Machine Learning for Jet Physics

America/Chicago
One West (WH1W) (Fermilab)

One West (WH1W)

Fermilab

Ben Nachman (University of California Berkeley (US)) , Kyle Stuart Cranmer (New York University (US)) , Pushpalatha Bhat (Fermi National Accelerator Lab. (US)) , Sergei Gleyzer (University of Florida (US)) , Tilman Plehn (Heidelberg University)
Description

Machine learning has become a hot topic in particle physics over the past several years. In particular, there has been a lot of progress in the area of particle and event identification, reconstruction, fast simulation and others. One significant area of research and development has focused on jet physics. In this workshop, we will discuss current progress in this area, focusing on new breakthrough ideas and existing challenges. The ML4Jets workshop will be open to the full community and will include LHC experiments as well as theorists and phenomenologists interested in this topic.  

For last year's workshop, please see https://indico.physics.lbl.gov/indico/event/546/.

Chairs:
Sergei Gleyzer (U Florida)
Ben Nachman (LBNL)


Organizing Committee:
Pushpa Bhat (Fermilab) 
Kyle Cranmer (NYU) 
Sergei Gleyzer (U Florida) 
Ben Nachman (LBNL) 
Tilman Plehn (Heidelberg)

Local Organizing Committee:
Gabriele Benelli (Brown U),
Javier Duarte (Fermilab)
Benjamin Kreis (Fermilab)
Nhan Tran (Fermilab)
Justin Pilot (UC Davis)

LPC Event Committee Chairs:
Gabriele Benelli (Brown U)
Benjamin Kreis (Fermilab)

LPC Coordinators:
Cecilia Gerber (UIC)
Sergo Jindariani (Fermilab)

