Lorentz group equivariant autoencoders.

Area of Science:

• High Energy Physics
• Machine Learning
• Particle Physics

Background:

• Machine learning models are increasingly used in high energy physics (HEP) for tasks like classification and anomaly detection.
• Existing models often lack inductive biases suitable for HEP data, such as equivariance to inherent symmetries, which can limit performance and interpretability.

Purpose of the Study:

• To develop a novel machine learning model, the Lorentz group autoencoder (LGAE), incorporating equivariance to the proper, orthochronous Lorentz group.
• To evaluate the performance of the LGAE against baseline models on HEP data, specifically on jet analysis at the Large Hadron Collider (LHC).

Main Methods:

• Developed an autoencoder architecture (LGAE) with a latent space structured according to the representations of the Lorentz group.
• Experimental validation on LHC jet data, comparing LGAE performance against graph and convolutional neural network baselines.

Main Results:

• The LGAE demonstrated superior performance compared to baseline models across various metrics, including compression, reconstruction, and anomaly detection.
• The equivariant nature of the LGAE facilitated improved analysis of the latent space, enhancing the explainability of discovered anomalies.

Conclusions:

• The Lorentz group autoencoder (LGAE) offers a powerful and interpretable approach for machine learning in high energy physics.
• Incorporating physical symmetries, such as Lorentz equivariance, is crucial for developing more effective and data-efficient models in HEP.