EVENT DETAILS AND ABSTRACT

Title: Emergent Behaviors in Large Neural Networks
Speaker: Boris Hanin
Speaker Info: Princeton University
Brief Description:
Special Note:
Abstract:

Neural network based algorithms have achieved specular results in domains like natural language processing (e.g. ChatGPT), computer vision (e.g. self-driving cares), and structural biology (e.g. AlphaFold). However, the complexity and scale of today's networks is unprecedented, leading to procedures of almost alchemical complexity for figuring out how to use them in practice. This is both practically inefficient and conceptually unsatifying. A key missing ingredient, which is actively being developed, is a rigorous mathematical theory for the emergent properties of large scale networks that would roughly do for neural what thermodynamics did for the study of engines. In this talk, I will highlight some key questions such a theory must address and their relation to problems in probability, functional analysis, and mathematical physics.