UC Irvine

Experimental physicists explore the fundamental nature of the universe by probing the properties of subatomic particles using specialized detectors. These detectors generate vast quantities of data that must undergo multiple stages of processing, such as dimensionality-reduction and feature extraction, before statistical analysis and interpretation. The processing steps are typically designed by physicists, using expert knowledge and intuition. However, this approach is human-limited and costly --- useful information is inevitably lost. Machine learning offers an alternative approach in which processing steps are instead learned from data. In particular, deep learning is the approach of learning many processing steps simultaneously. In this dissertation, we apply deep learning to a number of data analysis pipelines in experimental physics. We demonstrate that human-engineered processing steps can instead be learned from data, and that the deep learning approach can retain useful information from low-level data that is otherwise lost. Ultimately, this has the potential to both streamline data analysis pipelines and increase the statistical power of experiments, speeding up scientific discoveries.