Participants
  • Abhijith Gandrakota
  • Abhishek Das
  • Alejandro Gomez Espinosa
  • Alexx Perloff
  • Anastasia Karavdina
  • Anders Andreassen
  • Annapaola De Cosa
  • Aran Garcia-Bellido
  • Arjun Chhetri
  • Aviv Ruben Cukierman
  • Ben Kreis
  • Ben Nachman
  • Benjamin Tannenwald
  • Bryan Ostdiek
  • Caleb James Smith
  • Christine Angela Mc Lean
  • Christopher Madrid
  • Cilicia Uzziel Perez
  • Colin Fallon
  • Cristina Ana Mantilla Suarez
  • Daniel Arthur Faia
  • Daniel Marley
  • David Miller
  • David Shih
  • David Yu
  • Dipsikha Debnath
  • Dorival Gonçalves
  • Douglas Berry
  • Elliot Parrish
  • Emanuele Usai
  • Emil Sorensen Bols
  • Emily Ann Smith
  • Enrico Bothmann
  • Eric Metodiev
  • Florian Bury
  • Gabriel Perdue
  • Gabriele Benelli
  • Giorgia Rauco
  • Grace Haza
  • Gregor Kasieczka
  • Haifa Rejeb Sfar
  • Hannsjorg Weber
  • Hevjin Yarar
  • Hossein Afsharnia
  • Huilin Qu
  • Ian Gustafson
  • Imre Kondor
  • Isabel Raymundo
  • Ivan Pogrebnyak
  • Jack Collins
  • Jacob Julian Kempster
  • James Proudfoot
  • Javier Mauricio Duarte
  • Jennifer Ngadiuba
  • Jennifer Thompson
  • Jesse Thaler
  • Joey Huston
  • Joey Huston
  • John Alison
  • Jonathan Shlomi
  • Jordan Damgov
  • Joseph Earl Lambert
  • Julie Hogan
  • Justin Pilot
  • Justin Tan
  • Ka Hei Martin Kwok
  • Kevin Pedro
  • Konstantin Gizdov
  • Leonardo Giannini
  • Lily Asquith
  • Loukas Gouskos
  • Lucas Kang
  • Marat Freytsis
  • Marc Gabriel Weinberg
  • Marco Farina
  • Marguerite Belt Tonjes
  • Mark Neubauer
  • Matthew Dolan
  • Matthew Feickert
  • Mauro Verzetti
  • Melissa Hutcheson
  • Miaoyuan Liu
  • Michael Andrews
  • Miguel Garcia
  • Nabin Poudyal
  • Nadja Strobbe
  • Nathaniel Joseph Pastika
  • Nicholas Elsey
  • Patrick Komiske
  • Paulina Kulyavtsev
  • Pradeep Jasal
  • Pushpalatha Bhat
  • Raghav Kunnawalkam Elayavalli
  • Rishi Gautam Patel
  • Robin Erbacher
  • Sadia Khalil
  • Salvatore Rappoccio
  • Samuel Ross Meehan
  • Saptaparna Bhattacharya
  • Sebastian Macaluso
  • Sergei Gleyzer
  • Sezen Sekmen
  • Shih-kai Lin
  • Simi Ily
  • Sitong An
  • Souvik Das
  • Stephan Lammel
  • Steve Mrenna
  • Sudeshna Banerjee
  • Sudhir Malik
  • Suneel Dutt
  • Sung Hak Lim
  • Tae Kim
  • Tao Liu
  • Taoli Cheng
  • Taylor Childers
  • Thea Aarrestad
  • Tilman Plehn
  • Titas Roy
  • Tommaso Dorigo
  • Tongguang Cheng
  • Victor Daniel Elvira
  • Walter Hopkins
  • Weinan Si
  • Yang-Ting Chien
  • Yannik Alexander Rath
  • Yao Yao
  • Yuichiro Nakai
  • Zhenbin Wu
    • 1
      Registration
    • Introduction (Chairs: Sergei Gleyzer and Benjamin Nachman)
      • 2
        Welcome Remarks
        Speaker: Dr Joseph Lykken (Deputy Director, Fermilab)
      • 3
        Logistics
      • 4
        Jets and ML in Theory (30'+15')
        Speaker: Jesse Thaler (MIT)
      • 5
        Jets and ML in ATLAS (30'+15')
        Speaker: Walter Hopkins (Argonne National Laboratory (US))
      • 6
        Jets and ML in CMS (30'+15')
        Speaker: Mauro Verzetti (CERN)
    • 11:30 AM
      Lunch
    • Jet Tagging (Chairs: Gregor Kasieczka and Matt Dolan)
      • 7
        Introduction and overview (20'+10')
        Speaker: Gregor Kasieczka (Hamburg University (DE))
      • 8
        top and W tagging with ATLAS (20’+5’)
        Speaker: Samuel Ross Meehan (University of Washington (US))
      • 9
        top and W tagging with CMS (20’+5’)
        Speaker: Justin Pilot (University of California Davis (US))
      • 2:00 PM
        Short Break
      • 10
        b-tagging in ATLAS (25’+5’)
        Speaker: Matthew Feickert (Southern Methodist University (US))
      • 11
        ML Techniques for heavy flavour identification in CMS (25'+5')

        Jet flavour identification is a fundamental component for the physics program of the LHC-based experiments. The presence of multiple flavours to be identified leads to a multiclass classification problem. In this presentation we will present the performance on both simulated and real data of our latest resolved heavy flavour taggers, DeepCSV and DeepFlavour as well as the future prospects for the evolution of these techniques and the technical strategies adopted to deploy them in the harsh computing environment of a large-scale HEP computing software stack.

        Speaker: Emil Sorensen Bols (Vrije Universiteit Brussel (BE))
      • 12
        Deep Learning Strange Jets (20’+5’)

        By applying deep learning techniques, we explore the possibility of strange-quark tagging, which is the last missing piece among quark and gluon identifications in jets. The main difficulty here is of distinguishing strange-quark jets from down-quark jets. However, strange-quark jets are likely to contain more Kaons carrying large fractions of the jet $p_T$ than down-quark jets. A strategy for strange-quark tagging is then to concentrate on neutral Kaons, $K_L$ and $K_S$, which are expected to be discriminated from other hadrons as the $K_L$ and long-lived $K_S$ drop their energies only to the Hadron Calorimeter while other hadrons leave some trace in the tracker or the Electromagnetic Calorimeter. We create the pixel images of strange and down-quark jets with colors of the track $p_T$, hadronic energy and electromagnetic energy. The images are fed into Convolutional Neural Networks (CNNs). We find that the CNN tagger outperforms the best cut-based tagger we define.

