Research map / Architectures and efficiency / Neural architecture search
Neural Predictor for Neural Architecture Search
Wei Wen, Hanxiao Liu, Yiran Chen, Hai Li, Gabriel Bender, Pieter-Jan Kindermans
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a filtering classifier stage (two-stage predictor) (ablation) | Neural Predictor · NASBench-101 cross-validation | validation accuracy MSE: 1.95 -> 0.66 | Improved |
| replace random search with Neural Predictor | Neural Predictor · NASBench-101 | sample efficiency: 22.83 times faster to same test accuracy | Improved |
| replace Regularized Evolution with Neural Predictor to reach same validation accuracy | Neural Predictor · NASBench-101 | sample efficiency: 12.40 x faster | Improved |
| use Regularized Evolution instead of random search | Regularized Evolution · NASBench-101 2000 models | mean test accuracy: 93.66 -> 93.97 | Improved |
| use very small N (43 or 86) for predictor training (ablation) | Neural Predictor · NASBench-101 N vs K | search quality | Hurt |
| use large N=860 with small K at fixed budget (ablation) | Neural Predictor · NASBench-101 N vs K | search quality | Hurt |
| replace GCN with MLP predictor (ablation) | Neural Predictor · NASBench-101 | search efficiency | Hurt |
| replace GCN with CNN predictor (ablation) | Neural Predictor · NASBench-101 | search efficiency: near random search performance | Hurt |
| compare Neural Predictor against ProxylessNAS | Neural Predictor · ImageNet ProxylessNAS space | Pareto frontier test accuracy | No clear effect |
| use Neural Predictor instead of random baseline with same 256 models | Neural Predictor · ImageNet ProxylessNAS space | Pareto frontier test accuracy | Improved |
| use Neural Predictor instead of a single ProxylessNAS search | Neural Predictor · ImageNet compute cost | compute: 47.5 times as much compute | Hurt |
| increase K at larger budget on NASBench-101 (ablation) | Neural Predictor · NASBench-101 | test accuracy | Mixed |
Closest papers by meaning
- Neural Architecture Search via Proxy Validation (TPAMI 2022 · similarity 0.91)
- Speedy Performance Estimation for Neural Architecture Search (NeurIPS 2021 · similarity 0.91)
- Contrastive Neural Architecture Search With Neural Architecture Comparators (CVPR 2021 · similarity 0.91)
- Evaluating The Search Phase of Neural Architecture Search (ICLR 2020 · similarity 0.90)
- How Powerful are Performance Predictors in Neural Architecture Search? (NeurIPS 2021 · similarity 0.90)
- Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective (ICLR 2021 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
