Abstract: I will talk about distributionally robust machine learning, a principled approach for robust performance across subpopulations, and shifting distributions. We will focus on the superquantile, a.k.a. the Conditional Value at Risk (CVaR), which was popularized by the seminal work of UW’s own R. T. Rockafellar and co-authors in the field of computational finance and economics in the early 2000s.

We will first review the use of the superquantile for distributionally robust supervised learning. We will prove a generalization bound from first principles.

Second, we will discuss an application of the superquantile in the field of federated learning, i.e., the distributed training of machine learning models on mobile phones. We will quantify the extent to which a user conforms to the population distribution and show how the superquantile can be leveraged to improve performance on users who do not conform to the population. We will round of the discussion with a communication-efficient training algorithm and experimental results and heterogeneous datasets.

Bio: Krishna Pillutla is a 5th year Ph.D. student at the Paul G. Allen School of Computer Science and Engineering at the University of Washington, where he is advised by Zaid Harchaoui and Sham Kakade. Krishna is broadly interested in machine learning and optimization and works in the particular areas of structured prediction and federated learning. Krishna was a 2019-20 JP Morgan Ph.D. Fellow.