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

TL;DR

This work investigates scalable, lightweight CNN-based top-quark jet tagging by combining compact EfficientNet-S variants with global jet features. Using a large Pythia8/Delphes dataset to create 3-channel jet images (32×32, cropped variants) and extensive jet-level observables, the study compares LeNet and EfficientNet-S architectures with and without global features. The results show that augmenting images with global jet features consistently boosts performance, and EfficientNet-S achieves competitive accuracy and AUC at a fraction of the computational cost of heavier CNNs, suggesting practical viability for low-resource environments. The authors advocate for systematic architecture searches and explore the potential of hybrid ensembles that fuse multiple representations (image and non-image) for further gains in jet tagging efficacy.

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 the 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 for tagging top-quark jets in a sea of other light-quark and gluon jets.

Figures (3)

• Figure 1: Representative images of two jets for different resolutions and pixel standardization. Although images in our dataset are rendered in RGB, we present them here using the CEI 1931 XYZ color scheme xyzcolorCEI:1931, for better visual understanding. All images from the top panel are of one representative top jet, and those below are of one QCD jet. The first column shows the jet constituents as a $32\times 32$ image in the $\Delta\eta - \Delta\phi$ plane, as the channels are combined and rendered. The second panel in the same row contains the same image, with the pixels standardized (the mean image of the dataset is subtracted from the original and then divided by the standard deviation image). The third column displays the images in the second, but cropped to $28\times 28$. The fourth column contains the same jets, first rendered in $40\times 40$, then cropped to $32\times 32$.
• Figure 2: Schematics of network architectures used in this work. On the left is LeNet, used as the benchmark CNN model. On the Right is the simplest Efficient Net (small) used. Representative schematics of the M-Conv blocks and the squeeze and expand residual blocks within those are also shown.
• Figure 3: ROC curve (signal efficiency vs. background rejection) for some of the networks (with some metric highest among others; table (\ref{['tab:pointres']}) has them in bold-face). The solid lines are for models working on both images and global features, dashed ones correspond to those with only images as input.