Non-equilibrium physics of machine learning
The recent advances in artificial intelligence rely on large machine-learning models in the form of deep neural networks. In this lecture, we will study how and why such models work using tools from non-equilibrium statistical physics. To this end, we will treat neural networks as complex systems consisting of interacting agents and study collective phenomena thereof. The lecture will start with a general introduction to non-equilibrium physics, including stochastic processes, and machine learning. We will then introduce theoretical concepts from the theory of disordered systems and apply them to deep learning.