BARC talk by Alessio Mazzetto

Tuesday, March 4, 2025, Alessio Mazzetto, PhD student at Brown University, USA, will give a talk on "Learning with Drifting Input Distributions".

Abstract:
We develop and analyze a general technique for learning with unknown distribution drift. Given a sequence of independent observations from the last T steps of a distribution that evolves over time, our algorithm addresses a learning problem with respect to the distribution's state at the final time step. In this non-stationary setting, older data may provide outdated information due to drift, making the selection of an appropriate recent data window crucial. Unlike previous work that relies on a priori assumptions about the magnitude of the drift, our algorithm dynamically adapts its window size based on the input. In particular, at each step, it selects a window of past observations that minimizes the trade-off between increased variance in the learning process from using fewer samples, and greater drift-induced error from including older, less relevant data. The challenge is that estimating the drift is impossible, as we may have only a single sample from each distribution. Nonetheless, we show that without explicitly estimating the drift, our method solves the learning problem with nearly the same error as an algorithm that knows the drift magnitude in advance. We demonstrate applications of this technique and establish matching lower bounds for problems such as binary classification, discrete distribution estimation, and vector quantization. 

Bio:
Alessio is a Computer Science Ph.D. student at Brown University with a research focus in Theoretical Computer Science, where he is advised by Eli Upfal. In Spring 2024, Alessio's work has been supported by the Kanellakis Fellowship. In Fall 2023, he was a co-instrucor for Advanced Introduction to Probability for Computing and Data Science (CS145). In Summer 2023, Alessio was an intern at Yahoo! Research on the Scalable Machine Learning team.
Alessio's main research area is Machine Learning Theory. He am broadly interested in learning settings where there is access to a small amount of data for the target task. In his work, he has developed theoretically sound methods that can quantify and use the knowledge provided by different sources other than labeled data for weak supervision. 
Prior to coming to Brown, he completed a Master in Computer Science and a Bachelors in Information Engineering from University of Padua in Italy.

Host:
Rasmus Pagh and Amartya Sanyal