- This event has passed.
Sivaraman BALAKRISHNAN (Carnegie Mellon University) – "On computationally-efficient robust and non-convex estimation"
Time: 2:00 pm – 3:15 pm
Date: 28th of May 2018
Place: Room 3001.
Sivaraman BALAKRISHNAN (Carnegie Mellon University) – “On computationally-efficient robust and non-convex estimation“
Classical statistical estimators have long been understood to be extremely susceptible to outliers and model misspecification, and this insight led to the development of the field of robust statistics. While statistical issues in robust estimation, i.e. designing and characterizing optimal estimators, are mostly well-understood several open questions remain in designing computationally-efficient practical estimators.
Building on recent work, that provided novel estimators for robust mean estimation, we provide a new class of computationally-efficient class of estimators for risk minimization that are provably robust in a variety of settings. Our workhorse is a novel robust variant of gradient descent, and we provide conditions under which our gradient descent variant provides accurate and robust estimators in any general convex risk minimization problem. These results provide some of the first computationally tractable and provably robust estimators for many canonical statistical models.
If time permits I will connect some of our recent work on robust parametric estimation to past work on establishing rigorous global convergence guarantees for estimation in latent variable models via non-convex likelihood maximization.
Cristina BUTUCEA, Alexandre TSYBAKOV, Eric MOULINES, Mathieu ROSENBAUM