Exploring the Intersection of Physics Modeling and Representation Learning

Abstract

Representation Learning has evolved into a multi-purpose tool capable of solving arbitrary problems provided enough data. This thesis focuses on two primary directions: (1) Harnessing the power of deep learning for applications in fundamental physics and (2) using physicsinspired tools to improve and shed some light on otherwise large-scale, inscrutable black-box algorithms. We explore a collection of applications that improve different aspects of nuclear and particle physics research across its many stages, from online data selection to offline data analysis. We also tease out how deep learning can open up entirely new avenues of research through the lens of mechanistic interpretability to (re)derive fundamental theory as well as new ways to reinterpret physics measurements. Lastly, we study how physics tools can be useful to better understand the dynamics of deep learning as well as provide a solid foundation for algorithms and training paradigms that expand the frontier of machine learning.