proxylessnas

**ProxylessNAS** is a **NAS method that directly searches on the target hardware and target dataset** — eliminating the need for proxy tasks (smaller datasets, shorter training) that introduce a gap between the searched and deployed architecture. **How Does ProxylessNAS Work?** - **Direct Search**: Searches directly on ImageNet (not CIFAR-10 proxy) and on the target hardware (GPU, mobile, etc.). - **Path-Level Binarization**: At each step, only one path (operation) is active -> memory-efficient (don't need to run all operations simultaneously like DARTS). - **Latency Loss**: Includes a differentiable latency predictor in the search objective: $mathcal{L} = mathcal{L}_{CE} + lambda cdot Latency$. **Why It Matters** - **No Proxy Gap**: Architectures searched directly on the target task & hardware generalize better. - **Hardware-Aware**: Different architectures for GPU, mobile CPU, and edge TPU — each optimized for its platform. - **Memory Efficient**: Binary path sampling uses ~50% less memory than DARTS. **ProxylessNAS** is **searching where you deploy** — finding the best architecture directly on the target hardware and dataset without approximation.