BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Jim Halverson (Northeastern University) DTSTART:20200904T133000Z DTEND:20200904T143000Z DTSTAMP:20260423T022809Z UID:CMU-TP/17 DESCRIPTION:Title: Neural Networks and Quantum Field Theory\nby Jim Halverson (Northeaste rn University) as part of Carnegie Mellon theoretical physics\n\n\nAbstrac t\nWe propose a theoretical understanding of neural networks in terms of W ilsonian effective field theory. The correspondence relies on the fact tha t many asymptotic neural networks are drawn from Gaussian processes\, the analog of non-interacting field theories. Moving away from the asymptotic limit yields a non-Gaussian process and corresponds to turning on particle interactions\, allowing for the computation of correlation functions of n eural network outputs with Feynman diagrams. Minimal non-Gaussian process likelihoods are determined by the most relevant non-Gaussian terms\, accor ding to the flow in their coefficients induced by the Wilsonian renormaliz ation group. This yields a direct connection between overparameterization and simplicity of neural network likelihoods. Whether the coefficients are constants or functions may be understood in terms of GP limit symmetries\ , as expected from 't Hooft's technical naturalness. General theoretical c alculations are matched to neural network experiments in the simplest clas s of models allowing the correspondence. Our formalism is valid for any of the many architectures that becomes a GP in an asymptotic limit\, a prope rty preserved under certain types of training.\n LOCATION:https://researchseminars.org/talk/CMU-TP/17/ END:VEVENT END:VCALENDAR