How does renormalization group apply to modern data science?
The renormalization group has found surprising applications in data science and machine learning. The idea of coarse-graining—reducing complexity while preserving essential features—is central to deep learning, where layers of a neural network extract hierarchical features. In image recognition, for instance, you start with pixels and then combine them into edges, shapes, and objects—much like blocking spins in the Ising model. The concept of a fixed point appears in training dynamics, where the network converges to a universal behavior independent of initial conditions. The correlation length in physics becomes a measure of long-range dependencies in data. It's all about scaling: finding the relevant operators that determine the outcome. The physics is in the universality class, and that's a powerful way to think about complex systems, whether they're magnets or datasets.
Ask Robert Wilson the follow-up →