Deep Learning in Particle Physics: A Philosophical Analysis

• Applicant: Dr. Martin King, Ph.D.
• Subject Area: Theoretical Philosophy
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 507919469

The lack of new physics discoveries at the LHC has had dramatic effects on the efforts of the particle physics community. Both the predictions of the Standard Model (SM), like the existence of the Higgs boson, as well as those of models of physics beyond the Standard Model (BSM) have strongly guided the course of particle physics research over the past decades. However, due to the lack of new discoveries, the top-down approach of hypothesis testing is no longer seen to be as promising as it once was. Instead, many physicists are hoping for a bottom-up approach to give them some guidance on where new physics may be lurking. This shift in methodology brings the opportunity to study the philosophical impacts of the major changes this is bringing about, including the re-evaluation of long-cherished guiding principles, such as naturalness, and the increase in model-independent approaches, such as the use machine learning (ML). This project will consist in two major subprojects. One will characterise the methodological shift in high energy physics, including the increase in SM precision measurements, the use of the SM effective field theory framework, and the increasing use of ML, in particular, of unsupervised deep neural networks (DNN). The motivation behind this shift is the diminishing hope for new physics discoveries at current and planned experiments. Without dramatic increases in energy, physicists need to be creative if they are to find discoveries or anomalies in existing data. Deep learning (DL) is particularly well suited to these kinds of searches. The second work package then explores the philosophical issues in using deep learning in high energy physics.High energy physics makes for an excellent case study for DL as it is only beginning to be treated philosophically and is unique is key ways:the experiments are of a quantum nature, which presents challenges to the labelling practices of classifiers; it has very strong demands on the required precision and the time-sensitivity of the outputs, which puts the latest techniques to the test; and the field is already equipped with the best tools and techniques for simulations and has incredible amounts of high-quality training data. The case study is then an excellent opportunity to gain philosophical insights into the following questions about DL: -is the strong data-drivenness and inductivism aimed at in these model-independent approaches possible and is it capable of discovering new physics?- Are there advantages to unsupervised learning and does this tell us anything about the role of physical reasoning and guiding principles?- Can the philosophical literature on explanation provide insights into the quest for explainable AI?- Is there something inherent in DL that precludes understanding?- Do the many steps in the data pipeline from choice of architecture and data acquisition through to final analysis tell us about the nature of scientific representation and understanding.

• DFG Programme: Research Grants