Source-linked AI summary
Deep learning in color: towards automated quark/gluon jet discrimination
Patrick T. Komiske, Eric M. Metodiev, Matthew D. Schwartz
TL;DR
The paper asks whether deep convolutional networks can improve quark/gluon jet discrimination without relying on carefully designed observables, despite differences among simulations. It represents jets as colored images, compares the networks with physically motivated variables, and finds comparable or better performance plus similar behavior across Pythia and Herwig training samples.
Problem
Quark/gluon jet discrimination is difficult, and existing simulations differ substantially, motivating tests of whether learning from minimally processed jet data can match physicist-designed observables.
Method
The study applies deep convolutional neural networks to jet images augmented with charged-particle momentum, neutral-particle momentum, and charged-particle counts.
Results
The deep networks perform as well as or better than physically motivated observables, while network performance is the same whether trained on Pythia or Herwig samples.
Takeaways & Limitations
The networks may extract robust physical information despite imperfect simulations and without substantial input about the physics underlying quark/gluon differences.
Takeaways & Limitations
The study cannot completely refute the possibility that the networks use unphysical simulation features.
Abstract
from arXiv · showhide
Artificial intelligence offers the potential to automate challenging data-processing tasks in collider physics. To establish its prospects, we explore to what extent deep learning with convolutional neural networks can discriminate quark and gluon jets better than observables designed by physicists. Our approach builds upon the paradigm that a jet can be treated as an image, with intensity given by the local calorimeter deposits. We supplement this construction by adding color to the images, with red, green and blue intensities given by the transverse momentum in charged particles, transverse momentum in neutral particles, and pixel-level charged particle counts. Overall, the deep networks match or outperform traditional jet variables. We also find that, while various simulations produce different quark and gluon jets, the neural networks are surprisingly insensitive to these differences, similar to traditional observables. This suggests that the networks can extract robust physical information from imperfect simulations.
1 Introduction
The paper motivates quark/gluon jet discrimination as a difficult, practically important collider-physics task and investigates whether minimally processed data can support effective learning. It also frames simulation reliability and theoretical understanding as central concerns.
- Deep learning may distinguish quark and gluon jets using reasonably raw data rather than carefully constructed observables.
- Jet images treat calorimeter energy deposits as pixel intensities, enabling image-recognition algorithms to process collider data.
- Quark/gluon jet discrimination is challenging because their radiation patterns and non-perturbative effects are difficult to disentangle.
- Traditional discrimination uses complementary shape and count observables, including jet width, mass, energy correlations, charged-particle count, and triggered calorimeter cells.
- Current quark and gluon jet simulations are not completely trustworthy, while detector effects and hadronization are difficult to approach analytically.
- The paper assumes a future in which simulations agree well with data, while noting that theory and simulation methods are improving.
2 Deep Neural Networks
This section introduces neural-network fundamentals and explains why deep convolutional architectures are suited to image-like inputs. It covers layers, activations, supervised training, pooling, and convolutional filters.
- A neural network learns a function from fixed-size inputs to fixed-size outputs through input, hidden, and output layers.
- Non-input units combine weighted inputs and a bias before applying an activation function.
- ReLU(x) = max{0, x} is used in modern computer vision because it is fast and avoids saturation for positive inputs.
- Supervised classification training adjusts weights and biases using labeled examples and a loss function.
- Deep networks can identify low-level features in early layers and more abstract features in later layers.
- Convolutional neural networks apply learned n × n filters across image patches, while max-pooling downsamples feature maps.
3 Event Generation and Network Architecture
The study generates quark and gluon jets with Pythia and Herwig, converts them into preprocessed jet images, and trains convolutional networks. It extends grayscale images with three particle-information channels and examines preprocessing and architecture choices.
- 3 Event Generation: Events were generated with Pythia 8.219 and Herwig 7.0 for 13 TeV proton-proton collisions, using multiple jet-transverse-momentum ranges.
- 3.1 Pre-processing: Jet images are square (η, φ)-space arrays whose pixels contain deposited transverse momentum; R = 0.4 with 33 × 33 pixels gives ∆η = ∆φ = 0.024.
- 3.3 Jet images in color: The approach neglects measurement uncertainty, although neural networks can propagate errors when trained on replicated datasets.
- 3.1 Pre-processing: Preprocessing centers, crops, normalizes, zero-centers, and standardizes the images before training.
- 3.1 Pre-processing: Zero-centering and standardization improved performance by placing input pixels on an equal footing and facilitating feature learning.
