Title: Robustness in Sequential Learning: Demystifying Heavy Tails
Speaker: Dr. Shubhada Agrawal, Herbert A. Johnson Postdoctoral Fellow in ISyE at Georgia Tech.
Day, Date and Time: Monday 08th January, 2024, 10:30 am to 11:30 am
Venue: Seminar room IE 211, Second Floor, IEOR Building
Abstract: Numerous real-world machine learning applications involve making decisions sequentially through dynamic interactions with the environment in the presence of limited feedback. Tech giants like Google and Meta use sequential learning algorithms to generate billions in revenue via online advertising. While the potential applications of these algorithms in high-stakes fields such as medicine, defence, and autonomous vehicles are vast, their adoption to safety-critical domains is low due to a limited understanding of their behavior in diverse practical environments.
In this talk, we will look at designing robust algorithms that are safe for real-world applications. In particular, we will focus on the robustness against heavy-tailed uncertainty distributions. For concreteness, we will consider the classical regret minimization problem in the multi-armed bandit (MAB) setting with minimal assumptions about the arm distributions. We will discuss an optimal algorithm for this problem and the ideas involved in its design and analysis. A crucial component of such algorithms is the construction of tight confidence intervals for the mean using the collected samples. To this end, we will explore a novel concentration result for heavy-tailed distributions, which may be of independent interest to researchers and practitioners in related fields. Finally, I will conclude with some future directions I'm excited about and my broader research vision.
Bio: Dr. Shubhada Agrawal is a Herbert A. Johnson Postdoctoral Fellow in ISyE at Georgia Tech. Her academic journey began with an undergraduate degree in Mathematics and Computing from IIT Delhi, culminating in a PhD from the School of Technology and Computer Science at TIFR, Mumbai. She has extensively explored various problems within the multi-armed bandit (MAB) framework. Her doctoral work led to pioneering advancements in MAB problems, especially those involving heavy-tailed distributions. The impact of this work has been recognized internationally, with accolades including the Google PhD Fellowship in Machine Learning (2021), the Sarojini Damodaran Foundation Fellowship (2022), recognition as a Rising Star at the TCS-for-all Workshop during STOC 2023, and recognition as a US Junior Oberwolfach Fellow (2024). Lately, She is directing her efforts towards enriching the theoretical foundations of reinforcement learning (RL) algorithms and exploring the robustness of sequential decision-making algorithms across MAB and RL scenarios.