Research map / Architectures and efficiency / Neural architecture search
BRP-NAS: Prediction-based NAS using GCNs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use BRP-NAS instead of prior search methods | NAS-Bench-101 search on CIFAR-10 · NAS-Bench-101 | trained models to reach reported test accuracy: 140 trained models; 94.22 test accuracy | Helped |
| use BRP-NAS rather than DARTS | CIFAR-10 architecture search in the DARTS space · DARTS search space, small budget | test error: 2.76±0.09 -> 2.71±0.07 | Helped |
| increase BRP-NAS search budget from the medium to high budget | CIFAR-10 architecture search in the DARTS space · DARTS search space | test error: 2.66±0.09 -> 2.59±0.11 | Helped |
| replace the layer-wise latency predictor with an end-to-end GCN predictor | NAS-Bench-201 latency prediction · desktop CPU | latency predictions within ±5% of measured latency: 18.2±0.4 -> 85.2±1.8 | Helped |
| replace the layer-wise latency predictor with an end-to-end GCN predictor | NAS-Bench-201 latency prediction · desktop GPU | latency predictions within ±5% of measured latency: 17.1±0.3 -> 85.9±1.9 | Helped |
| replace the layer-wise latency predictor with an end-to-end GCN predictor | NAS-Bench-201 latency prediction · embedded GPU | latency predictions within ±5% of measured latency: 29.7±0.8 -> 82.5±1.5 | Helped |
| initialize the accuracy GCN from a pretrained desktop GPU latency predictor (ablation) | NAS-Bench-201 accuracy prediction with 100 samples · NAS-Bench-201 | predictions within ±5% of actual accuracy: 76.9±2.4 -> 80.5±2.2 | Helped |
| initialize the accuracy GCN from a pretrained FLOPS predictor (ablation) | NAS-Bench-201 accuracy prediction with 100 samples · NAS-Bench-201 | predictions within ±5% of actual accuracy: 76.9±2.4 -> 80.5±2.7 | Helped |
| use binary pairwise accuracy comparisons instead of predicting absolute accuracy (ablation) | BRP-NAS accuracy predictor · NAS-Bench-201 | ranking correlation | Helped |
| use iterative data selection focused on high-ranked candidates (ablation) | BRP-NAS binary relation predictor · NAS-Bench-201 | NAS search performance | Helped |
| use iterative data selection rather than no iterative selection (ablation) | BRP-NAS binary relation predictor · NAS-Bench-201 | global ranking correlation | Hurt |
| reduce latency predictor training data from 900 models to 100 models (ablation) | GCN latency predictor · NAS-Bench-201 | latency prediction accuracy | Hurt |
Closest papers by meaning
- Hardware-adaptive Efficient Latency Prediction for NAS via Meta-Learning (NeurIPS 2021 · similarity 0.90)
- HW-NAS-Bench: Hardware-Aware Neural Architecture Search Benchmark (ICLR 2021 · similarity 0.90)
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search (ICLR 2020 · similarity 0.89)
- FP-NAS: Fast Probabilistic Neural Architecture Search (CVPR 2021 · similarity 0.89)
- NAS-Bench-101: Towards Reproducible Neural Architecture Search (ICML 2019 · similarity 0.88)
- HEP-NAS: Towards Efficient Few-shot Neural Architecture Search via Hierarchical Edge Partitioning (AAAI 2025 · similarity 0.87)
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.