Skip to content

jkhlim125/Constraint-Aware-Neural-Network-Systems

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

43 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Constraint-Aware Neural Network Systems

UROP 1100 research (HKUST, Advisor: Prof. Wei Zhang) on hardware-efficient neural networks — training and pruning networks so that FPGA cost (pruned fan-in, packable LUT structure) is part of the objective, not bolted on after. Three connected parts:

Folder What it covers Headline result
LutNet/ Analysis + pruning-correctness fix for a LUT-CNN framework (CIFAR-10) Fixed a tie-breaking pruning bug (682 → 691, exact target); mapped the accuracy–hardware frontier (Spearman ρ = −0.96), up to 86% pin / 81% slice reduction
RadioML/ Conventional 1-D CNN modulation classification + structured failure analysis Baseline 60.36%, two-branch V3 61.90%; diagnosed AM-SSB "sink" (72%) and WBFM → AM-DSB confusion (63.6%)
RadioML-LUT/ Porting the LUT hardware-aware workflow onto RadioML 1-D IQ Ablation: pruning-first 40% vs no-prune control 29% under full hardening (+11 pts); ≈38% pin pruning

The common question across all three:

How should model structure and feature representation be designed to improve both learning behavior and hardware efficiency?

LutNet/ establishes the hardware-aware method and its trade-offs; RadioML/ builds intuition for a new signal domain and where models fail; RadioML-LUT/ brings the two together by running the hardware-aware workflow on that domain.


My contributions

  • Reverse-engineered and analyzed an existing LUT-CNN training/pruning pipeline.
  • Found and fixed a sensitivity-pruning correctness bug caused by threshold ties.
  • Ran a 40+ configuration sweep of the accuracy ↔ hardware trade-off and produced the analysis figures.
  • Ported the LUT operators and staged schedule to RadioML 1-D IQ, and ran the pruning-vs-no-prune ablation.
  • Built the failure-analysis tooling (class/SNR breakdowns, confusion, confidence-margin analysis).

Scope note

This repository emphasizes analysis, adaptation, and experiment structure. It does not include the full original collaborative LUT framework; some code is reorganized to highlight my direct contributions. All result files and figures are my own experiment outputs.

Repository layout

LutNet/        analysis_code/  figures/  results/            + README
RadioML/       models/ data/ training/ analysis/ figures/ results/   + README
RadioML-LUT/   figures/ results/                             + README

About

Exploring constraint-aware neural network design through LUT-based FPGA pruning and RadioML signal classification analysis.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages