MSc. Konrad Becker and MSc. Maximilian Gaschler

T a l k : 5. June 2025

Statistical Description of Neural Networks

In this talk, we will present some basic concepts, which could bring us closer towards the ultimate goal of describing the dynamics of neural networks analytically. An important result in this direction is the DNN to Gaussian process correspondence for neural networks at infinite width. In this limit, the neural network dynamics can be fully described by a stochastic process, i.e., a collection of random experiments. We will explain the necessary details of neural networks and Gaussian processes to understand this result. In particular, we will introduce path integral methods for neural networks, which allow for a statistical field theory description of the dynamics.