        Speaker: Yuichiro Nakai (Rutgers University)
      • 3:45 PM
        Pre-Colloquium Coffee Break
    • 13
      Fermilab Colloquium (Graph Networks and Physics Applications)
      Speaker: Risi Condor (University of Chicago)
    • 8:30 AM
      Coffee
    • Representing Jets (Chairs: Mauro Verzetti and David Shih)
      • 14
        Introduction and overview (20’+10’)
        Speaker: David Shih (Rutgers University)
      • 15
        Energy Flow Networks: Deep Sets for Particle Jets (20’+5’)

        Collider events are naturally described as sets of particles which have variable size and are inherently permutation symmetric. Machine learning architectures operating on collider events should ideally be able to handle variably sized inputs and be manifestly symmetric with respect to particle ordering. Building off of the recently developed Deep Sets paradigm, which is designed for learning from point clouds, I will introduce Energy Flow Networks (EFNs) and their more general counterparts Particle Flow Networks (PFNs), which explicitly have these desired properties. Using the task of discriminating different kinds of jets as an example, I will demonstrate how the EFNs and PFNs have excellent classification performance and allow for fascinating visual interpretations of what the model is learning.

        Speaker: Patrick Komiske (Massachusetts Institute of Technology)
      • 16
        Jet as a particle cloud (20’+5’)

        How to represent a jet is one of the key aspects of machine learning algorithms for jet physics. Motivated by recent progress in machine learning community on point cloud recognition, we propose a new approach that represents a jet as an unordered set of particles with their measured properties, effectively a "particle" cloud. Specialized algorithms for point cloud recognition, e.g., Dynamic Graph CNN, are explored, and the performance on jet classification is compared to alternative approaches using state-of-the-art 2D CNN models on jet images and 1D CNN models on jet constituent particles.

        Speaker: Huilin Qu (Univ. of California Santa Barbara (US))
      • 17
        Spectral Analysis of Color Charge in Two-Prong Jets with Neural Networks (20’+5’)

        We discuss signatures in the two-point correlation spectrum $S_{2}(R)$ on the angular scale $R$ for identifying color charge in two-prong jets. In a two-prong jet, the radiation pattern is correlated with the color charge of originating partons and the decay topology of the jet so that we need a strategy considering those effects simultaneously. The spectral analyses with $S_{2}(R)$ and neural network provide us with a visual framework for studying two-prong substructure as well as color superstructure in terms of the angular scale $R$. Furthermore, we can design neural networks with interpretable weights in this framework. The interpretable weights help us understand how the prediction from the neural network came out. We show our results in the context of classification among Higgs, Sgluon, and QCD jets.

        Speaker: Sung Hak Lim (KEK)
      • 18
        ML@QCD efforts in Sherpa: shower variations and phase-space sampling (20’+5’)

        QCD calculations that resum soft-collinear logarithms by a parton-shower algorithm can not currently be used in PDF fits. This is due to the high computational cost of generating Monte-Carlo events for each variation of the PDFs, and reduces the number of data points available for the fits. We propose an approximation based on training a NN to predict the effect of varying the shower input parameters and present proof-of-principle results for strong-coupling variations.

        Another challenge in QCD Monte-Carlo simulations is effective phase-space sampling. The efficiency of generating unweighted events can easily drop to $10^{-4}$ and less for some state-of-the-art calculations. We present results for using ML to improve the generation of phase-space points both globally and in the local Markov-Chain steps of the $(\text{MC})^3$ method.

        Speaker: Enrico Bothmann (University of Edinburgh)
    • 11:30 AM
      Lunch
    • Representing Jets (Chairs: Mauro Verzetti and David Shih)
      • 19
        Quarks vs. Gluons for Higgs->invisible searches (20'+5')

        Quark-gluon discrimination could greatly improve the sensitivity of a
        number of analyses at the LHC, and as such has received a significant
        amount of investigation. Because the differences between quark and
        gluon jets are largely contained in the jet substructure and are often
        very subtle, this problem lends itself to machine learning techniques.
        We explore this question in the LoLa framework, and demonstrate that
        we see good discrimination for pure quark and gluon jets, both at
        particle level and after including a fast detector simulation. Next,
        we apply our network to a physics problem, a monojet Higgs -> invisible
        signal (gluon dominated) with a Z+monojet background (quark dominated).
        We investigate how this differs from the pure quark-gluon case and how
        the jet transverse momentum affects the network performance.

