Dynamics of Information in Neural Networks

A central theme is the interpretation of training and inference as dynamical systems. Parameters evolve during optimization, while representations evolve through depth, creating a natural framework for studying neural networks through time-dependent processes rather than only static input-output maps.

In this perspective, neurons or features can often be modeled as interacting particles whose collective behavior is shaped by architecture, attention mechanisms, and training algorithms. In suitable scaling limits, these systems give rise to continuum flow descriptions that capture the macroscopic behavior of deep networks.

The workshop aims to deepen the connection between learning theory and tools from analysis, probability, statistical mechanics, and mathematical physics, and to encourage progress toward a unified mathematical theory of learning and inference in modern neural network architectures.