Machine Learning and Deep Learning

Knowhow and experience derived from early adoption of neural network techniques by particle physics community.

CERN's Know-How:

  • Particle physicists were among the first to use machine learning (ML) in software for analysis & simulations
  • First AI HENP seminar in 1990
  • Already in 2010, the CMS and LHCb experiments successfully introduced machine learning algorithms to its trigger system
  • Higgs boson discovery earlier than expected (2012), also with help of ML

Facts & Figures:

  • <10 μsec: ML applied for extremely fast decision making in CERN detector trigger systems
  • AUC> 90%: High true positive / low false positive rates achieved even in sparse images with little datapoints
  • >1000 times faster: Convolutional neural networks have dramatically decreased computing time in physics (vs traditional computing)
  • ~100% efficient: Highly reliable trace reconstruction algorithms using tailored neural network techniques, even with multiple tracks in one sensor

Value Proposition:

Read more about Machine Learning and Deep Learning here

Key Competences

Designing & Training Neural Networks:

CERN has a long history in the design and training of neural networks in for example classification, filtering, event and particle detection, regression, clustering and anomaly detection. Most of the ML/DL codes are tailor made using C++, Phyton, TensorFlow and Keras and applied in software or hardware (FPGAs).

Fast neural network inference in FPGAs:

CERN needs ultra fast machine learning interference (execution in μsec), requiring compact code for FPGAs. A companion compiler package for this work is developed based on High-Level Synthesis (HLS) called hls4ml to build machine learning models in FPGAs, allowing for fast prototyping and shorter time to results.

Key Applications

TMVA as Open Source ML / DL Toolkit:

The open source Toolkit for Multivariate Data Analysis (TMVA) developed by CERN provides a machine learning environment for the processing and evaluation of multivariate classification, both binary and multi class, and regression techniques. It is integrated in ROOT, a modular scientific software toolkit (C++).