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

This paper demonstrates that a lightweight and scalable EfficientNet architecture, augmented with global jet features, achieves competitive top-quark jet tagging performance while significantly reducing computational costs compared to Transformer-based and standard Graph Neural Network approaches.

The Gist

The Big Picture: Finding a Needle in a Haystack

Imagine the Large Hadron Collider (LHC) as a massive, high-speed particle smasher. When it smashes protons together, it creates a chaotic explosion of debris. Most of this debris is just "junk" (light quarks and gluons), but occasionally, a heavy, rare particle like a Top Quark is born.

The Top Quark is special. It's the heaviest known particle and decays so fast it doesn't even have time to form a stable atom. Instead, it explodes into a tight, high-speed bundle of other particles. Physicists call this bundle a "Fat Jet."

The Problem: The LHC produces billions of these jets. Finding the rare "Top Quark Fat Jets" hidden among the billions of ordinary "junk" jets is like trying to find a specific, glowing needle in a giant, dark haystack.

The Old Way vs. The New Way

For a long time, physicists used "hand-crafted" rules to find these needles. They looked for specific shapes or weights, like a detective looking for a specific fingerprint.

Then, Machine Learning (AI) came along. Instead of giving the computer rules, we let it learn by looking at pictures of the jets.

• The Heavyweights: Recently, the best AI models (like Transformers and Graph Neural Networks) have been like super-computers. They are incredibly smart and accurate, but they are also expensive. They require massive data centers and huge amounts of electricity to run. It's like using a nuclear-powered tank to hunt for a rabbit.
• The Goal of This Paper: The authors asked, "Can we build a lightweight, fuel-efficient car that is just as good at catching the rabbit?" They wanted a model that is fast, cheap to run, and doesn't need a supercomputer, but still finds the Top Quarks accurately.

The Solution: The "EfficientNet" and the "Global Cheat Sheet"

1. The Camera Lens (CNNs)

The team used a type of AI called a Convolutional Neural Network (CNN). Think of this as a camera that looks at the jet as a 2D image.

• They take the particles flying out of the jet and map them onto a grid (like a photo).
• Brighter pixels mean more energy.
• The AI looks at this "photo" to decide: "Is this a Top Quark jet or just junk?"

2. The "Efficient" Engine

Most high-accuracy AI models are huge and slow. The authors chose a specific architecture called EfficientNet.

• The Analogy: Imagine a standard deep learning model is a giant, heavy truck with a massive engine. It can carry a lot, but it guzzles gas. EfficientNet is like a hybrid sports car. It uses a special design (called "compound scaling") that makes it deeper and wider without making it heavier. It gets the same mileage (accuracy) but uses a fraction of the fuel (computing power).
• They created a "mini" version of this car (called EffNet-S) specifically for their small, low-resolution jet images.

3. The "Global Cheat Sheet" (Global Features)

Here is the clever twist. The AI looks at the "photo" of the jet, but sometimes the photo is blurry or hard to read.

• The authors decided to give the AI a cheat sheet alongside the photo.
• This cheat sheet contains Global Features: simple numbers describing the whole jet, like its total weight, its speed, and how "spread out" the particles are.
• The Metaphor: Imagine you are trying to identify a suspect from a grainy security camera photo (the image). It's hard. But if you also tell the detective, "The suspect is 6 feet tall and wearing a red hat" (the global features), the job becomes much easier.

What Did They Find?

1. Small but Mighty: Their lightweight "hybrid car" (EffNet-S) was almost as good at finding Top Quarks as the massive "super-computers," but it was much faster and required much less computing power.
2. The Cheat Sheet Works: Adding the "global features" (the cheat sheet) boosted the performance of even the smallest models significantly. It was like giving a small detective a magnifying glass and a list of suspects.
3. The Trade-off: Interestingly, when they added the cheat sheet, the size of the AI model mattered less. A small model with the cheat sheet performed just as well as a giant model with the cheat sheet. This suggests that the "global features" do a lot of the heavy lifting, so you don't need a massive brain to process the image.

Why Does This Matter?

In the future, the LHC will be running even faster (High Luminosity LHC), producing even more data.

• If we use the "nuclear-powered tanks" (huge AI models), we might run out of electricity or computing time.
• By using these lightweight, efficient models, physicists can run these analyses in real-time, on smaller computers, or even on the detectors themselves.

The Bottom Line

The authors proved that you don't need a supercomputer to find the most interesting particles in the universe. By using a smart, compact AI design and giving it a little bit of extra context (global features), you can get top-tier results with a fraction of the cost. It's a win for physics and a win for the environment!

Technical Summary

1. Problem Statement

In the High-Luminosity Large Hadron Collider (HL-LHC) era, identifying "fat jets" (jets resulting from the hadronic decay of massive, boosted particles like the top quark) amidst a sea of light quark and gluon jets is critical for new physics searches.

• The Challenge: While Transformer-based models and Graph Neural Networks (GNNs) currently offer the highest accuracy for jet tagging, they are computationally expensive. They often require constructing fully connected graphs (pairwise interactions) or massive foundational models, making them difficult to deploy on standard hardware or for real-time inference.
• The Goal: The authors aim to develop a computationally lightweight and scalable Convolutional Neural Network (CNN) architecture that achieves competitive performance in top-quark jet tagging without the heavy computational overhead of Transformers or large GNNs.

