Source-linked AI summary
JEDI-net: a jet identification algorithm based on interaction networks
Eric A. Moreno, Olmo Cerri, Javier M. Duarte, Harvey B. Newman, Thong Q. Nguyen, Avikar Periwal, Maurizio Pierini, Aidana Serikova, Maria Spiropulu, Jean-Roch Vlimant
TL;DR
The paper asks whether interaction networks can identify high-momentum heavy-particle decays while avoiding representation assumptions used by other jet taggers. JEDI-net models jets as fully connected graphs of particle constituents and learns pairwise interaction representations. Across the studied jet categories, it outperforms the compared dense, convolutional, and recurrent models, while retaining practical representation advantages but remaining costly for real-time deployment.
Problem
Jet tagging must distinguish ordinary quark and gluon jets from compact, multi-pronged jets produced by high-momentum heavy-particle decays.
Method
JEDI-net represents each jet as a fully connected graph of particle constituents and learns a representation from one-to-one particle interactions for classification.
Results
JEDI-net achieves better performance than dense, convolutional, and recurrent neural networks trained and optimized on the same data set.
Takeaways & Limitations
JEDI-net directly processes constituent features without constituent ordering or detector-geometry assumptions, making graph-network deployment an interesting prospect for future LHC runs.
Takeaways & Limitations
The current implementation is costly for online LHC selection because of its large computational resource demand and inference time.
Abstract
from arXiv · showhide
We investigate the performance of a jet identification algorithm based on interaction networks (JEDI-net) to identify all-hadronic decays of high-momentum heavy particles produced at the LHC and distinguish them from ordinary jets originating from the hadronization of quarks and gluons. The jet dynamics are described as a set of one-to-one interactions between the jet constituents. Based on a representation learned from these interactions, the jet is associated to one of the considered categories. Unlike other architectures, the JEDI-net models achieve their performance without special handling of the sparse input jet representation, extensive pre-processing, particle ordering, or specific assumptions regarding the underlying detector geometry. The presented models give better results with less model parameters, offering interesting prospects for LHC applications.
1 Introduction
High-momentum heavy-particle decays create multi-pronged jets that differ from ordinary quark and gluon jets. JEDI-net applies interaction networks to learn from pairwise constituent relations without relying on particle ordering or detector-specific representations.
- High-momentum boson decays form single jets with two prongs, while top decays form jets with three prongs.
- Jet tagging distinguishes ordinary quark and gluon cascades from distinctive heavy-particle momentum flows and large invariant masses.
- JEDI-net represents jet constituents as vertices in a fully connected graph and learns one-to-one particle interaction representations for classification.
- The interaction-network formulation is studied as an alternative that avoids assumptions about constituent ordering and detector geometry.
- Existing deep-learning taggers use high-level features, ordered particle lists, or rectangular jet images, requiring additional representation choices.
2 Related work
Related jet-tagging work has applied dense, convolutional, recurrent, recursive, and graph-based architectures to particle or jet representations. Graph approaches address the sparsity of image-based jet representations and have shown strong categorization performance.
- Deep-learning jet taggers have used DNNs, CNNs, recurrent layers, and recursive layers with physics-motivated quantities or reconstructed particle momenta.
- Graph neural networks have been explored for jet tagging as a way to circumvent sparse image-based jet representations.
- Graph-network methods have also been applied to event topology classification, particle tracking, pileup subtraction, and particle reconstruction.
3 Data set description
The study uses simulated approximately 1 TeV jets from five categories in 13 TeV proton-proton collisions. It compares high-level features, pixelated images, and constituent-particle representations for alternative classifiers and JEDI-net.
- The data set contains simulated jets from light quarks, gluons, W bosons, Z bosons, and top quarks at pT ≈1 TeV in √s = 13 TeV proton-proton collisions.
- Jets are clustered from reconstructed particles with the anti-kT algorithm using jet-size parameter R = 0.8.
- The DNN receives 16 high-level features, while the CNN receives 100×100 images covering Δη = Δφ = 2R and filled with scalar pT sums.
- Each particle representation contains 16 momentum-derived features, including momentum components, energy, pT, angular coordinates, distances, and relative quantities.
- The RNN and JEDI-net use the 150 highest-pT particles, with zero padding when fewer than 150 particles are reconstructed.
- The graph example illustrates three fully connected vertices connected by six directional edges.
4 JEDI-net
JEDI-net represents a jet as a fully connected graph of particle constituents and learns interaction-based features for jet classification. Its architecture processes sender–receiver particle features through trainable functions, while alternative models use dense, image-based, or recurrent representations.
- Graph representation: JEDI-net applies an interaction-network architecture to learn a graph representation of jet constituents for tagging.The input graph contains particles as vertices, and the learned representation is used for classification.
- Graph representation: Each jet graph fully connects its NO particle vertices through directional edges, producing NE = NO × (NO − 1) particle-to-particle interactions.The labeling convention does not affect the network’s representation.
- Interaction processing: The input matrix I stores P features for each of the NO particle vertices, while receiving and sending matrices encode which vertices receive or send each edge.Multiplying I by these matrices and concatenating the results forms the edge-feature matrix B with dimension 2P × NE.
- Interaction processing: A trainable function fR maps sender–receiver edge features to hidden interaction representations, whose effects are summed over incoming edges for each constituent.The aggregated interaction features are appended to the original particle features before post-interaction processing.
- Classification representation: A trainable function fO converts each constituent’s combined input and interaction features into the post-interaction matrix O, which is then supplied to φC for classification.The architecture can use summed O features or flattened O elements, with the summed formulation preserving independence from graph labeling.
