-
Notifications
You must be signed in to change notification settings - Fork 352
🎉 New Accepted Solution: Top Quark Tagging Using Deep CNN #151
-
🎉 New Accepted Solution: Top Quark Tagging Using Deep CNN
Submitted by from .
Domain: High-Energy Physics Method: Deep CNN Result: ~90 % Accuracy
approaches this challenge with impressive ambition: tackling the classification of top-quark jets in high-energy collisions using a custom convolutional neural network developed in MATLAB. The challenge asked participants to design a robust workflow for detecting signatures of the top quark in collider-style data, and embraced the full pipeline—from data preparation to model design, evaluation, and deployment.
In the project, begins by ingesting a large dataset drawn from CERN’s Zenodo database, transforming raw jet constituent information (energy, momentum, spatial features) into multi-channel "jet-image" inputs for the CNN. (github.com) The model architecture integrates grouped convolutions (Aggregated Residual Transformations) together with Squeeze-and-Excite blocks for channel-level attention—a thoughtful adaptation of modern network design to a physics-inspired image representation. (github.com) Training workflows with ~90k samples achieved ~90.87% test accuracy and ~90.33% validation accuracy; even with 60k samples, ~89.4% validation accuracy was achieved. (github.com)
Beyond modeling, documented the workflow clearly: data preprocessing, image construction via MATLAB scripts, model architecture built in Deep Network Designer, training with Adam optimizer and learning-rate schedules, then finally deployment via HDL code generation and FPGA target support. (github.com) The result is not only high-performance but also reproducible and suited for real-time event-filtering contexts. Huge congratulations to for delivering a creative, technically rigorous, and well-documented solution!
Important
Why it was accepted: The solution excels in clarity (well-structured repo and documentation), originality (custom CNN architecture applied to HEP jet images), reproducibility (MATLAB scripts, clear dataset usage, deployment workflow), and strong application of MATLAB/Simulink capabilities (image generation, deep network designer, HDL codegen).
Tip
Technical highlights:
- Use of grouped convolution blocks (Aggregated Residual Transformations) with Squeeze-and-Excite channel attention.
- Generation of multi-channel jet-image data in MATLAB (including energy/pT skewness, kurtosis etc). (github.com)
- Training hyper-parameters: e.g., initial learning rate 5e-3, piecewise schedule, batch size 64, max epochs 12 (for 90k samples). (github.com)
- Achieved ~90.9% test accuracy with ~90k training images, ~89.4% validation accuracy with ~60k samples. (github.com)
- Deployment pathway: hdl_code_gen.m and deploy_on_fpga.m scripts for HDL generation and FPGA integration. (github.com)
- Strong documentation including architecture diagram, visualization of jet-images, and clear structure of image_generation → model → deploy folders. (github.com)
Open in MATLAB Online
View Repository
At a glance
- Challenge: Design a deep-learning classifier to detect top-quark jets (vs background) from collision-style jet-constituent data
- Dataset/Models: Jet-constituent data (energy, momentum_x, momentum_y, momentum_z) for ~200 constituents per jet; jet-images generated via MATLAB scripts; CNN model built in Deep Network Designer (github.com)
- Toolboxes/Methods: MATLAB Deep Learning Toolbox, Deep Network Designer, HDL Coder/Deep Learning HDL Toolbox (for FPGA)
- Key results: ~90.87% test accuracy, ~90.33% validation accuracy (90k samples); ~89.4% validation (60k samples) (github.com)
- Deployment: HDL code generation + FPGA deployment workflow via deploy_on_fpga.m and hdl_code_gen.m (github.com)
Accepted to the MATLAB & Simulink Challenge Project Hub. Congratulations again to from for their remarkable contribution!
Beta Was this translation helpful? Give feedback.