- 3.2 Network architecture: The deep architecture uses three convolution–ReLU–maxpooling iterations, followed by a dense ReLU layer and a two-unit softmax output.
- 3.2 Network architecture: Only moderate architecture optimization and minimal hyperparameter tuning were performed, so systematic exploration might improve network performance.
- 3.3 Jet images in color: Color jet images use charged-particle transverse momentum, neutral-particle transverse momentum, and local charged-particle multiplicity as separate channels.
4 Fisher jets and a look inside the networks
The paper connects Fisher’s linear discriminant and shallow dense networks to image-based jet observables, then examines the features these models learn. The learned observables emphasize both broad jet geometry and transverse momentum near the jet core.
- Fisher jets: A single-unit network with weighted image inputs is equivalent to Fisher’s Linear Discriminant, whose weight image acts as a filter.The Fisher jet image is trained to maximize discrimination between quark and gluon samples.
- Fisher jets: After zero-centering and standardization, no regularization is needed for a sensible Fisher jet, while a log transformation with r′ = 10^-3 significantly improves FLD performance.Without these preprocessing steps, regularization is needed to limit overfitting and sensitivity to noisy outer image regions.
- Shallow dense network: The shallow dense network generalizes the FLD by using 16 fully connected ReLU units to learn additional discriminating features.It uses L2 regularization, dropout, and the log-transformed inputs.
- Fisher jets: The Fisher jet visualization compares its learned weight pattern with standardized quark- and gluon-jet images after projection.The figure uses 200 GeV Pythia jets and includes the additional log-transformation preprocessing step.
- Learned observables: The learned shallow-network observables divide into features sensitive to large-scale jet geometry and features sensitive to transverse momentum near the jet core.The geometric features include moments, annular structure, and left/right asymmetries.
5 Network performance
Deep convolutional networks applied to jet images match or outperform physics-motivated observables and their boosted decision-tree combination. Adding color helps most at higher jet transverse momenta, while performance remains comparatively robust across image sizes and event generators.
- N95 is the strongest physics-motivated observable in several cases, but deep convolutional networks notably outperform it.N95 measures the minimum number of pixels containing 95% of the jet transverse momentum.
- Deep convolutional networks outperform traditional jet variables and match or exceed a BDT combining all five observables.The five observables are girth, charged particle multiplicity, two-point moment, xmax, and N95.
- 5.2 Merge layers: Adding CPM modestly improves performance at 500 GeV and 1000 GeV, whereas adding N95 yields little or no improvement.This indicates that the network learns geometric observables containing information similar to N95.
- 5.4 Event generator dependence: Networks and baseline jet variables perform worse on Herwig than on Pythia, but a Pythia-trained network has comparable Herwig performance to one trained directly on Herwig.The generators disagree primarily on gluon distributions, and the network is surprisingly insensitive to the generator when transferred between samples.
6 Conclusions
Deep convolutional networks address the difficult problem of quark/gluon jet discrimination using minimally processed jet-image data and can match or exceed physically motivated observables. Colored inputs capture charge and multiplicity information, while generator-insensitive performance suggests the networks learn distinctions shared across simulations, although simulation reliability still limits immediate application to data.
- Motivation: Quark/gluon discrimination is practically important at the LHC because many signals are quark-rich while backgrounds are gluon-dominated.The task is difficult because radiation-pattern correlations and hadronization effects are hard to disentangle.
- Performance: Deep convolutional networks using essentially no hand-crafted physics input perform as well as or better than the best physically motivated observables.The comparison is reported against a collection of observables from earlier studies.
- Colored inputs: Colored jet images encode charged-particle transverse momentum, neutral-particle transverse momentum, and local charged-particle multiplicity as separate channels.Adding these channels gives marginal improvement at 200 GeV but substantial improvement for very high-pT jets, where multiplicity is higher.
- Generator dependence: Network performance is surprisingly insensitive to the event generator: Pythia-trained and Herwig-trained networks perform similarly across the generator samples.This suggests the networks identify underlying physical distinctions shared by the simulations, despite poorer labeling performance on Herwig than on Pythia.
- Implications and limitations: The networks may be more trustworthy on data than the imperfect simulations used for training, but the authors do not recommend applying them to data yet.The remaining concern is that current generator simulations cannot yet be trusted sufficiently.
- Interpretability: Convolutional filters and merge layers provide ways to inspect learned radiation features and test whether particular observables have been learned.The authors note that some lack of low-level understanding is unavoidable for multivariate methods, including boosted decision trees.