Chapter 15
Introduction and the JetNet Package

In the final Part of the dissertation, we discuss some more developments in machine learning and high energy physics, focusing primarily on jets. We first present the JetNet Python package in Chapter 15. In the same spirit as the eponymous dataset introduced in Chapter 9, it aims to increase the accessibility and reproducibility in our field by providing a standardized interface for accessing HEP datasets and benchmarking ML algorithms, as well as general utilities for ML and HEP. Since its introduction in $2021$, it has become widely adopted by the community, with over 50,000 downloads, and has been used extensively for many exciting developments in the field, as described below. By providing a common framework for jet datasets and evaluation metrics, it has also facilitated easy benchmarking and comparisons between different algorithms, particularly in the area of ML-based fast simulations, as discussed in Part IV.

We then conclude this Part, and the dissertation, by presenting the first Lorentz-group-equivariant autoencoder (LGAE) in Chapter 16. As detailed in Chapter 7.2, equivariant neural networks are extremely useful in the physical sciences, where data from sources such as molecules and high energy collisions naturally possess intrinsic physical symmetries, such as rotations, translations, and Lorentz-boosts. Incorporating such inductive biases of our data can lead to more data-efficient, interpretable, and performant AI algorithms. Indeed, we find that the LGAE outperforms baseline, non-Lorentz-equivariant, models on tasks of compression and anomaly detection for jets, provides a more interpretable latent space, and achieves high performance with a small fraction of the data needed to train CNNs.