On Tuesday 15th and Wednesday 16th of August, the CERN openlab 2023 summer students will present their work at two dedicated public Lighting Talk sessions.
In a 5-minute presentation, each student will introduce the audience to their project, explain the technical challenges they have faced and describe the results of what they have been working on for the past two months.
It will be a great opportunity for the students to showcase the progress they have made so far and for the audience to be informed about various information-technology projects, the solutions that the students have come up with and the potential future challenges they have identified.
Please note
The Openlab network leverages a range of computing resources, spanning cloud-based quantum computing and AI infrastructure, on-site servers, and personal devices. My project aims to streamline the tasks of systems administrators by developing a web portal for centralized control and data visualization of these nodes.
The hls4ml project is a mature library for deployment of neural networks on FPGAs used by the L1 trigger systems of the LHC experiments. With its support for multiple neural network architectures and FPGAs from multiple vendors, the library has recently seen an increase in adoption among the LHC experiments with multiple projects in various stages of development. To meet the latency constraints of the trigger systems the neural networks need to be compressed and fine-tuned for the target FPGA hardware, requiring multiple time-consuming firmware synthesis runs. To facilitate further rapid prototyping of neural networks with hls4ml in this project we will aim to speed up the synthesis flow. My project is on enhancing the internals of hls4ml to support dividing the neural network into blocks that can be synthesized in parallel and reassembled into the final firmware for deployment on hardware. The result of this work will make hls4ml a more usable library, reduce the development lifecycle and foster adoption by the wider scientific community.
The interpretability of machine learning models remains a critical yet elusive aspect of contemporary computational science. In this presentation, I specifically explore the interpretability of machine learning algorithms by applying self-attention maps to the AtmoRep large-scale weather prediction model. By leveraging self-attention mechanisms, I present a method to analyze the internal structure and dependencies within the model's layers. This technique enables me to interpret the intricate relationships between meteorological variables and the resultant predictions. The application of self-attention maps presents an essential step towards a more transparent and scientifically rigorous approach to interpreting large-scale weather modeling, offering potential implications for advancements in climate science and meteorological forecasting.
Relevant buzzwords: AI, ML, HPC, Cloud Computing, Transformer, All You Need, Digital Twin, Computer Vision, Big Data
Irrelevant buzzwords: Blockchain, IOT, VR, AR, Quantum Computing, QML, FCC, Exascale, NLP, Beyond the Standard Model
The project focuses on improving usability and inclusivity of the single sign on for individuals with disabilities.
The approach employed ensures compliance with Web Content Accessibility Guidelines(WCAG).
REANA is a reusable and reproducible research data analysis platform that allows researchers to run declarative computational workflows on a remote compute cloud. REANA currently uses native Kubernetes Job API to schedule user workloads on Kubernetes clusters. This study evaluates the features and performance of the Kubernetes native batch scheduling systems Kueue, as a possible alternative for REANA. The findings of this study leverage representative particle physics model analyses to provide valuable insight into the strengths and limitations of the Kueue scheduling system for its future integration into the REANA ecosystem.
Pentaho is a service offered for creating and designing reports, generating analytics and data integration. The current system exists as two instances- one with SSO authentication and another with basic authentication. My project in this internship was to create an authentication plugin which would streamline the process, creating a single entry point for both cases. This was done using the OIDC plugin in combination with the CERN SSO.
Implementing and benchmarking functions already implemented in the CPU for the GPU, which can have speedups of x50. Memory management, shared memory, limitations...
Since the ATLAS Detector is exposed to an intense environment during Run-3 and additionally due to its age, the operation of the detector becomes even more challenging. These challenges introduce difficulties in ensuring high data quality standards. In order to counteract against that, identifying the emerging problems in the Data Acquisition (DAQ) and Detector Control System (DCS) plays a crucial role. Therefore, a Machine Learning based anomaly detection method is employed. This method detects outliers of various time series data coming from the DAQ and DCS, to identify emerging problems before they impact the data quality. This talk will present first results of feasibility studies of using such methods in the ATLAS Pixel Detector as an example use case.