Abstract

Following up on the overview talk on stability and privacy, I will discuss connections between another notion of stability, called forward stability, and privacy. Specifically, we will see how differentially private learning algorithms can be converted to online learning via a generalization of "follow the perturbed leader" framework. This leads to new algorithms for private analysis, as well as to new online learning algorithms. 

Based on work to appear at NIPS 2013.

Video Recording