Neural Spectral Bias and Conformal Correlators

by Prof. Andreas Stergiou (King's College London)

Thursday 14 May 2026, 15:00 → 17:00 Europe/Rome

Can a simple neural network deduce the full functional form of a conformal field theory (CFT) correlation function from almost nothing? In this talk, I will demonstrate that by training a neural network using only crossing symmetry, correlator asymptotics, and a single “anchor point”, one can successfully recover correlators across a wide range of theories. I will argue that this success is driven by the spectral bias of neural networks, which inherently focuses on maximally smooth functions. This suggests a potential new aspect of CFTs: their correlators are among the smoothest crossing-symmetric functions.