Reproduce “EH-DNAS: End-to-End Hardware-aware Differentiable Neural Architecture Search” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qian Jiang, Xiaofan Zhang, Deming Chen, Minh N. Do, Raymond A. Yeh
EH-DNAS incorporates a learned, differentiable predictor trained on end-to-end hardware performance data into neural architecture search, enabling architectures to be optimized for both task accuracy and target-device efficiency. Across evaluated accelerators and processors, it generally reduces latency relative to comparison methods; E2E-Perf estimates customized-accelerator performance with average errors of 1.15% and 2.17% for two accelerator paradigms.
Key insights
- Optimizing with hardware feedback can find architectures that trade accuracy and latency more effectively than search guided by FLOPs or lookup-table estimates alone.
- Predictor quality is platform-dependent: errors were low on the tested Edge GPU and Edge TPU, but substantially higher on the Raspberry Pi 4 and Pixel 3 CPUs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.