- Model comparison: The study compares JEDI-net with dense, convolutional, and recurrent benchmark models using the same data set and Bayesian optimization of network architectures.The JEDI-net optimization varies particle counts, hidden-layer sizes, activations, and optimizer choices; the reference result is selected from the optimized configurations.
5 Results
The optimized JEDI-net models are evaluated against DNN, CNN, and GRU taggers using cross-validated ROC curves and TPR values at 10% and 1% FPR. They outperform the alternatives in almost all cases, with specific exceptions and a small trade-off between the two JEDI-net variants.
- Performance comparison: JEDI-net outperforms the DNN, CNN, and GRU architectures in almost all tested tagging cases at the evaluated operating points.Performance is assessed using TPR values at reference FPR values of 10% and 1%, alongside ROC curves and AUC values.
- Performance comparison: At the tight top-jet working point, the DNN achieves a TPR about 2% higher than JEDI-net, while CNN and GRU perform much worse.This is identified as the only notable exception to the broader comparison.
- JEDI-net variants: The two JEDI-net variants have TPR values within 1% of each other, except at tight W- and Z-jet working points.For those exceptions, the model using summed O features shows an approximately 4% TPR drop.
- JEDI-net variants: Despite the small W- and Z-jet TPR loss, the summed-O model is preferred because it reduces model complexity and is independent of particle and edge labeling conventions.The passage explicitly connects this preference to reduced complexity and labeling independence.
6 What did JEDI-net learn?
JEDI-net learns engineered interaction features that capture some established jet-substructure information, while other learned quantities encode additional information. Several learned quantities show correlations with constituent multiplicity and N-subjettiness variables.
- JEDI-net engineers high-level features that encode information beyond the quantities used by the DNN model.
- The O1 sum correlates with jet-constituent multiplicity because comparable O magnitudes make the aggregate proportional to the number of constituents.
- Correlation coefficients between O1 and τ (β=2) are 0.69–0.97 across jet classes, a relationship described as less expected.
- Aggregated constituent-momentum information allows JEDI-net to learn a quantity very close to τ (β=2)₁, while O2 shows an intermediate correlation with τ (β=1)₁.
- The two considered O sums correlate with corresponding substructure quantities with smaller coefficients, within 0.48 and 0.77.
7 Resource comparison
JEDI-net uses fewer parameters than the CNN in its best-performing form, and its summed-O variant uses fewer parameters than the other compared models. However, the interaction-network models require substantially more operations and are currently costly for online LHC deployment.
- The best-performing JEDI-net has more than twice as many trainable parameters as the DNN and GRU, but about six times fewer than the CNN.
- The summed-O JEDI-net achieves comparable performance with about four times fewer parameters than the best-performing JEDI-net and fewer than the DNN and GRU.
- The summed-O model uses 458 MFLOP for a 150-particle input sequence, with the operation count driven mainly by graph-vertex scaling.
- GPU inference time for both interaction-network models is much larger than for the other models because they perform more operations.
- The current interaction-network algorithm is costly to deploy in a typical LHC online selection environment.
8 Conclusions
JEDI-net outperforms matched dense, convolutional, and recurrent models on the studied jet-tagging dataset while directly processing constituents without ordering or detector-geometry assumptions. Its learned quantities show both connections to established jet-substructure physics and unexplained patterns, but deployment remains computationally demanding.
- JEDI-net achieves better performance than dense, convolutional, and recurrent neural networks trained and optimized on the same dataset and procedure.
- The model directly processes constituent features, is insensitive to constituent ordering, and makes no assumptions about detector geometry.
- The current implementation requires large computational resources and inference time, limiting real-time selection without deployment optimization.
- Engineered interaction-network quantities correlate with established jet-substructure variables, indicating that the model learns relevant physics.
A Alternative models
The alternative DNN, CNN, and GRU benchmarks are selected through Bayesian hyperparameter optimization, with architectures and input representations varied across models. The optimized DNN provides a reported reference accuracy of 0.76.
- The benchmark models are derived through Bayesian optimization of their hyperparameters.The optimization uses the GpyOpt library and early stopping during training.
- DNN model: The DNN alternates dense and dropout layers, with its architecture selected by optimizing depth, width, activations, dropout, batch size, and optimizer.
- DNN model: The optimized DNN uses three hidden layers of 80 ELU-activated neurons, dropout 0.11, batch size 50, and Adam optimization.
- DNN model: The optimized DNN reaches loss 0.66 and accuracy 0.76.
- CNN model: The CNN uses two-dimensional convolutions with batch normalization, max pooling, and dense layers, while varying convolutional and dense architecture hyperparameters.
- GRU model: The GRU model feeds a recurrent layer into dense layers and optimizes recurrent width, dense architecture, dropout, batch size, and optimizer.Its best performance is obtained when the particle list is truncated at 50 elements.
B Performance on public top tagging data set
JEDI-net is retrained and evaluated on a public top-tagging dataset using Bayesian hyperparameter optimization. Its accuracy, AUC, and fixed-efficiency background rejection approach state-of-the-art performance reported for that dataset.
- JEDI-net is evaluated on a public dataset for distinguishing top-quark jets from light-quark jets.
- The optimization scans interaction-network dimensions, activation functions, and optimizers, including Adam and Adadelta with an initial learning rate of 10^-3.
- Performance is reported using accuracy, area under the ROC curve, and background rejection at 30% signal efficiency.
- JEDI-net performance approaches state-of-the-art results on the public top-tagging dataset.