Jordan's bigdata talk in Math/Stats series

Really excellent talk by Professor Michael Jordan from the Berkeley Statistics department visiting us at Davis yesterday. Haven’t seen a talk cover technical content with remarkable clarity for the algorithms and insights involved, while also covering such a breadth of material. My notes didn’t really keep up, but transcribing some of my scribbling into electronic form will have to wait until I have more time. (Their pubs are probably a better reference anyway.)

But to summarize,I enjoyed the bag of little bootstraps – a very simple, elegant, and powerful idea. Divide-and-conquer meets Efron.Kleiner et. al. (2011) 

I hadn’t realized NetFlix prize could be thought of a matrix completion problem with justifiable assumption of low-rank, and was impressed both by how simple the best algorithm is and by the 10 fold speedup obtained by their approximate divide-and-conquer solution. Mackey et. al. (2012) 

Highlights the use of Stein’s method of exchangeable pairs for giving you a differential-equation approach to proving things like central limit theorem, and a remarkably simple proof for a Bernstein-like large deviation result for matrices sounds pretty clever.