[Paper Review] Application of a Convolutional Neural Network for image classification to the analysis of collisions in High Energy Physics

Celia Fernández Madrazo, Ignacio Heredia+2 more|arXiv (Cornell University)|Aug 23, 2017

Particle physics theoretical and experimental studies4 citations

TL;DR

This paper proposes a novel image-based representation of high-energy physics collisions—converting particle momenta and jet data into 224×224×3 RGB images—enabling the application of pre-trained Convolutional Neural Networks (CNNs), such as ResNet50, for event classification. The method achieves ~95% classification efficiency in distinguishing top-quark pair (ttbar) events from background processes (Drell-Yan and W+jets), outperforming traditional feedforward neural networks in signal-vs-background discrimination.

ABSTRACT

The application of deep learning techniques using convolutional neural networks to the classification of particle collisions in High Energy Physics is explored. An intuitive approach to transform physical variables, like momenta of particles and jets, into a single image that captures the relevant information, is proposed. The idea is tested using a well known deep learning framework on a simulation dataset, including leptonic ttbar events and the corresponding background at 7 TeV from the CMS experiment at LHC, available as Open Data. This initial test shows competitive results when compared to more classical approaches, like those using feedforward neural networks.

Motivation & Objective

• To explore the feasibility of applying deep learning techniques, specifically convolutional neural networks (CNNs), to particle collision classification in high-energy physics.
• To develop an intuitive visual representation of particle physics events by transforming momentum and jet data into 224×224×3 RGB images for CNN input.
• To evaluate the performance of CNNs against classical machine learning methods, such as feedforward neural networks, in classifying complex collision events.
• To assess the potential of CNNs for signal-background discrimination in ttbar events using real simulation data from the CMS experiment at the LHC.

Proposed method

• Transforming particle-level observables (momenta, jet information) from simulated proton-proton collisions into 224×224×3 RGB images using a custom Python pipeline.
• Employing a pre-trained ResNet50 architecture with ImageNet-preinitialized weights for transfer learning on particle physics image data.
• Training the CNN for 40 epochs using the Adam optimizer on high-performance GPUs, with data augmentation (rotation, translation, mirroring) applied during training.
• Using a standard 80/10/10 split for training, validation, and test sets to ensure robust model evaluation and prevent overfitting.
• Applying image resizing to standardize inputs to 224×224 pixels, compatible with the ResNet50 input requirement.
• Comparing results with a 5-layer feedforward neural network (FNN) with 500 units per layer and 50% dropout to evaluate relative performance.

Experimental results

Research questions

• RQ1 Can a convolutional neural network effectively classify high-energy physics collision events when particle data is represented as images?
• RQ2 How does the performance of a CNN compare to a traditional feedforward neural network in distinguishing ttbar events from background processes?
• RQ3 To what extent does the image-based representation preserve the physical information necessary for accurate classification?
• RQ4 Does the use of transfer learning with ImageNet-pretrained weights improve classification performance on simulated LHC data?

Key findings

• The CNN achieved a signal-vs-background classification efficiency of 95.4% when distinguishing ttbar+jets events from Drell-Yan and W+jets backgrounds in a 50/50 split.
• The CNN correctly classified 95% of ttbar+jets events and misclassified only 4% of W+jets and 5% of Drell-Yan events as signal.
• In comparison, the feedforward neural network achieved a lower signal-vs-background efficiency of 94.6%, indicating superior performance of the CNN for this specific task.
• The CNN outperformed the FNN in signal-vs-background discrimination, despite the FNN showing better performance on simpler classification tasks like invariant mass tagging.
• The image representation method allowed for intuitive handling of variable-length events by simply adding circles for additional particles, unlike vector-based FNNs that required zero-padding.
• The preprocessing pipeline, including JSON-based data extraction and image generation, proved effective and reusable, with code publicly available.