        Speaker: Jennifer Thompson (ITP Heidelberg)
      • 20
        Top tagging with Lorentz Boost Networks and simulation of electromagnetic showers with a Wasserstein GAN (20'+5')

        In this talk, we present two applications of deep learning in the areas of top quark identification and electromagnetic shower generation.
        As deep learning methods are adopted for high energy physics, increasing attention is given to the development of dedicated architectures incorporating physical knowledge. We introduce a model that utilizes our knowledge of particle combinations and directly integrates Lorentz boosting, and apply this model to separate hadronic top-quark decays from light quark and gluon jets. We also investigate the trained combinations and boosts to gain insights into what the network is learning.
        Generative models have recently been applied to physics simulations, in particular for calorimeter showers. They promise a speed-up of several orders of magnitude compared to full simulations. We present results on the generation of electromagnetic showers in a multi-layer calorimeter using a Wasserstein Generative Adversarial Network (WGAN), emphasizing on the comparison to a traditional simulation using GEANT4. Initial conditions of the simulation are incorporated through a dedicated architecture based on constrainer networks.

        [M. Rieger, D. Schmidt, Winners presentation at the IML Machine Learning
        Working Group: sequential models, CERN, Geneva, Jun. 2018, https://indico.cern.ch/event/722319/contributions/ ]
        [M. Erdmann, J. Glombitza, T. Quast, arXiv:1807.01954]

        Speaker: Yannik Alexander Rath (RWTH Aachen University (DE))
    • Experimental/Practical aspects of learning with jets (Chairs: Ben Hooberman and Daniel Elvira)
      • 21
        Introduction and overview (20'+10')
      • 22
        Machine Learning for Jet Calibration in ATLAS (20'+5')
        Speaker: Aviv Ruben Cukierman (SLAC National Accelerator Laboratory (US))
      • 23
        End-to-end jet identification for quark/gluon discrimination using CMS Open Data (20'+5')

        Jet identification is a very active area of applied machine learning research in particle physics, benefitting from a wide array of ideas and algorithms. Among these is the idea of building jet “images". However, many image-based implementations have struggled to compete with the current state-of-the-art classifiers that are dominated by specialized networks that rely on higher-level inputs. In this talk, we describe the application of the end-to-end approach (https://arxiv.org/abs/1807.11916) developed for particle and event identification at CMS to the challenge of identifying jets.

        We describe end-to-end jet reconstruction using the full granularity of the CMS detector. CMS Open Data samples are used that take advantage of the full Geant4-based detector simulation. Using quark vs. gluon classification as a reference, we demonstrate competitive performance with existing state-of-the-art jet identification algorithms. Furthermore, we offer insights into the role of various sub-detectors for jet identification, and describe how end-to-end techniques can be useful for event-level classification for events containing jets.

        Speaker: Michael Andrews (Carnegie-Mellon University (US))
      • 3:30 PM
        Coffee break an group photo
      • 24
        Pulling Out All the Tops with Computer Vision and Deep Learning (20'+5')

        We apply computer vision with deep learning -- in the form of a convolutional neural network (CNN) -- to build a highly effective boosted top tagger. Previous work (the ``DeepTop" tagger of Kasieczka et al) has shown that a CNN-based top tagger can achieve comparable performance to state-of-the-art conventional top taggers based on high-level inputs. Here, we introduce a number of improvements to the DeepTop tagger, including architecture, training, image preprocessing, sample size and color pixels. Our final CNN top tagger outperforms BDTs based on high-level inputs by a factor of $\sim 2$--3 or more in background rejection, over a wide range of tagging efficiencies and fiducial jet selections.
        As reference points, we achieve a QCD background rejection factor of 500 (60) at 50% top tagging efficiency for fully-merged (non-merged) top jets with $p_T$ in the 800--900 GeV (350--450 GeV) range.
        Our CNN can also be straightforwardly extended to the classification of other types of jets, and the lessons learned here may be useful to others designing their own deep NNs for LHC applications.

