 
The talk will be accessible to a general mathematical audience. We will not assume any background in learning theory or submodular functions. 

This is joint work with Maria-Florina Balcan at Georgia Tech.