I work on randomized algorithms for scalable machine learning. By replacing expensive exact algorithms with lightweight approximate methods, we can substantially reduce the resources needed to run a program. Machine learning is an ideal application area because learning algorithms can adapt to the noise introduced by the approximation.
My current work is on efficient approximate algorithms for low-level building blocks of machine learning, such as kernel sums and near-neighbor search, as well as fast training and inference. I am particularly interested in simple methods with theoretical guarantees that also work well in a web-scale production environment.