neural architecture search (nas)

Neural Architecture Search (NAS) automatically discovers optimal model architectures instead of manual design. **Motivation**: Architecture design requires expertise and intuition. Automate to find better architectures efficiently. **Search space**: Define possible operations (conv sizes, attention types), connectivity patterns, depth/width ranges. **Search methods**: **Reinforcement learning**: Controller network proposes architectures, trained on validation performance. **Evolutionary**: Population of architectures, mutate and select best. **Gradient-based**: Differentiable architecture, learn architecture parameters (DARTS). **Weight sharing**: Train supernet containing all possible architectures, evaluate subnets. **Compute cost**: Early NAS required thousands of GPU-days. Modern methods reduce to GPU-hours through weight sharing. **Notable success**: EfficientNet family found by NAS, outperformed manual designs. AmoebaNet, NASNet. **For transformers**: AutoML searches over attention patterns, FFN sizes, layer configurations. **Search vs transfer**: Once good architecture found, transfer to new tasks. NAS is research tool. **Current status**: Influential for initial architecture discovery, but recent trend toward scaling simple architectures (plain transformers) rather than complex search.