Seminar Details

Mean estimation is a simple statistical task, solved optimally by a simple algorithm: the empirical mean. The story changes dramatically when the estimator is required to be robust. What exactly it means for an estimator to be robust can depend on the situation. It can be robust to model misspecification or adversarial contamination. It can be robust to outliers caused by heavy-tailed data. Or it can be robust in the sense of differential privacy. We will discuss how, surprisingly, we can design computationally efficient algorithms subject to all these types of robustness using the same underlying technical ideas, resulting in powerful estimators that are qualitatively quite different from the empirical mean. These connections offer glimpses of a broader landscape of robustness in statistical estimation.

A survey article accompanying this talk is published in IEEE BITS Magazine (https://arxiv.org/abs/2412.02670).