Physics Extraction Using Deep Learning with Uncertainty Quantification

Deep learning methods with uncertainty quantification were evaluated for the extraction of the mass and the Breit-Wigner parameters of the rho mass distribution in the context of two pion channels in photon-production. Our simplified model takes into consideration the relativistic Briet-Wigner distribution peak as well as background noise. This provides the final step in an end-to-end deep learning framework that is being actively developed by scientists at the Theory Center at Thomas Jefferson National Accelerator Facility for analysis of data from photon-hadron collisions. Deep neural networks combined with data resampling were used to quantify uncertainty in the mass (m_r) and width (Γ) predictions, and this method was compared to deep learning methods that incorporate uncertainty quantification into the training, such as approximate Bayesian Neural Networks and quantile neural networks. Results and comparisons of these methods will be presented.