Career Fair and Career Events

Title: Learning in Dynamic Environments

Abstract: Much of modern machine learning deals with decision making in static environments: after choosing some fixed prediction rule based on past data, this rule is deployed to make decisions in the future. However, this paradigm does not capture many(and perhaps even most) practical applications. For example, past data frequently does not perfectly model the future - imagine a book recommendation service for which user interests might change over time. Even more disturbingly, it can easily occur that the decisions made in the past actually affect the distribution of future data. In these situations, we need learning algorithms that can quickly and accurately adapt to a changing environment. In this talk I will describe recent work on algorithms that address this and other needs in sequential decision making.

Bio: Ashok Cutkosky is an assistant professor at Boston University in the ECE department since the fall of 2020. Previously, he was a research scientist at Google. He earned aPhD in computer science at Stanford University in 2018 under the supervision of Kwabena Boahen, as well as a masters in medicine. His current research focuses on machine learning and stochastic optimization theory. He designs training algorithms that achieve statistically optimal convergence rates without requiring manual hyper parameter tuning or prior knowledge of any statistical properties of the data and yet are efficient enough to be of practical use on the gigantic datasets and models in use today.