Jet Energy Corrections with GNN Regression using Kubeflow @ CERN
summary
Graph neural network and deep‑set regressors are trained on a 10 GB S3 dataset via Kubeflow pipelines with Katib random search for hyperparameters, multi‑GPU distributed training and TensorBoard monitoring, exported to ONNX and served with NVIDIA Triton through KServe. Results show 10% resolution improvement over the non‑ML baseline and a factor of three reduction in flavor dependence. The production pipeline, hyperparameter trials and serving endpoint are reproducible; only minor issues arose around the Triton image version and TensorBoard S3 secret handling.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:22About CERN
- 0:54Large Hadron Collider
- 2:35Machine Learning Applications
- 4:51Challenges
- 6:53Jet Energy Corrections
- 8:05Particle Cloud
- 11:03Kubeflow
- 11:34Training
- 12:46Inference
- 13:28Live Demo
- 18:29Results
- 21:38Conclusion
related talks
29mA Scalable Platform for Training and Inference Using Kubeflow
32mFrom Edge To Insight: Cybersecurity Intelligence With Kubeflow...
51mMachine Learning using Kubeflow and Kubernetes
21mIntroducing Kubeflow (w. Special Guests Tensorflow and Apache Spark)
24mA Better and More Efficient ML Experience for CERN Users
31mBuilding and Managing a Centralized ML Platform with Kubeflow at...
37mTraining and Optimisation of Large Transformer Models: An ATLAS and CERN Use Case - Ricardo Rocha, CERN & Maxence Draguet, University of Oxford
42mDeep learning beyond the learning