Abstract: Modern machine learning has achieved remarkable success across science and technology. Yet progress remains largely empirical, lacking a unified theory that explains how and why these systems work. In parallel, experimental neuroscience now enables large-scale recordings of brain activity, calling for quantitative theories that link microscopic neural interactions to macroscopic behavior. In this talk, I will use statistical physics as a unifying language to tackle these challenges through data-driven yet analytically tractable models. These models allow to identify interpretable summary statistics that capture collective neural computation. I will present applications of these methods to study the inductive biases of neural architectures, the structure of real datasets, and the learning curves of training algorithms viewed as controlled high-dimensional dynamical processes.

Host: Prof. Ernst Wit