Date/Time
17 January 2027 - 22 January 2027
Organized by
Andrea Agazzi (UNIBE), Borjan Geshkovski (INRIA Paris), and Philippe Rigollet (MIT)
Event page & registration
https://indico.global/event/16529/
Description
Deep neural networks, and in particular transformer architectures, have become a cornerstone of modern machine learning, driving unprecedented advances in Artificial Intelligence (AI). Yet the mechanisms by which these models learn and process information through their depth remain poorly understood from a mathematical standpoint. This workshop aims to provide an interface for researchers in machine learning theory and mathematical physics/mathematics to investigate the mathematical foundations of information processing in deep neural networks, with a particular focus on transformer architectures. The natural point of contact between these fields lies in interpreting the evolution of parameters during training and of data representations during inference as dynamical systems, where neurons or features evolve as particles coupled through interaction terms resulting from the choice of network architecture or training algorithm. In appropriate scaling limits, these interacting particle dynamics give rise to continuous flow descriptions capturing the emergent, macroscopic behavior of deep networks. This connection provides a natural bridge between learning theory and established areas of mathematical physics, statistical mechanics, and analysis, which the workshop aims to strengthen by fostering collaborations toward a unified mathematical theory of learning and inference dynamics in modern neural network architectures
Location
SwissMAP Research Station, Les Diablerets, Switzerland
