binary junipr: An Interpretable Probabilistic Model for Discrimination.

Area of Science:

• Particle Physics
• Machine Learning
• Unsupervised Learning

Background:

• Jets in particle physics are complex phenomena.
• Unsupervised learning offers a powerful framework for analyzing particle collision data.
• Probabilistic models can capture the intricate structure of particle jets.

Purpose of the Study:

• To refine the JUNIPR unsupervised learning approach for improved jet classification.
• To optimize discrimination power in particle physics event analysis.
• To provide physical insights into classification mechanisms.

Main Methods:

• Developed Binary JUNIPR, a classification-context refined JUNIPR model.
• Trained separate JUNIPR models for different event or jet types.
• Utilized relative probabilities from probabilistic models for discrimination.

Main Results:

• Binary JUNIPR achieves state-of-the-art performance in quark-gluon discrimination.
• Binary JUNIPR demonstrates state-of-the-art performance in top tagging.
• Trained models provide physical insights into classification strategies.

Conclusions:

• Refined JUNIPR models (Binary JUNIPR) significantly enhance classification power.
• The approach offers a novel method for analyzing and understanding particle jets.
• Enables detailed exploration of differences between particle jet types and simulation methods.