Fall 2021 Kickoff Event on Robustness in High-dimensional Statistics and Machine Learning

Tuesday, Sept. 21, 2021

The Fall 2021 Special Quarter on Robustness in High-dimensional Statistics and Machine Learning is sponsored by The Institute for Data, Econometrics, Algorithms, and Learning (IDEAL), a multi-discipline, multi-institution collaborative institute that focuses on key aspects of the theoretical foundations of data science. The special-quarter activities include mini-workshops, seminars, graduate courses, and a reading group.  The research goal is to explore several theoretical frameworks and directions towards designing estimators and learning algorithms that are tolerant to errors, contamination, and misspecification in data.

The kick-off event for this quarter will be held on Tuesday, September 21, 2021 at 3 pm Chicago/Central time. We will briefly introduce the institute, the key personnel and information about the various activities during the special quarter. There will also be short research talks by the organizers of the special quarter. You can register for this event using the registration form. Please join us at the kick-off event!

Logistics

  • Date: Tuesday, Sept. 21st, 2021 from 3:00-5:00 PM
  • Location: Gather.town with Panopto streaming
  • Registration: Registered participants will get a link to the workshop by email.
  • WATCH FULL EVENT HERE

Schedule

Titles and Abstracts

 
Speaker: Chao Gao (Statistics, UChicago)
Title: Robust Regression with Contamination
Abstract: We study regression with contaminated observations. We will discuss different results depending on whether both the responses and the covariates are contaminated or only the responses are contaminated. General minimax rates are derived based on regression depth functions when both the responses and the covariates are contaminated. The result implies consistency is possible only if the contamination proportion is vanishing. In comparison, we show that when the covariates are clean, consistent robust regression is actually possible even when the contamination proportion approaches one. A near-optimal procedure in this case is the simple median regression. Applications of the second setting in model repair problems will also be discussed.
 
Speaker: Yu Cheng (UIC) 
Title: High-Dimensional Robust Statistics: Faster Algorithms and Optimization Landscape

Abstract: We study the fundamental problem of high-dimensional robust estimation when a constant fraction of the input samples are adversarially corrupted.  Recent work gave the first polynomial-time robust algorithms for a wide range of statistical and machine learning tasks with dimension-independent error guarantees.

In this talk, we will discuss two exciting new directions in the area of high-dimensional robust statistics: (1) designing faster algorithms for robust estimation, with the ultimate goal of matching the runtime of the fastest non-robust algorithms if possible, and (2) exploring more direct non-convex formulations of robust estimation and analyzing their optimization landscape. This talk is mostly based on joint works with Ilias Diakonikolas, Rong Ge, and Mahdi Soltanolkotabi.

 
Speaker: Aravindan Vijayaraghavan (TCS, NU)
Title: Algorithmic aspects of adversarial robustness
Abstract: Adversarial robustness measures the susceptibility of a machine learning algorithm to small perturbations made to the input at test time or at training time. This has attracted much interest on the empirical side however our theoretical understanding of adversarial robustness is limited. Requiring robustness to adversarial perturbations for many machine learning tasks leads to novel, challenging problems. I’ll give a flavor of these questions in the context of some basic problems in classification and unsupervised learning. I’ll follow this up with some questions and directions for future work that I am excited about.

Join Our Newsletter