CNN on `Top': In Search of Scalable & Lightweight Image-based Jet Taggers

arXiv:2512.05031 · hep-ph, physics.comp-ph, physics.data-an · Submitted 2025-12-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "CNN on `Top'".

Mira: Here's the gist of the paper. CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets amidst other light quark and gluon jets 'Fat jet tagging' is a critical task for identifying the hadronic decays of massive particles, such as the Higgs boson, top quark, or a W or Z-boson that are produced with high momentum or ‘boost’ in a particular direction The work also sheds light on the importance of global features for both the accuracy and the apparent redundancy of the network’s complexity Identifying top quark jets as fat jets in a sea of other light quark and gluon jets, therefore, is a rather important task Network Architectures and Representation Different types of neural networks have been developed over the last couple of decades, each tailor-made to embed input tensors of a specific type A Convolutional neural network (CNN) uses 2D convolutions on an image to do this A recurrent neural network (RNN) is best suited for data presented as sequences, e.g., audio files, time series, and natural language A graph neural network (GNN) works best on data with distinct global, node/vertex, and edge features The jet is now treated as a graph where each particle is a ‘node’, and the connections between them are ‘edges’ The advent of the self- and cross-attention mechanisms gave us the famously parallelizable and scalable Transformer architecture Data Preprocessing and Feature Engineering We pre-process the data to create 3-channel images with transverse momentum (pT), mass (m), and energy (E) in each channel, respectively To create each channel from the jet constituents, the hardest constituent is considered to be the center of the channel, and the rest of the image is constructed by binning the remaining constituents in two dimensions: the difference in pseudo-rapidity (∆η) and that in the azimuthal angle (∆ φ) Standardization of input data has become common in ML, as it has been found to improve the performance of generic neural networks We perform this standardization by creating a ‘mean’ and a ‘standard deviation’ image using a large number of random samples from the dataset This clearly isolates a fairly circular region in the ∆η −∆ φ plane containing the bulk of the jet structure Model Comparison and Performance Metrics Table 2 lists all single-number performance metrics for the networks explored in this analysis The smallest EffNet-S, while applied only to images, achieves accuracy similar to the largest LeNet, while being 1/8-th in size (the total number of parameters) Adding global features is beneficial here as well, but the effect is less prominent for larger networks In figures (3a) and (3b), accuracies of the nets are compared, with their model complexity in the x-axis We have also added the optimal possible ROC curve obtained from likelihood ratio analysis using generative models to showcase the space allowed for future progress We observe that adding global features to the models, along with the images, results in better accuracy This exploration leads to three observations: first, there needs to be a systematic architecture search for versions of scalable models such as EfficientNet, tuned for jet classification second, adding global jet features, in addition to the constituent information, makes it easier for smaller models to perform better last but not least, as jets can have multiple types of representation, e.g., image, constituent four-momenta, and global jet features, one can combine the strengths of different network architectures and build a hybrid ‘mixture of experts’ ensemble of lightweight networks with the potential for superior performance Page 1 --- CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets amidst other light quark and gluon jets Page 2 --- CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets amidst other light quark and gluon jets Page 3 --- CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets amidst other light quark and gluon jets We observe that adding global features to the models, along with the images, results in better accuracy This exploration leads to three observations: first, there needs to be a systematic architecture search for versions of scalable models such as EfficientNet, tuned for jet classification second, adding global jet features, in addition to the constituent information, makes it easier for smaller models to perform better last but not least, as jets can have multiple types of representation, e.g.

Paper summary: Kai: We talked about the core idea of "CNN on Top: In Search of Scalable and Lightweight Image-based Jet Taggers". The paper is looking at using a lightweight and scalable version of the EfficientNet architecture alongside global features from the jet data to get competitive performance when tagging top quark jets in a sea of light quark and gluon jets.

Mira: Basically, they are exploring how to achieve this without needing the huge computational resources that Transformer-based or standard Graph Neural Networks typically demand. They argue that this approach is computationally inexpensive but still capable of achieving good tagging accuracy for top quark jets compared to other light quark or gluon jets.

Lev: I remember thinking about how much parameter count matters for real hardware deployment, and this paper seems to be specifically interested in finding a model that balances low complexity with usable performance on the task.

