top of page

WP4 news: Collaborative efforts of T4.1

This news item from WP4 highlights the collaborative efforts of T4.1 with CERN openlab, whose focus of work coincides with the themes of CoE RAISE on many fronts. Some great examples of this come in the form of RAISE members presenting their work carried out in T4.1 at several CERN openlab lectures during the past months. CERN openlab is a unique public-private partnership between CERN and leading tech companies, which works to drive innovation in the computing technologies needed by CERN’s research community. It also includes the CERN openlab summer student programme which provides (under-)graduate students with an opportunity to work on an R&D project for nine weeks under the supervision of CERN experts.

Let us now take a closer look at a few of the lectures given by our colleagues in T4.1 of WP4. Firstly, Juan Pablo García Amboage, a CERN Technical Student, presented work that was carried out in the framework of T4.1 for his bachelor's thesis in computer science in the openlab series of the CERN Computing Seminars. The talk was titled Accelerating hyperparameter optimization using performance prediction on a heterogeneous  High Performance Computing system and covered a new algorithm for Hyperparameter Optimization (HPO) that we developed specifically to take advantage of both Quantum and Classical compute infrastructure. The algorithm, called Swift-Hyperband, builds upon the well known Hyperband algorithm, but is able to improve on both speed and model performance as well as enabling quantum-classical workflows. Swift-Hyperband can run both in a fully classical way on a normal computer as well as in a hybrid mode, where model performance prediction is carried out on a Quantum Annealer (QA) and the training on Deep Learning (DL) models is performed on GPUs in an HPC center.


Eric Wulff giving a lecture on HPO for DL models using HPC at the CERN openlab summer student lecture program

Secondly, Eric Wulff, Fellow at CERN and Task Leader of T4.1, gave a lecture on HPO for DL models using HPC at the CERN openlab summer student lecture program. The lecture started by giving an introduction to HPO, covering the what and why of tuning hyperparameters as well as some of the most popularly used hyperparameter search algorithms and trial scheduling strategies. Much attention was given to how Bayesian Optimization can be used to aid the HPO process. After this, Eric continued by presenting how HPO has been applied to optimize the Machine-Learned Particle Flow (MLPF) algorithm within T4.1 of CoE RAISE and shared his experience of implementing, using and evaluating different HPO techniques.

Lastly, David Southwick presented an overview of the challenges and integration efforts of HPC at CERN, again in the openlab series of the CERN Computing Seminars. The talk touched upon areas of work such as benchmarking, data transfers, job scheduling, and authentication & authorization. In addition, it gave an outlook on incorporating HPC centers and the heterogeneous resources they provide into HEP workflows.


All three lectures were recorded and the videos, along with the presenters’ slides, are available at the provided links for anyone interested in learning more about any of the fascinating topics that were discussed.

bottom of page