Elchanan Mossel, Professor of Mathematics, MIT

Title: Data and Deep Learning

Abstract: Deep learning challenges much of the traditional thinking in statistics and machine learning, in particular the assumption that data is i.i.d. from families of distributions that are easy to describe. In the first (more scientific) part of the talk I will review some models of hierarchical or correlated data that may explain the need for depth and the efficiency of gradient based methods. In the second (more philosophical) part of the talk I will discuss the importance of data in assessing the extent to which the extraordinary performance of LLMs may be explained by novel reasoning versus plagiarism

3:30pm - Pre-talk meet and greet teatime - 219 Prospect Street, 13th floor, there will be light snacks and beverages in the kitchen area.