Machine Learning Seminar

Speaker: Dr Ilija Vukotic, Computational Scientist, The University of Chicago and Enrico Fermi Institute

Title: ATLAS ML platform – physics and analytics applications

Refreshments available at 2:00 pm.

Date: Mon, 12 Feb 2018, 2:15 pm – 3:15 pm
Type: Seminar
Location: 1400 BPS Bldg.

Abstract:
ATLAS experiment is looking forward to its RUN3 and the HL-LHC both of which will dramatically increase luminosity delivered. This means not only a lot more data to handle but also much more complex data that will be harder to simulate and reconstruct. With a flat computing budget and break down of the Moore’s Law, it is clear that the new ways of doing things are needed. Plan is to use faster, ML methods, in data simulation, reconstruction, and analysis but also increase operational efficiency of the ATLAS distributed computing. For this reason we developed a ML platform that enables development of the new methods and can be used as a test ground for the new services. I will present one specific application that utilizes Generative Adversarial Networks to efficiently generate calorimeter showers. During the SC2017 we used it for hyperparameter tuning over hundreds of points, generated and analyzed billions of shower images on a distributed set of nodes equipped with consumer grade GPUs.

Speaker Bio:
Ilija Vukotic is a computational scientist at the University of Chicago and Enrico Fermi Institute. He did his HEP PhD at Humboldt University, Berlin, where he looked for evidence of quark-gluon plasma at HERA-B experiment, DESY. While at Linear Accelerator Laboratory, Orsay, France he looked for Higgs to four lepton decays in ATLAS experiment and got steeped in all aspects of ATLAS computing. Moving to University of Chicago, he started working on ATLAS Distributed Computing tasks. Currently he is a coordinating ADC analytics group, helping mine mountains of data we collect from all our computing systems. By trying to utilize experience in Physics, ML and distributed computing he hopes to help ATLAS deliver best possible physics results.