Omar-Camara/nas-differential-evolution

Neural Architecture Search with Differential Evolution

Interactive demo of automated neural network architecture discovery

🎯 Overview

This Space demonstrates a production-ready Neural Architecture Search (NAS) system that automatically discovers optimal neural network architectures using Differential Evolution.

📊 Key Results

• —Best Architecture Found: [21, 48, 11] (3-layer hourglass pattern)
• —Test Accuracy: 69.91%
• —Improvement over Baseline: +2.67% (67.24% → 69.91%)
• —Beat Random Search: +0.88% with same computational budget
• —Search Time: 33.6 minutes on Tesla T4 GPU

🔬 Experiments Included

1. Random Search Baseline

Compared DE against 144 random architectures to prove optimization works. Result: DE found 0.88% better architecture.

2. Ablation Study

Systematically tested each component's impact. Key Finding: Single-trial evaluation outperformed multi-trial averaging at short horizons (+0.24% accuracy, -45% time).

3. Statistical Validation

Final evaluation over 5 independent trials with confidence intervals.

🛠️ Technologies

• —PyTorch 2.0+ with CUDA
• —Differential Evolution optimization
• —UCI Adult Income dataset (39K samples)
• —GPU acceleration (Tesla T4)
• —Statistical validation

📚 Learn More

• —Full Code: Google Colab Notebook
• —GitHub: Repository
• —Author: Omar | Syracuse University

🎓 About

Built by Omar Camara

📄 License

MIT License - Free to use with attribution

Built with ❤️ using PyTorch, Gradio, and Differential Evolution