Research map / Detection and segmentation / CNN features
Efficiency Follows Global-Local Decoupling
Zhenyu Yang, Gensheng Pei, Tao Chen, Yichao Zhou, Tianfei Zhou, Yazhou Yao, Fumin Shen
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 |
|---|---|---|---|
| add per-stage neural memory global branch to local-only (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 78.2 -> 80.0 | Improved |
| apply memory at every block instead of per stage (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 80.0 -> 80.4 | No clear effect |
| replace neural memory with CBAM (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 80.0 -> 78.4 | Hurt |
| replace neural memory with Non-local (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 80.0 -> 79.1 | Hurt |
| replace neural memory with self-attention (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 80.0 -> 79.8 | Hurt |
| replace neural memory with IR-RWKV (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 80.0 -> 77.6 | Hurt |
| replace neural memory with Titans-style memory (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 80.0 -> 79.6 | Hurt |
| use gating instead of additive fusion (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 79.6 -> 80.0 | Improved |
| use gating instead of concatenation (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 79.5 -> 80.0 | Improved |
| raise memory dim from 64 to 128 (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 78.6 -> 80.0 | Improved |
| raise memory dim from 128 to 256 (ablation) | ConvNeur-M3 · ImageNet-1K | top-1 accuracy: 80.0 -> 80.3 | No clear effect |
Closest papers by meaning
- Register and [CLS] tokens induce a decoupling of local and global features in large ViTs (NeurIPS 2025 · similarity 0.88)
- OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels (CVPR 2025 · similarity 0.87)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.87)
- Recognize Any Regions (NeurIPS 2024 · similarity 0.86)
- From Local Details to Global Context: Advancing Vision-Language Models with Attention-Based Selection (ICML 2025 · similarity 0.85)
- Role of Locality and Weight Sharing in Image-Based Tasks: A Sample Complexity Separation between CNNs, LCNs, and FCNs (ICLR 2024 · similarity 0.85)
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.
