Research map / Architectures and efficiency / CNN design
WeightNet: Revisiting the Design Space of Weight Networks
Ningning Ma, Xiangyu Zhang, Jiawei Huang, Jian Sun
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 WeightNet (1x) | ShuffleNetV2 0.5x · ImageNet val | Top-1 error: 39.7 -> 36.7 | Improved |
| add WeightNet (1x) | ShuffleNetV2 1x · ImageNet val | Top-1 error: 30.9 -> 28.8 | Improved |
| raise lambda from 1 to 8 (ablation) | WeightNet on ShuffleNetV2 0.5x · ImageNet val | Top-1 error: 36.7 -> 34.0 | Improved |
| raise lambda from 4 to 8 (ablation) | WeightNet on ShuffleNetV2 2x · ImageNet val | Top-1 error: 23.5 -> 24.0 | Hurt |
| raise G from 1 to 4 (ablation) | WeightNet on ShuffleNetV2 0.5x · ImageNet val | Top-1 error: 37.18 -> 36.37 | Improved |
| apply WeightNet to stages 3 and 4 only instead of all stages (ablation) | WeightNet on ShuffleNetV2 0.5x · ImageNet val | Top-1 error: 35.47 -> 35.30 | Improved |
| add WeightNet to stage 4 only instead of stage 2 only (ablation) | WeightNet on ShuffleNetV2 0.5x · ImageNet val | Top-1 error: 39.13 -> 36.43 | Improved |
| use layer-wise GAP instead of stage-wise GAP (ablation) | WeightNet (2x) on ShuffleNetV2 0.5x · ImageNet val | Top-1 error: 37.01 -> 35.04 | Improved |
| replace SE with WeightNet (2x) | ShuffleNetV2 1.5x · ImageNet val | Top-1 error: 26.4 -> 25.2 | Improved |
| replace CondConv (4x) with WeightNet (2x) | ShuffleNetV2 1.5x · ImageNet val | Top-1 error: 26.1 -> 25.2 | Improved |
| add WeightNet (1x) instead of SE | ResNet50 · ImageNet val | Top-1 error: 22.8 -> 22.5 | Improved |
| add WeightNet (4x) as backbone module | RetinaNet ShuffleNetV2 0.5x · COCO minival | mAP: 22.5 -> 27.1 | Improved |
Closest papers by meaning
- Weight Excitation: Built-in Attention Mechanisms in Convolutional Neural Networks (ECCV 2020 · similarity 0.91)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.87)
- ScaleNet - Improve CNNs through Recursively Rescaling Objects (AAAI 2020 · similarity 0.87)
- Generative Modeling of Weights: Generalization or Memorization? (CVPR 2026 · similarity 0.87)
- Learning to Auto Weight: Entirely Data-Driven and Highly Efficient Weighting Framework (AAAI 2020 · similarity 0.87)
- Towards Scalable and Versatile Weight Space Learning (ICML 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.