2. Methodology

A. Data Preparation

• Dataset: A pre-existing dataset of 2 million events (1M top-quark jets, 1M QCD background jets) generated at $\sqrt{s} = 14$ TeV using Pythia8 and simulated with Delphes.
• Jet Selection: Top jets are selected with $p_T \in [550, 650]$ GeV and radius $R=0.8$. All decay products must be within $\Delta R = 0.8$.
• Image Representation:
  • Constituents are binned into 3-channel images representing Transverse Momentum ($p_T$), Mass ($m$), and Energy ($E$).
  • Resolution: Two primary resolutions were tested: $35 \times 35$ and $40 \times 40$.
  • Preprocessing: Images were standardized (mean subtraction and division by standard deviation) and center-cropped to $28 \times 28$ (from 35) and $32 \times 32$ (from 40) to isolate the jet core.
  • Augmentation: Random flipping along axes was used; random cropping was avoided to preserve the jet center.

B. Global Features

To enhance tagging efficiency, the authors incorporated a set of global jet-level features alongside the images. These include:

• Jet four-momentum ($p_T, \eta, \phi, m_{jet}$).
• Number of constituents ($N_{cons}$).
• N-subjettiness ratios ($\tau_{21}, \tau_{32}, \dots$).
• Energy Correlation Functions (C-series, D-series, U-series, M-series, N-series, L-series).
• Note: Jets with fewer than 4 constituents were discarded to ensure all features could be computed.

C. Model Architectures

The study compares two main CNN architectures:

1. LeNet (Benchmark): A traditional, simple CNN used as a baseline. It was tested with varying input resolutions.
2. EfficientNet-Small (EffNet-S): A lightweight version of the EfficientNet architecture.
   • Scaling Strategy: Instead of using the standard high-resolution inputs ($224 \times 224$) required by original EfficientNets, the authors applied compound scaling with negative scaling parameters ($\phi$) to downscale the architecture for low-resolution ($36 \times 36$) inputs.
   • Structure: Utilizes Mobile Inverted Bottleneck (MBConv) blocks with depthwise and pointwise convolutions to reduce computational cost while maintaining performance.
   • Integration: Global features were concatenated with the output of the network's aggregation layer (after flattening) and fed into a Multi-Layer Perceptron (MLP) block.

D. Training Configuration

• Hardware: Trained on a single desktop PC (Intel i9, 64GB RAM, NVIDIA RTX A2000).
• Procedure: Data was processed in rounds (300k samples) to fit memory. Models were trained until validation loss plateaued.
• Optimization: ADAM optimizer with cross-entropy loss. Batch sizes were adjusted based on GPU memory constraints.

3. Key Contributions

1. Lightweight Architecture for Jet Physics: The successful adaptation of EfficientNet for low-resolution jet images, creating a "EffNet-S" variant that is significantly smaller than standard CNNs (e.g., ResNet-50) while maintaining high accuracy.
2. Global Feature Integration: Demonstrating that combining image-based CNNs with global kinematic features significantly boosts performance, particularly for background rejection.
3. Efficiency vs. Accuracy Trade-off Analysis: Providing a comprehensive benchmark showing that lightweight models can achieve performance comparable to much larger, computationally expensive models (like ResNeXt-50) when global features are utilized.
4. Systematic Scaling Study: Investigating how input resolution and model depth/width interact in the context of sparse jet images, revealing that standard scaling rules for natural images do not always translate directly to jet physics without modification.

4. Results

• Performance Metrics:

  • Accuracy: The best EffNet-S model (with global features) achieved an accuracy of ~93.38% and an AUC of ~98.25%.
  • Comparison: This performance rivals or exceeds larger models like ResNeXt-50 (93.60% accuracy) and DeepTop (93.00% accuracy).
  • Background Rejection: At 50% signal efficiency, the best models achieved a QCD background rejection of ~98.2%.

• Computational Efficiency:

  • Parameter Count: The best EffNet-S models have ~422K parameters, which is roughly 1/7th the size of ResNeXt-50 (~1.46M parameters) and significantly smaller than standard EfficientNets (B0 has 5.3M).
  • Inference Speed: EffNet-S models are approximately twice as fast as ResNeXt-50. The inference time for 100 jets was roughly 17ms for the largest EffNet-S compared to 302ms for ResNeXt-50.

• Impact of Global Features:

  • Adding global features improved accuracy and background rejection across all architectures.
  • Interestingly, for models with global features, performance became less dependent on network complexity (depth/width), suggesting the global features provide a strong signal that compensates for simpler architectures.

• Resolution Sensitivity:

  • For LeNet, higher resolution inputs ($40 \times 40$) generally improved performance.
  • For EffNet-S, performance was optimal at lower resolutions ($35 \to 28$) and degraded slightly at higher resolutions ($40 \to 32$). The authors hypothesize that the "bottleneck" structure of MBConv blocks struggles with the sparsity of high-resolution jet images compared to full convolutions.

5. Significance and Conclusion

This paper demonstrates that computationally efficient, image-based CNNs are a viable alternative to heavy Transformer or GNN models for jet tagging in high-luminosity environments.

• Practicality: The proposed models can run on standard desktop hardware, making them accessible for a wider range of researchers and potentially suitable for online trigger systems where latency is critical.
• Feature Synergy: The work highlights that the "redundancy" of complex network architectures can be mitigated by the inclusion of physics-inspired global features.
• Future Directions: The authors suggest that future work should focus on a systematic architecture search specifically tuned for jet images (rather than adapting ImageNet models) and exploring hybrid "Mixture of Experts" ensembles that combine the strengths of different network types (images, sequences, and global features).

In summary, the authors successfully proved that lightweight, scalable CNNs augmented with global features can achieve state-of-the-art top-quark tagging performance with a fraction of the computational cost of current leading models.