AI Paper Atlas

Research map / Architectures and efficiency / Neural architecture search

Powering One-shot Topological NAS with Stabilized Share-parameter Proxy

Ronghao Guo, Chen Lin, Chuming Li, Keyu Tian, Ming Sun, Lu Sheng, Junjie Yan

ECCV 2020 · Architectures and efficiency › Neural architecture search · approach: One-shot and few-shot supernets · 19 citations (OpenAlex)

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 changedWhereMeasuredEffect
add 4 random skip residual blocks to ResNet-18 at equal FLOPs (topology complexity)ResNet-18 · ImageNet valtop-1 accuracy (%): 70.2 -> 71.5 / 72.0 / 69.6Mixed
estimate architecture score by averaging SGLD-sampled shared parameters instead of a single checkpoint (ablation)ST-NAS hyper-network · Ranking of 12 architecturesKendall tau: 0.71 -> 0.84Improved
estimate architecture score by averaging accuracy over SGLD-sampled shared parameters (ablation)ST-NAS hyper-network · Ranking of 12 architecturesKendall tau: 0.71 -> 0.81Improved
add independent fine-tuning of each architecture from hyper-network weights (ablation)ST-NAS hyper-network · Ranking of 12 architecturesKendall tau: 0.71 -> 0.64Hurt
parameter expectation versus accuracy expectation (ablation)SGLD sampling · Ranking of 12 architecturesKendall tau: 0.84 vs 0.81No clear effect
use ST-NAS-A topology-augmented searched backbone instead of MobileNetV2 at equal MAddsFPN detector · COCOmAP: 31.7 -> 33.2Improved
use ST-NAS-B backbone instead of ResNet18FPN detector · COCOmAP: 32.2 -> 35.3Improved
use channel-scaled ST-NAS-B* backbone instead of ResNet50 with a quarter of the MAddsFPN detector · COCOmAP: 36.9 -> 37.7Improved
search in topology augmented space instead of chain-structured space (ST-NAS-A vs SPOS)one-shot NAS · ImageNet valtop-1 accuracy (%): 74.7 -> 76.4Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy