Research map / Architectures and efficiency / CNN design
Aggregated Residual Transformations for Deep Neural Networks
Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, Kaiming He
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 |
|---|---|---|---|
| replace the ResNet-50 bottleneck with aggregated transformations at cardinality 32, keeping complexity similar (ablation) | ResNet-50 · ImageNet-1K val | top-1 error: 23.9 -> 22.2 | Improved |
| increase cardinality while reducing bottleneck width to preserve complexity (ablation) | ResNet-50 · ImageNet-1K val | top-1 error: 23.9 -> 23.0 -> 22.6 -> 22.3 -> 22.2 | Improved |
| increase cardinality while reducing bottleneck width to preserve complexity (ablation) | ResNet-101 · ImageNet-1K val | top-1 error: 22.0 -> 21.7 -> 21.4 -> 21.3 -> 21.2 | Improved |
| increase cardinality to 64 at approximately 2× complexity (ablation) | ResNeXt-101 (32×4d) · ImageNet-1K val | top-1 error: 21.2 -> 20.4 | Improved |
| spend approximately 2× complexity on increasing cardinality rather than depth or width (ablation) | ResNet-101 · ImageNet-1K val | top-1 error: ResNet-200: 21.7; wider ResNet-101: 21.3; ResNeXt-101 (2×64d): 20.7 | Improved |
| remove residual shortcuts (ablation) | ResNeXt-50 (32×4d) · ImageNet-1K val | top-1 error: 22.2 -> 26.1 | Hurt |
| remove residual shortcuts (ablation) | ResNet-50 · ImageNet-1K val | top-1 error: 23.9 -> 31.2 | Hurt |
| replace ResNet with ResNeXt-50 (32×4d) at similar complexity | ResNet-50 · ImageNet-5K train, 5K-way evaluation | top-1 error: 45.5 -> 42.3 | Improved |
| replace ResNet with ResNeXt-101 (32×4d) at similar complexity | ResNet-101 · ImageNet-5K train, 5K-way evaluation | top-1 error: 42.4 -> 40.1 | Improved |
| use cardinality 16 rather than a single wide residual path | CIFAR-10 and CIFAR-100 models · CIFAR test sets | test error: CIFAR-10: 4.17 -> 3.58; CIFAR-100: 20.50 -> 17.31 | Improved |
| replace ResNet-50 with ResNeXt-50 (32×4d) at similar complexity | Faster R-CNN with ResNet-50 · COCO minival | AP: 26.5 -> 27.5 | Improved |
| replace ResNet-101 with ResNeXt-101 (32×4d) at similar complexity | Faster R-CNN with ResNet-101 · COCO minival | AP: 29.8 -> 30.0 | Improved |
Closest papers by meaning
- Residual Alignment: Uncovering the Mechanisms of Residual Networks (NeurIPS 2023 · similarity 0.87)
- Deep Residual Learning for Image Recognition (CVPR 2016 · similarity 0.87)
- ResMLP: Feedforward Networks for Image Classification With Data-Efficient Training (TPAMI 2022 · similarity 0.87)
- Scaling White-Box Transformers for Vision (NeurIPS 2024 · similarity 0.86)
- Revisiting ResNets: Improved Training and Scaling Strategies (NeurIPS 2021 · similarity 0.86)
- LAuReL: Learned Augmented Residual Layer (ICML 2025 · 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.