        Speaker: Sebastian Macaluso
      • 25
        Fast inference of jet substructure classifiers with FPGAs (20'+5')

        Recent results at the Large Hadron Collider (LHC) have pointed to enhanced physics capabilities through the improvement of the real-time event processing techniques. Machine learning methods are ubiquitous and have proven to be very powerful in LHC physics, and particle physics as a whole. However, exploration of the use of such techniques in low-latency, low-power FPGA hardware has only just begun. FPGA-based trigger and data acquisition (DAQ) systems have extremely low, sub-microsecond latency requirements that are unique to particle physics. We present a case study for neural network inference in FPGAs focusing on a classifier for jet substructure which would enable, among many other physics scenarios, searches for new dark sector particles and novel measurements of the Higgs boson. While we focus on a specific example, the lessons are far-reaching. We develop a package based on High-Level Synthesis (HLS) called hls4ml to build machine learning models in FPGAs. The use of HLS increases accessibility across a broad user community and allows for a drastic decrease in firmware development time. We map out FPGA resource usage and latency versus neural network hyperparameters to identify the problems in particle physics that would benefit from performing neural network inference with FPGAs. For our example jet substructure model, we fit well within the available resources of modern FPGAs with a latency on the scale of 100 ns.

        Speaker: Zhenbin Wu (University of Illinois at Chicago (US))
      • 26
        Modern jet machine-learning classification in real time (20'+5')

        We demonstrate the ability to create drone from a wide range of classifiers, with a particular emphasis on the application to modern jet classification. Machine learning is increasingly dominating the preferred tool for the classification of jets. However, as experiment data rates increase by orders of magnitude, such technologies become expensive in terms of time and performance. In light of this, we present a method and toolkit for creating a drone classifier from any machine learning classifier, preserving the accuracy, precision, and specificity, but greatly improving on algorithm execution performance.

        Speaker: Konstantin Gizdov (The University of Edinburgh (GB))
      • 27
        Heavy flavour identification for boosted resonances and large cone jets in CMS (20'+5')

        Physics with boosted objects has been an increasingly interesting topic in the last years. Modern machine learning techniques, and Deep Learning in particular, have changed the landscape providing new taggers with significant performance boost. The presentation will focus on the taggers for the boosted regime, DeepAK8, DeepDoubleB, and DeepDoubleC and the strategies to measure their performance in real data.

        Speaker: Javier Mauricio Duarte (Fermi National Accelerator Lab. (US))
    • 28
      Dinner (Masala)
    • 29
      History of Machine Learning for High Energy Physics (20'+10')
      Speaker: Pushpalatha Bhat (Fermi National Accelerator Lab. (US))
    • 9:30 AM
      Coffee break
    • Simulation Independent Methods (Chairs: Tommaso Dorigo and Bryan Ostdiek)
      • 30
        Introduction and overview (20'+10')
        Speaker: Bryan Ostdiek (University of Oregon)
      • 31
        Disentangling Jet Categories at Colliders (20'+5')

        The “jet topics” framework identifies (or defines) underlying classes of jets directly from data with little to no input from simulation or theory. Due to a mathematical connection between mixed samples of jets and emergent themes in documents, methods from topic modeling and blind source separation can be used to extract jet topics from data. Any machine-learned jet tagger, treated as a likelihood-ratio approximator, can be directly applied as a jet topic extractor. I apply the jet topics method to extract quark and gluon distributions and fractions from simulated Z+jet and dijet samples, and I discuss the potential for fully data-driven training and calibration of jet taggers.

