Research map / Architectures and efficiency / Neural architecture search
Neural Architecture Search for Lightweight Non-Local Networks
Yingwei Li, Xiaojie Jin, Jieru Mei, Xiaochen Lian, Linjie Yang, Cihang Xie, Qihang Yu, Yuyin Zhou, Song Bai, Alan L. Yuille
Check 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 the proposed LightNL blocks to MobileNetV2 at the reported manual configuration | MobileNetV2 · ImageNet | Top-1 accuracy: 73.4% to 75.0% | Improved |
| Add LightNL blocks | MobileNetV2 backbone in Deeplabv3 · PASCAL VOC 2012 | mIoU: 70.6 to 72.9 | Improved |
| Replace the conventional non-local block with LightNL (ablation) | MobileNetV2 · ImageNet | Top-1 accuracy: 75.2% vs. 75.0% | No clear effect |
| Use the full LightNL design instead of conventional non-local computation (ablation) | MobileNetV2 · ImageNet | FLOPs increase: +6.2G vs. +15M | Improved |
| Search the placement and compactness of LightNL blocks with AutoNL | MobileNetV2-family mobile architecture · ImageNet 2012 Val | Top-1 accuracy: 76.3 to 77.7 | Improved |
| Use AutoNL-L rather than MobileNetV3-L (1.25x) | MobileNetV3-L (1.25x) · ImageNet 2012 Val | Top-1 accuracy: 76.6 to 77.7 | Improved |
| Use AutoNL-S rather than MobileNetV2 | MobileNetV2 · ImageNet 2012 Val | Top-1 accuracy: 72.0 to 76.5 | Improved |
| Share and lighten feature transformations, remove the separate g transform, and use depthwise convolution for the residual transform (ablation) | LightNL block · ImageNet | Top-1 accuracy: 75.2, 75.1, 75.0, 75.0 | No clear effect |
| Use compact features to compute affinity matrices (ablation) | LightNL block · ImageNet | FLOPs: +119M, +93M, +66M, +38M, +15M | Improved |
Closest papers by meaning
- MnasNet: Platform-Aware Neural Architecture Search for Mobile (CVPR 2019 · similarity 0.88)
- MobileODE: An Extra Lightweight Network (NeurIPS 2025 · similarity 0.87)
- Rethinking Mobile Block for Efficient Attention-based Models (ICCV 2023 · similarity 0.86)
- AutoViT: Achieving Real-Time Vision Transformers on Mobile via Latency-aware Coarse-to-Fine Search (IJCV 2025 · similarity 0.86)
- MobileNetV4: Universal Models for the Mobile Ecosystem (ECCV 2024 · similarity 0.85)
- RepViT: Revisiting Mobile CNN From ViT Perspective (CVPR 2024 · similarity 0.84)
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.
