Deep learning for deep underground physics

Gabriel Perdue, Fermilab

Neutrino event reconstruction is challenging. It is difficult to properly define the important features of an interaction a priori, which makes this problem a good candidate for new Deep Learning techniques, rather than the traditional method of specifying algorithms in advance and testing them on data.

Machine Learning (ML) refers to a family of techniques for inferring a program from data. ML algorithms allow the researcher to use data to construct a new algorithm designed to solve a specific problem. Deep Learning (DL) is a subset of ML that involves the application of many-layered neural networks — a family of ML algorithms inspired by the structure of the brain  — to solving problems. By using many layers of “neurons” in a network, DL allows the algorithm to build a hierarchical model of the problem. This largely relieves the need to do “feature engineering,” which is the process by which the algorithm designer decides what “features” of a problem are important for the ML algorithm.

On Monday, November 13, a group of us held a workshop at Fermilab on Deep Learning for the DUNE experiment in order to start a dialogue and begin collaborating on tools to better enable the use of these techniques. The workshop chairs included Alexander Radovic (College of William and Mary), Robert Sulej (Fermilab/NCBJ Warsaw), Dorota Stefan (CERN/NCBJ), Kazuhiro Terao (SLAC), and myself. In all, about a dozen collaborators participated.

Fueled by successful technology demonstrations from NOvA, MicroBooNE, and ProtoDUNE, this meeting proved very productive.

Courtesy Kazuhiro Terao (SLAC) and MicroBooNE.

NOvA has published very high-profile physics results with DL for event selection, which has really validated the technique in our field. ProtoDUNE demonstrated great success in integrating DL into the reconstruction chain and showed great progress in the technical integration of DL libraries into LArSoft, along with interesting speed and performance benchmarks. The experiments used open source ML libraries for the really heavy math and built their own application code on top.

Finally, MicroBooNE showcased both a suite of software tools that are well-documented and look to generalize extremely well into the DUNE software stack, and impressive results in what we call “semantic segmentation” – or, in physics parlance, labeling each hit in an event by the source particle that created the deposition. The ProtoDUNE results were very encouraging for this sort of reconstruction as well.

Now well informed about on-going efforts, we emerged from the workshop confident that event reconstruction in DUNE will be able to rely heavily on DL. There have just been too many successes!