IDT-UM common meeting of subject areas C & D 28-Mar-2019
Minutes: Florian Bernlochner, Martin Erdmann
C=Deep Learning, knowledge acquisition through data-driven methods
D=Event reconstruction, cost- and energy-efficient utilization of computing resources
1) Scope for both areas: generate progress in the experiment-specific challenges, transfer results into generalized & easy to digest ways to make results available to all partners.
For C we need small examples with source code & data to be executed on compute resources. These examples can be made available through VISPA at RWTH Aachen (presentation by Jonas Glombitza). At a later stage, all results can be moved to a Jupyter type system to be build e.g. at DESY. Martin will talk to Volker Gülzow.
For D the cooperative ACTS is the appropriate delivery space. Their resources are shared on Gitlab at CERN, but a discussion has started to move to an open platform as well.
2) Generative modeling: Thorben Quast presented the R&D results on simulating an electromagnetic calorimeter (T. Comput Softw Big Sci (2019) 3: 4). Using the Wasserstein distance, the generator not only learns to produce energy deposits in the different layers, but also correlations between layers. Interpolations between trained electron beam energies works well. Extrapolation outside the trained electron energies does not work. Small energy depositions get underestimated by the generator. Markus Elsing suggests to compare production times of generative modeling with human made parameterizations.
3) KIT progress: Florian presents the current efforts on algorithms in Karlsruhe. They actively participated in a Hackathon with the goal of track reconstruction. They will streamline results to get results into the ACTS system. They also succeeded in producing a benchmark in generative modeling based on the work of Thorben Quast et al. The goal is to now adapt this to a Belle II related example.
4) Discussion on common grounds:
Workpackage C1: Mainz is the only funded partner here, working on FPGA based jet reconstruction for ATLAS trigger. There is great interest within the IDT-UM community in a later tutorial how to get deep networks.
Workpackage C2/C3: here benchmark data sets will be very useful to measure the performance of the various network efforts. Markus Elsing: ATLAS could provide tracking events based on the Machine Learning Challenge using the fastgen feature of ACTS. Thorben Quast: could provide electromagnetic calorimeter events (GEANT4). Thomas Kuhr: can provide BelleII event four vectors, Martin expresses interest in cooperating wrt the new Lorentz Boost Network (arXiv:1812.09722). For generative modeling in Belle II the Cherenkov detector may be a good initial candidate (reduced complexity compared to calorimeter with timing information).
Workpackage C4: here exchange is expected on the methods to gain insights in the networks which is central to all efforts.
Workpackage D1: No news since the kick-off.
Workpackage D2: Also no news since the recent kick-off. Florian will start organizing with Markus Elsing, Andreas Salzburger, Heather Gray, and Simone Paganini the next tracking workshop / Hackathon, which will take place in Germany.