Robust Jet Classifiers through Distance Correlation.

Area of Science:

• High Energy Physics
• Machine Learning
• Data Science

Background:

• Deep learning excels at supervised classification tasks in particle physics.
• Network prediction stability against input variations and systematic uncertainties is crucial for practical applications.
• Current decorrelation methods can be complex and difficult to train.

Purpose of the Study:

• To develop a simpler and more stable deep learning method for classification tasks.
• To improve the stability of network predictions against feature changes and systematic perturbations.
• To apply and evaluate a novel decorrelation technique in particle physics.

Main Methods:

• A new method based on distance correlation, a measure of nonlinear correlations, is proposed.
• The method is applied to recast a recent ATLAS study on boosted, hadronic W tagging.
• Distance correlation is explored as a regularization technique for convolutional neural networks and hadronic top tagging.

Main Results:

• The proposed distance correlation method achieves performance comparable to state-of-the-art adversarial decorrelation networks.
• The method demonstrates significantly improved training stability.
• The feasibility of using distance correlation for convolutional neural networks and top tagging is confirmed.

Conclusions:

• Distance correlation offers a simpler and more stable alternative to existing decorrelation methods in deep learning for particle physics.
• This technique enhances the reliability of deep learning models in high energy physics applications.
• The method shows promise for improving W and top tagging efficiency and robustness.