AI Paper Atlas

Research map / Architectures and efficiency / Neural architecture search

BossNAS: Exploring Hybrid CNN-Transformers With Block-Wisely Self-Supervised Neural Architecture Search

Changlin Li, Tao Tang, Guangrun Wang, Jiefeng Peng, Bing Wang, Xiaodan Liang, Xiaojun Chang

ICCV 2021 · Architectures and efficiency › Neural architecture search · approach: Bi-level differentiable NAS · 112 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
replace naive unsupervised bootstrapping with ensemble bootstrapping for supernet training (ablation)BossNAS · MBConv search space ratingKendall tau: 0.12 -> 0.65Improved
use unsupervised evaluation metric instead of supervised linear evaluation after ensemble bootstrapping (ablation)BossNAS · MBConv search space ratingKendall tau: 0.55 -> 0.65Improved
use unsupervised ensemble bootstrapping instead of supervised distillation (ablation)BossNAS · MBConv search space ratingKendall tau: 0.62 -> 0.65Improved
use unsupervised ensemble bootstrapping instead of supervised classification training (ablation)BossNAS · MBConv search space ratingKendall tau: 0.46 -> 0.65Improved
use a teacher unlike the candidates (MobileNetV1 instead of EfficientNet-style teacher) in DNADNA · MBConv search space ratingKendall tau: 0.62 -> 0.23Hurt
search with BossNAS instead of supervised DNA-T on HyTraHyTra search · ImageNet val (HyTra search space)top-1 accuracy (%): 80.3 -> 80.8Improved
search with BossNAS instead of random selection in HyTraHyTra search · ImageNet val (HyTra search space)top-1 accuracy (%): 76.7 -> 80.8Improved
search with BossNAS instead of UnNAS in HyTraHyTra search · ImageNet val (HyTra search space)top-1 accuracy (%): 79.8 -> 80.8Improved
replace relative position encoding in BoTBlock with depthwise conv implicit position encoding (BoT50 to BoT50-T)BoT50 · ImageNet val (HyTra search space)top-1 accuracy (%): 78.3 -> 79.5Improved
add SE and SiLU to BossNet-T0 (ablation)BossNet-T0 · ImageNet val (HyTra search space)top-1 accuracy (%): 80.5 -> 80.8Improved
test BossNet-T0 on larger input (288) without finetuningBossNet-T0 · ImageNet val (HyTra search space)top-1 accuracy (%): 80.8 -> 81.6Improved
remove downsampling in last stage of BossNet-T0 (BossNet-T1)BossNet-T0 · ImageNet val (HyTra search space)top-1 accuracy (%): 80.8 -> 82.2Improved

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