The use of adversaries for optimal neural network training

TL;DR

This paper introduces an adversarial neural network approach to improve signal-background discrimination in particle physics data, reducing correlations and outperforming traditional methods like Boosted Decision Trees and NeuroBayes.

Contribution

It presents a novel use of adversarial training to mitigate correlations in deep neural networks for particle physics analysis, enhancing discrimination performance.

Findings

Adversarial neural networks reduce correlation with the $E$ variable.

The adversarial approach outperforms Boosted Decision Trees.

The method surpasses NeuroBayes in discrimination accuracy.

Abstract

B-decay data from the Belle experiment at the KEKB collider have a substantial background from $e^{+}e^{-}\to q \bar{q}$ events. To suppress this we employ deep neural network algorithms. These provide improved signal from background discrimination. However, the deep neural network develops a substantial correlation with the $\Delta E$ kinematic variable used to distinguish signal from background in the final fit due to its relationship with input variables. The effect of this correlation is reduced by deploying an adversarial neural network. Overall the adversarial deep neural network performs better than a Boosted Decision Tree algorithimn and a commercial package, NeuroBayes, which employs a neural net with a single hidden layer.