Kai: The work shows the efficacy of their network when tagging top quark jets within a sea of other light quark and gluon jets, while also highlighting how important global features are for both the accuracy and the apparent redundancy of the network's complexity.

Mira: They aren't just showing a single setup; they are comparing different ways these models can be applied independently to different data representations, like pixels from an image, vectors, or graphs. This shows how flexible these network architectures can be when you change how you represent the input data.

Lev: That flexibility is key because in real-world physics experiments, the way we represent a particle decay—as an image or a set of nodes—can drastically change what kind of model works best for us to train.

Kai: The paper lays out that they start by representing calorimeter energy deposits as a 2D image grid where pixel intensity represents the energy or transverse momentum <ref:2512.05031#pg1>. This is the starting point for using CNNs in jet embedding and sub networks, followed by dense or linear networks.

Mira: They then move on to creating these standardized images, showing how cropping and standardization steps affect the input resolution, as seen in figures like one which compare different image resolutions <ref:2512.05031#pg1>.

Lev: So they are showing the practical pipeline from raw energy deposits to a standardized image that feeds into their network architecture, and it’s interesting how they vary the resolution to see what happens.

Kai: They also discuss using different inputs altogether, like constituent four-momenta or those global jet features as nodes and edges in a graph, which is another way to structure the data for a GNN approach.

Mira: So this paper is really showing that there isn't one single best way to do jet tagging; you have multiple representations—image, constituent information, and global features—and you can use different network architectures depending on what representation you choose.

Lev: And from an engineering standpoint, it’s important to see how much overhead each of these different input structures adds when we try to run them on actual hardware.

Conclusion: Kai: So wrapping up this discussion on "CNN on Top: In Search of Scalable and Lightweight Image-based Jet Taggers". The authors are essentially proving that you can achieve competitive performance by using a lightweight EfficientNet architecture plus global jet features.

Mira: They highlight the main implications around two things: the first is that there needs to be a systematic search for scalable models like EfficientNet specifically tuned for jet classification, and the second is that adding those global jet features on top of constituent information makes smaller models perform better.

Lev: From a practical standpoint, this means we need to prioritize finding ways to build these specific model versions efficiently so they can run reliably in an experimental environment.

Kai: And the final point they make is that because jets offer different kinds of representation—images, four-momenta, and those global features—we should build hybrid "mixture of experts" ensembles of lightweight networks that can combine the strengths of these different approaches for better results.

Mira: It suggests we don't have to pick one single representation and stick with it; you can mix and match architectures to build something that leverages the specific strengths each data type offers.

Lev: So the main implication is shifting focus toward building these hybrid systems rather than just optimizing one monolithic network, which is a direction that seems very practical for tackling these kinds of problems on real hardware.

Department of Physics, SEAS, Bennett University · Department of Physics, Bangabasi Evening College

hep-ph, physics.comp-ph, physics.data-an

Submitted: 2025-12-04

Updated: 2026-11-04

Comments: 15 pages, 5 figures, 3 tables, version accepted by Eur. Phys. J. C

Code: https://github.com/tensorflow/tpu

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 70/100

The gist: CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets…

Key concepts

Fat Jet Tagging
'Fat jet tagging' is the crucial task of identifying hadronic decays from massive particles like top quarks or Higgs bosons that are produced with high momentum. Distinguishing these 'fat jets' from lighter quark and gluon jets is important for particle identification in high-energy physics.
Graph Neural Network (GNN)
A GNN treats a jet as a graph where individual particles are nodes and connections between them are edges. This structure is used to model the complex relationships within the jet constituents, allowing the network to capture both local particle information and overall structural features of the jet.
Global Features
Global features refer to information that summarizes the entire jet structure rather than just individual particles or local patterns. In this context, these global features help models better distinguish top quark jets from other types of jets by capturing broader characteristics of the jet's overall shape and distribution.

