Research map / Architectures and efficiency / Neural architecture search
DNA Family: Boosting Weight-Sharing NAS With Block-Wise Supervisions
Guangrun Wang, Changlin Li, Liuchun Yuan, Jiefeng Peng, Xiaoyu Xian, Xiaodan Liang, Xiaojun Chang, Liang Lin
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 |
|---|---|---|---|
| use concurrent block-wise distillation instead of progressive strategy A (ablation) | DNA · ImageNet val (MBConv) | top-1 accuracy (%): 77.0% -> 77.49% | Improved |
| use concurrent block-wise distillation instead of progressive strategy B (ablation) | DNA · ImageNet val (MBConv) | top-1 accuracy (%): 77.15% -> 77.49% | Improved |
| add multi-cell search (3 cells per block) (ablation) | DNA · ImageNet val (MBConv, no constraint) | top-1 accuracy (%): 77.49% -> 77.84% | Improved |
| add multi-cell search (3 cells per block) under 5.3M parameter constraint (ablation) | DNA · ImageNet val (MBConv, constrained) | top-1 accuracy (%): 77.21% -> 77.38% | Improved |
| use EfficientNet-B7 teacher instead of EfficientNet-B0 teacher (ablation) | DNA · ImageNet val (MBConv) | top-1 accuracy (%): 77.8% vs 77.8% | No clear effect |
| distill architecture from EfficientNet-B0 teacher (student vs teacher) (ablation) | DNA-from-B0 · ImageNet val (MBConv) | top-1 accuracy (%): 76.3% -> 77.8% | Improved |
| scale DNA-c to teacher size (DNA-c7 vs EfficientNet-B7) (ablation) | DNA-c · ImageNet val (224x224) | top-1 accuracy (%): 77.8% -> 79.9% | Improved |
| add progressive generations (DNA+ generation 1) | DNA+ · ImageNet val (MBConv) | top-1 accuracy (%): 76.3 -> 78.1 | Improved |
| add progressive generation 2 (DNA+) | DNA+ · ImageNet val (MBConv) | top-1 accuracy (%): 78.1 -> 78.3 | Improved |
| add generation 3 of DNA+ | DNA+ · ImageNet NAS Bench | Kendall tau: 0.66 -> 0.66 | No clear effect |
| replace DNA with self-supervised DNA++ (no KD) | DNA-T · ImageNet val (ViT space) | top-1 accuracy (%): 81.2 -> 82.0 | Improved |
| replace DNA with self-supervised DNA++ (with KD) | DNA-T · ImageNet val (ViT space) | top-1 accuracy (%): 82.9 -> 83.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DARTS — DARTS-: Robustly Stepping out of Performance Collapse Without Indicators (ICLR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Block-Wisely Supervised Neural Architecture Search With Knowledge Distillation (CVPR 2020 · similarity 0.91)
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search (ICLR 2020 · similarity 0.90)
- Understanding Architectures Learnt by Cell-based Neural Architecture Search (ICLR 2020 · similarity 0.90)
- NAS-Bench-101: Towards Reproducible Neural Architecture Search (ICML 2019 · similarity 0.89)
- Towards Neural Architecture Search through Hierarchical Generative Modeling (ICML 2024 · similarity 0.88)
- Efficient Few-Shot Neural Architecture Search by Counting the Number of Nonlinear Functions (AAAI 2025 · similarity 0.88)
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.
