Machine learning uncertainties with adversarial neural networks.

Area of Science:

• Particle Physics
• Machine Learning
• Computational Physics

Background:

• Machine learning excels at identifying correlations in complex data.
• Predicting outcomes from theoretical models is challenging, especially with unknown dynamics.
• Systematic and theoretical uncertainties are crucial in data analysis.

Purpose of the Study:

• To develop a machine learning approach for reliable event classification.
• To enable novel methods for parameter fitting in particle physics data.
• To integrate known uncertainties directly into the training process.

Main Methods:

• Utilized adversarial networks for machine learning model training.
• Incorporated a priori known systematic and theoretical uncertainties during training.
• Applied the method to analyze Higgs boson production in association with jets within effective field theory extensions.

Main Results:

• Achieved more reliable event classification on an event-by-event basis.
• Demonstrated novel approaches for performing parameter fits on particle physics data.
• Successfully showcased the benefits of the adversarial network approach in a specific physics scenario.

Conclusions:

• Adversarial networks offer a robust framework for handling uncertainties in machine learning for physics.
• The proposed method enhances the precision of event classification and parameter estimation.
• This approach facilitates deeper insights into theoretical models like effective field theories.