I spent Spring 2015 visiting colleagues at Microsoft Research, New York, which is basically a data science lab. I'll talk mostly about research results, but also a bit about the experience of being in an industry lab in contrast to an academic position (for students/postdocs who are considering this choice).

On the research side, I'll talk about progress we made in solving "learning to search" problems in a "bandit" setting. In particular, suppose you are designing the landing page for some web page, and you have to automatically choose what stories to show, what layout to use, how much space to allocate each story, what fonts to use, etc. This is a structured prediction problem. However, perhaps the only feedback you get from a user is something like clickthrough, which only gives you a very distal sense of whether you did a good job or not, to say nothing of which parts of your predicted layout were good. We design algorithms that can solve this problem efficiently with a small(ish) number of samples. Furthermore, we show that you can easily program models trained using these algorithms (somewhat akin to probabilistic programming, but without the "probabilistic") and that when used in a full information setting (as opposed to a bandit setting), this approach is state-of-the-art for several structured prediction tasks.

This is joint work with: Alekh Agarwal, Alina Beygelzimer, Kai-Wei Chang, He He, Akshay Krishnamurthy, John Langford, Paul Mineiro and Stephane Ross.