Mean-Field Theory Insights into Neural Feature Dynamics - Lecture 2

Explore advanced theoretical insights into neural network behavior through mean-field theory in this lecture delivered by Cengiz Pehlevan at the International Centre for Theoretical Sciences. Delve into the mathematical foundations that govern neural feature dynamics and understand how infinite-width neural networks behave through the lens of statistical physics and probability theory. Learn how mean-field approaches provide powerful analytical tools for understanding the training dynamics and feature learning capabilities of deep neural networks. Examine the theoretical connections between neural network optimization and statistical mechanics, gaining insights into how networks evolve during training and how features emerge and interact. Discover how these theoretical frameworks can inform the design of more robust and adaptable machine learning systems. This presentation forms part of the comprehensive Data Science: Probabilistic and Optimization Methods II program, which brings together cutting-edge research in probability, optimization, and their applications to modern machine learning challenges.