-
Notifications
You must be signed in to change notification settings - Fork 353
🎉 New Accepted Solution: Top Quark Tagging Using Deep CNN #149
-
🎉 New Accepted Solution: Top Quark Tagging Using Deep CNN
Submitted by Adit Smoak from [School Name].
Domain: High-Energy Particle Physics Method: Deep CNN + MATLAB Result: ~90.9% Accuracy Project Description
Adit Smoak tackles this challenge by building a custom deep convolutional neural network in MATLAB’s Deep Network Designer to classify jets produced in high-energy collisions — specifically distinguishing top quark events from background. His solution uses jet "images" derived from particle-level features (energy, momentum, pT, etc.) and channels that capture skewness, kurtosis and radial profiles. (github.com) Through this work, Adit demonstrates careful architecture design (aggregated residual transforms + squeeze-and-excite blocks) and obtains around 90.87% test accuracy on ~90 k samples. (github.com) Congratulations to Adit for delivering a creative, well-documented and high-impact solution! 🎊
Important
Why it was accepted: The submission combines strong clarity (complete README + architecture description), technical quality (custom CNN blocks and real-world physics dataset) and practical deployment awareness (HDL code generation + FPGA target). These qualities aligned well with the Challenge’s criteria.
Tip
Technical highlights:
- Multi-channel jet-image generation (energy/pT stats, radial profiles) using MATLAB scripts. (github.com)
- Custom CNN architecture with grouped convolutions (aggregated residual blocks) and Squeeze-and-Excite channel attention. (github.com)
- Training on ~90 k samples, achieving ~90.87% test accuracy and ~90.33% validation accuracy. (github.com)
- HDL code generation & FPGA deployment workflow via MATLAB’s Deep Learning HDL Toolbox. (github.com)
Open in MATLAB Online View Repository
At a glance
- Challenge: Real-time classification of particle jets for top-quark tagging in high-energy physics. (github.com)
- Dataset/Models: Jet data from CERN’s Zenodo repository (200 constituents per jet, zero-padded). (github.com)
- Toolboxes/Methods: MATLAB Deep Learning Toolbox, Deep Network Designer, grouped convolutional blocks, SE blocks, image generation via MATLAB scripts. (github.com)
- Key results: ~90.87% test accuracy (90 k samples) and ~90.33% validation accuracy; ~89.4% validation accuracy with 60 k samples. (github.com)
- Deployment: HDL code generation and FPGA deployment using MATLAB’s HDL code generation tools. (github.com)
Accepted to the MATLAB & Simulink Challenge Project Hub. Congratulations again to Adit Smoak from [School Name] for their remarkable contribution!
Beta Was this translation helpful? Give feedback.