Abstract: Machine learning has been used in high-energy physics for decades, and today we see plentiful examples of neural networks designed for detector operation and analysis tasks. However, as these tools find their ways into more and more physics analyses, questions about their functionality and interpretability remain.

In this talk, I will give an overview of the Lorentz Group Network, a new neural network design. By explicitly respecting the symmetries of the Lorentz group—which underpin particle physics—this network provides a generalized, physics-inspired architecture for use in classification and regression tasks, with improved prospects of physical interpretability. I will outline the network’s design, and present studies and results of its performance in top-quark tagging.