Terminology

Summary

CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets amidst other light quark and gluon jets

Motivation for the Study

'Fat jet tagging' is a critical task for identifying the hadronic decays of massive particles, such as the Higgs boson, top quark, or a W or Z-boson that are produced with high momentum or ‘boost’ in a particular direction The work also sheds light on the importance of global features for both the accuracy and the apparent redundancy of the network’s complexity Identifying top quark jets as fat jets in a sea of other light quark and gluon jets, therefore, is a rather important task

Network Architectures and Representation

Different types of neural networks have been developed over the last couple of decades, each tailor-made to embed input tensors of a specific type A Convolutional neural network (CNN) uses 2D convolutions on an image to do this A recurrent neural network (RNN) is best suited for data presented as sequences, e.g., audio files, time series, and natural language A graph neural network (GNN) works best on data with distinct global, node/vertex, and edge features The jet is now treated as a graph where each particle is a ‘node’, and the connections between them are ‘edges’ The advent of the self- and cross-attention mechanisms gave us the famously parallelizable and scalable Transformer architecture

Data Preprocessing and Feature Engineering

We pre-process the data to create 3-channel images with transverse momentum (pT), mass (m), and energy (E) in each channel, respectively To create each channel from the jet constituents, the hardest constituent is considered to be the center of the channel, and the rest of the image is constructed by binning the remaining constituents in two dimensions: the difference in pseudo-rapidity (∆η) and that in the azimuthal angle (∆ φ) Standardization of input data has become common in ML, as it has been found to improve the performance of generic neural networks We perform this standardization by creating a ‘mean’ and a ‘standard deviation’ image using a large number of random samples from the dataset This clearly isolates a fairly circular region in the ∆η −∆ φ plane containing the bulk of the jet structure

Model Comparison and Performance Metrics

Table 2 lists all single-number performance metrics for the networks explored in this analysis The smallest EffNet-S, while applied only to images, achieves accuracy similar to the largest LeNet, while being 1/8-th in size (the total number of parameters) Adding global features is beneficial here as well, but the effect is less prominent for larger networks In figures (3a) and (3b), accuracies of the nets are compared, with their model complexity in the x-axis We have also added the optimal possible ROC curve obtained from likelihood ratio analysis using generative models to showcase the space allowed for future progress

Conclusions and Future Directions

We observe that adding global features to the models, along with the images, results in better accuracy This exploration leads to three observations: first, there needs to be a systematic architecture search for versions of scalable models such as EfficientNet, tuned for jet classification second, adding global jet features, in addition to the constituent information, makes it easier for smaller models to perform better last but not least, as jets can have multiple types of representation, e.g., image, constituent four-momenta, and global jet features, one can combine the strengths of different network architectures and build a hybrid ‘mixture of experts’ ensemble of lightweight networks with the potential for superior performance

--- Page 1 ---

CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets amidst other light quark and gluon jets

--- Page 2 ---

CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets amidst other light quark and gluon jets

--- Page 3 ---

CNN on ‘Top’ explores using a lightweight and scalable version of EfficientNet architecture along with global features of the jet to achieve competitive performance in tagging top-quark jets amidst other light quark and gluon jets

We observe that adding global features to the models, along with the images, results in better accuracy This exploration leads to three observations: first, there needs to be a systematic architecture search for versions of scalable models such as EfficientNet, tuned for jet classification second, adding global jet features, in addition to the constituent information, makes it easier for smaller models to perform better last but not least, as jets can have multiple types of representation, e.g.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Lightweight Network Construction for Jet Tagging: Implement lightweight versions of EfficientNet following prescribed scaling rules to build networks applicable for low resolution jet images without increasing RAM or VRAM demands significantly. This allows for deployment on more constrained computational resources while maintaining competitive performance against larger models like ResNeXt-50.

  2. Global Feature Integration Strategy: Systematically explore the way global jet features are fed into CNNs, either directly to the images or to the network output, as this makes it easier for smaller models to perform better. This exploration should focus on finding the optimal way these features interact with constituent-level information.

  3. Hybrid Mixture of Experts Ensemble: Combine different network architectures—such as CNNs (for spatial arrangement), particle four-momenta (for constituent information), and global jet features—to build a hybrid ensemble of lightweight networks for superior signal to background discrimination. This approach leverages the strengths of different representations, potentially leading to better performance than any single architecture.

Abstract

While Transformer-based and standard Graph Neural Networks (GNNs) have proven to be the best performers in classifying different types of jets, they require substantial computational power. We explore the scope of using a lightweight and scalable version of EfficientNet architecture, along with global features of the jet. The end product is computationally inexpensive but is capable of competitive performance. We showcase the efficacy of our network in tagging top-quark jets in a sea of other light quark and gluon jets. The work also sheds light on the importance of global features for both the accuracy and the apparent redundancy of the network's complexity.

Sources

Related papers