QCD-Aware Recursive Neural Networks for Jet Physics

Gilles Louppe
Kyunghyun Cho
Cyril Becot
Kyle Cranmer

Recent progress in applying machine learning for jet physics has been built upon an analogy between calorimeters and images. In this work, we present a novel class of recursive neural networks built instead upon an analogy between QCD and natural languages. In the analogy, four-momenta are like words and the clustering history of sequential recombination jet algorithms is like the parsing of a sentence. Our approach works directly with the four-momenta of a variable-length set of particles, and the jet-based tree structure varies on an event-by-event basis. Our experiments highlight the flexibility of our method for building task-specific jet embeddings and show that recursive architectures are significantly more accurate and data efficient than previous image-based networks. We extend the analogy from individual jets (sentences) to full events (paragraphs), and show for the first time an event-level classifier operating on all the stable particles produced in an LHC event.

Please cite using the following BibTex entry:

@article{louppe2017qcdrecnn,
           author = {{Louppe}, G. and {Cho}, K. and {Becot}, C and {Cranmer}, K.},
            title = "{QCD-Aware Recursive Neural Networks for Jet Physics}",
          journal = {ArXiv e-prints},
    archivePrefix = "arXiv",
           eprint = {1702.00748},
     primaryClass = "hep-th",
             year = 2017,
            month = feb,
}

Instructions

Requirements

python 2.7
autograd
scikit-learn
click

Data

(mirror to be released)

Usage

Classification of W vs QCD jets:

# Training
python train.py data/w-vs-qcd/final/antikt-kt-train.pickle model.pickle

# Test
# see notebooks/04-jet-study.ipynb

Classification of full events:

# Training
python train_event.py data/events/antikt-kt-train.pickle model.pickle

# Test
python test_event.py data/events/antikt-kt-train.pickle data/events/antikt-kt-test.pickle model.pickle 100000 100000 predictions.pickle
# see also notebooks/04-event-study.ipynb

Rebuilding the data

Additional requirements:

fastjet
cython
rootpy
h5py

Building W vs QCD jet data:

Load data from h5 files. Execute cells 1-3 from notebooks/01-load-w-vs-qcd.ipynb. Adapt cell 3 as necessary.
Split the data into train and test. Execute cells 1-3 from notebooks/02-split-train-test.ipynb. Adapt cell 3 as necessary.
Preprocess the data. Execute cells 1-6 from notebooks/03-preprocessing.ipynb. Adapt cell 6 as necessary. Run one of the following cells depending on the desired topology.
(Optional) Repack the train and test using the last cell of notebooks/03-preprocessing.ipynb. This is necessary for historical reasons to exactly reproduce the paper data.

Building full event data:

Load data from pickle files. Execute cells 1-2, 4-7. Adapt cell 7 as necessary.
Split the data into train and test. Execute cells 1-2 and the last of notebooks/02-split-train-test.ipynb.

ym1495 / recnn