Optimization, Learning and the Universality of Mirror Descent

Seminar
Tuesday, May 01, 2012
11:00am

I will discuss deep connections between Statistical Learning, Online Learning and Optimization.  I will show that there is a tight correspondence between the sample size required for learning and the number of local oracle accesses required for optimization, and thesame measures of "complexity" (e.g. the fat-shattering dimension or Rademacher complexity) control both of them.  Furthermore, I will show how the Mirror Descent method, and in particular its stochastic/online variant, is in a strong sense "universal" for online learning,statistical learning and optimization.  That is, for a general class of convex learning/optimization problems, Mirror Descent can alwaysachieve a (nearly) optimal guarantee.  In the context of statistical learning, this also implies that for a broad generic class of convexproblems, learning can be done optimally (in the worst-case agnostic-PAC sense) with a single pass over the data.

Speaker