Research map / Architectures and efficiency / CNN design
SkipNet: Learning Dynamic Routing in Convolutional Networks
Xin Wang, Fisher Yu, Zi-Yi Dou, Trevor Darrell, Joseph E. Gonzalez
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 hybrid RL on top of supervised pre-training with recurrent gates (ablation) | SkipNet-110 · CIFAR-10 | computation reduction at matched accuracy: 50% | Improved |
| add hybrid RL on top of supervised pre-training with recurrent gates (ablation) | SkipNet-152 · SVHN | computation reduction at matched accuracy: 86% | Improved |
| add hybrid RL on top of supervised pre-training with recurrent gates (ablation) | SkipNet-101 · ImageNet | computation reduction at matched accuracy: 30% | Improved |
| add supervised pre-training before hybrid RL (ablation) | SkipNet-38+HRL · CIFAR-10 | accuracy | Improved |
| add supervised pre-training before pure RL (ablation) | SkipNet-38 · CIFAR-10 | accuracy: roughly 10% | Improved |
| use hard gating instead of soft gating in pre-training (ablation) | SkipNet-38 · CIFAR-10 | accuracy: 66.67 -> 90.83 | Improved |
| use hard gating instead of soft gating in pre-training (ablation) | SkipNet-74 · CIFAR-10 | accuracy: 52.29 -> 92.38 | Improved |
| use hard gating instead of soft gating in pre-training (ablation) | SkipNet-110 · CIFAR-100 | accuracy: 9.84 -> 63.66 | Improved |
| use recurrent gate instead of feed-forward gates (ablation) | SkipNet · CIFAR-10, CIFAR-100, SVHN | computation reduction | Improved |
| increase cost weight alpha (ablation) | SkipNet-34/50/101 · ImageNet | accuracy vs skipped layers | Mixed |
| replace random block skipping (SDV) with learned skipping policy | SkipNet · CIFAR-10 / CIFAR-100 | accuracy at same FLOPs | Improved |
Closest papers by meaning
- Skip-Convolutions for Efficient Video Processing (CVPR 2021 · similarity 0.90)
- Skip-Attention: Improving Vision Transformers by Paying Less Attention (ICLR 2024 · similarity 0.87)
- Beyond Skip Connection: Pooling and Unpooling Design for Elimination Singularities (AAAI 2025 · similarity 0.87)
- Adaptive Depth Networks with Skippable Sub-Paths (NeurIPS 2024 · similarity 0.86)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.85)
- The Surprising Effectiveness of Skip-Tuning in Diffusion Sampling (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.
