Ivan Butakov - Data Science
Date: - Location: Eurecom
Abstract: "In recent years, information theory has gained significant traction within the deep learning community, driving rigorous and profound analysis across various fields. This talk aims to provide a concise overview of some information-theoretic frameworks used in machine learning. The first part will be dedicated to the information bottleneck principle and related fundamental results (the "fitting-compression" hypothesis, grokking). Then, information-theoretic approaches to unsupervised and self-supervised learning will be discussed. The talk will also cover novel neural-based estimators and their relevance in current research." Bio: Ivan Butakov is a Junior Research Scientist and a PhD student at Skoltech, where he is currently working under the supervision of Professor Alexey Frolov. Ivan earned his Master's degree in Applied Mathematics and Physics from the Moscow Institute of Physics and Technology in 2024. Ivan's research focuses on the intersection of information theory, deep learning, and numerical mathematics. In his recent works, various applications of information theory to deep learning are explored, including the information bottleneck principle, mutual information estimation, and self-supervised learning techniques.
