Philosophy of Physics Seminar (Thursday - Week 7, MT25)

Abstract: Recent advances in machine learning, and neural networks in particular, raise important questions about the epistemic status of scientific understanding. In philosophy of physics, reduction has often been taken as a privileged route to explanatory depth: we understand one theory better when we can derive it from, or embed it within, another more fundamental framework. Yet neural networks achieve striking predictive successes without offering such derivations. This talk explores whether and how neural networks might contribute to the reductive understanding of physical theories. After reviewing canonical accounts of reduction and explanation in physics, I examine case studies where neural networks have been used to model inter-theoretic relations, for example, mapping between lattice models and continuum field theories, or learning renormalization group flows. I argue that while these methods fall short of the transparency traditionally associated with explanation, they nonetheless reveal structural features of theories that can support a novel, data-driven form of reductive understanding. The talk concludes by considering the implications of this perspective for debates about the nature of scientific understanding and the role of computational tools in contemporary physics.