Exploring Raw HEP Data using Deep Neural Networks at NERSC

Abstract

High Energy Physics has made use of artificial neural networks for some time. Recently, however, there has been considerable development outside the HEP community, particularly in deep neural networks for the purposes of image recognition. We describe the deep-learning infrastructure at NERSC, and analyses built on top of this. These are capable of revealing meaningful physical content by transforming the raw data from particle physics experiments into learned high-level representations using deep convolutional neural networks (CNNs), including in unsupervised modes where no input physics knowledge or training data is used. Here we describe in detail a project for the Daya Bay Neutrino Experiment showing both unsupervised learning and how supervised convolutional deep neural networks can provide an effective classification filter with significantly better accuracy than other machine learning methods. These approaches have significant applications for use in other experiments triggers, data quality monitoring or physics analyses.