CS5720 - Week 12
Slide 224 of 240

Model Architecture Search (NAS)

Automating Architecture Design

Neural Architecture Search (NAS) automates the design of neural network architectures, using algorithms to discover optimal structures that often outperform hand-crafted designs.
Key Components:

Search Space - Possible architectures to explore
Search Strategy - How to explore the space
Performance Estimation - Evaluate architectures efficiently
Hardware Constraints - Memory, latency requirements
🚀 Why Use NAS?
Discovers novel architectures that achieve better accuracy-efficiency trade-offs than human designs, especially for specific datasets and hardware constraints.

NAS Approaches

  • 🎮
    Reinforcement Learning NAS
    Controller network learns to design architectures
  • 🧬
    Evolutionary NAS
    Evolve populations of architectures over generations
  • 📈
    Differentiable NAS (DARTS)
    Gradient-based optimization of architecture parameters
  • One-Shot NAS
    Train supernet once, then search efficiently

NAS Process Visualization

Search Space
• Conv layers
• Skip connections
• Pooling types
• Activations
Controller
(RL/Evolution)
Evaluation
Train & Test
Architecture
Performance
Click on any method or component to learn more!
Prepared by Dr. Gorkem Kar