Modern high energy physics sits at a unique intersection: we have strong theory priors, reliable simulators, and effectively “infinite” datasets—precisely the ingredients that have powered recent breakthroughs in other fields through the use of artificial intelligence (AI). In this talk, I argue for closing the loop between physics and machine learning (ML). Employing theoretical physics to understand what makes deep learning work, and using scalable learning systems to push beyond today’s bottlenecks in inference, analysis and theory calculation.

 I will present (i) ML theory-driven insight into optimization and robustness, that explains when and why networks learn to de-noise and generalize under data corruption with applications to physics problems (ii) a field-theoretic perspective on wide networks in which learning curves and spectral bias emerge from a continuum free-field limit, with extensions that treat finite-sample fluctuations as a form of random geometry (iii) generative modeling for first-principles physics, using jet data as…