Steven Farrell

Biographical Sketch

Steve is a Machine Learning Engineer in the Data and Analytics Services group at NERSC. He supports machine learning and deep learning workflows on the NERSC supercomputers and collaborates with scientists for applied ML research.

Background

Steve's background is in high energy experimental particle physics. As an undergrad in Minnesota, he worked on the MINOS experiment, SNEWS, and CLEAR. As a Ph.D. student at UC Irvine, he joined the ATLAS experiment at CERN, where he worked on searches for Supersymmetry. Finally, as a Postdoc at Berkeley Lab in the Physics Division, Steve worked on software and computing for the ATLAS experiment and machine learning R&D for HEP.

Supporting Deep Learning at NERSC

Steve maintains the Deep Learning software stack at NERSC, including Intel-optimized Tensorflow and PyTorch, scalable libraries for training such as Horovod and the Cray PE ML Plugin, and Jupyter notebook solutions for distributed ML on the Cori supercomputer. He is also compiling and maintaining a set of Deep Learning science benchmark applications for NERSC, to characterize the supercomputer systems and to guide optimization efforts to ensure that scientific applications run smoothly and efficiently. Finally, Steve provides training to the community through documentation, blog posts, workshops, and tutorials.

Deep Learning for Science

Keynote presentation at SEA 2019 conference: https://sea.ucar.edu/event/deep-learning-science-capabilities-and-challenges-transforming-scientific-workflows

Slides: https://drive.google.com/open?id=1blxnMrcTFW0OcJOhOHFJ-8JQdd5VoM-Z

Deep Learning for HEP Analysis and Simulation

Deep Neural Networks for Physics Analysis on low-level whole-detector data at the LHC, https://arxiv.org/abs/1711.03573

Next generation generative neural networks for HEP, my plenary talk at CHEP 2018: https://indico.cern.ch/event/587955/contributions/2937509/

Deep Learning for Particle Track Reconstruction

I'm a member of the HEP.TrkX project (https://heptrkx.github.io/) and have developed a Graph Neural Network application for finding tracks in LHC experiments.

A few select references:

The TrackML Kaggle Challenge, https://www.kaggle.com/c/trackml-particle-identification

Novel Deep Learning Methods for Track Reconstruction, a contributed talk at CTD 2018, https://indico.cern.ch/event/658267/contributions/2881175/. Paper: https://arxiv.org/abs/1810.06111

“Convolutional Neural Networks for Particle Tracking”, invited talk at The 3rd International Workshop on Data Science in High Energy Physics, Fermilab.

S. Farrell et al., “The HEP.TrkX Project: deep neural networks for HL-LHC online and offline tracking,” EPJ Web Conf. 150, 00003 (2017).