AutoML for Networks
Automatically design neural networks.
Search Space
Cells: building blocks. Operations: conv, pooling, attention. Connectivity: how cells connect.
Search Strategies
Random search: baseline. Grid search: inefficient. Evolutionary: mutate, select. Reinforcement: reward accuracy.
Efficient Methods
Weight sharing: share weights across architectures. Early stopping: prune unpromising. DARTS: differentiable search.
Key Takeaways
- NAS automates architecture design
- Search space defines possible networks
- Efficient methods needed