        Speaker: Eric Metodiev (Massachusetts Institute of Technology)
      • 32
        JUNIPR: a Framework for Unsupervised Machine Learning in Particle Physics (20'+5')

        In applications of machine learning to particle physics, a persistent challenge is how to go beyond discrimination to learn about the underlying physics.
        In this talk, we will present a new framework: JUNIPR, Jets from UNsupervised Interpretable PRobabilistic models, which uses unsupervised learning to learn the intricate high-dimensional contours of the data upon which it is trained, without reference to pre-established labels.
        In order to approach such a complex task, JUNIPR is structured intelligently around a leading-order model of the physics underlying the data.
        In addition to making unsupervised learning tractable, this design actually alleviates existing tensions between performance and interpretability.
        Applications to discrimination, data-driven Monte Carlo generation and reweighting of events will be discussed.

        Speaker: Anders Andreassen (UC Berkeley)
    • 11:30 AM
      Lunch
    • Simulation Independent Methods (Chairs: Tommaso Dorigo and Bryan Ostdiek)
      • 33
        CWoLa Hunting: Enhancing the Bump Hunt with Machine Learning (20'+5')

        Machine learning (ML) has rapidly become a core tool for LHC physics, due to the great volume and complexity of the data that this machine collects. Given that it is not a-priori known what form new physics (if any) might take, there has been a surge of interest in the past year in approaches that would enable an ML algorithm to look for new physics directly in the LHC data without reference to any simulated signal sample. This talk will focus on a concrete example called 'CWoLa Hunting' (Classification Without Labels), in which it assumed that the signal is localized in some window in one variable (e.g. a resonance in an invariant mass) in which the background is smooth, but no additional assumptions are made about the morphology of the signal in some orthogonal set of 'auxiliary' variables. The ML algorithm searches for an unusual population of events in the signal window using these auxiliary variables. I will use as a case study a dijet resonance search in which a resonant signal might form a bump in the dijet invariant mass distribution, while the ML algorithm searches for a localized population of events with unusual jet substructure.

        Speaker: Jack Collins (University of Maryland and Johns Hopkins University)
      • 34
        QCD or What: Deep autoencoder based searches for new physics (20'+5')

        In the current era of high energy particle collider experiments, we are faced with an overwhelming amount of data and the limiting uncertainty in new physics searches can often come from theory and not experiment. In our efforts to develop new approaches to extract complex signals from large backgrounds, BDTs, neural networks and other machine learning techniques are becoming increasingly significant. These tools allow us to find patterns in data that would be impossible to identify with a simple cut-and-count approach. In this work we show how unsupervised learning approaches based on deep autoencoders can be directly trained on data and used for model-independent searches for new physics. Beyond autoencoder we will discuss progress in applying and understanding the four-vector based Lorentz-layer approach to new challenges.

        Speaker: Gregor Kasieczka (Hamburg University (DE))
      • 35
        Searching for New Physics with Autoencoders (20'+5')

        Autoencoders as tools for new physics discovery. The key idea of the autoencoder is that it learns to map background events back to themselves, but fails to reconstruct anomalous events that it has never encountered before. The reconstruction error can then be used as an anomaly threshold. An illustrative example of background QCD jets versus tops will be discussed.

        Speaker: Marco Farina (Rutgers University)
      • 36
        Novelty Detection Meets Collider Physics (20'+5')

        Novelty detection is the machine learning task to recognize data belonging to an unknown pattern. Complementary to supervised learning, it allows to analyze data without a priori knowledge on signal or model-independently. In this talk, we would demonstrate the potential role of novelty detection in collider physics, using autoencoder-based deep neural network. Explicitly, we develop a set of density-based novelty evaluators, which are sensitive to the clustering of unknown-pattern data/signal events, for optimizing detection algorithms. We also study the generic influence of the known-pattern data fluctuations on detection sensitivity which arise from non-signal regions in the feature space (Look Elsewhere Effect). Strategies to address it are proposed. For proof of concept, the algorithms are applied to detecting signal/novel events which are defined by fermionic di-top partner and resonant di-top productions at LHC, and by exotic Higgs decays of two specific modes at future e+e- collider, respectively. With parton-level analysis, we show that the signal of new-physics benchmarks could be recognized with high efficiency.​

        Speaker: Tao Liu (The Hong Kong University of Science and Technology (HK))
    • 37
      Discussion and Closeout
    • 38
      Wine and Cheese Seminar (Deep Learning for HEP)