Research map / Architectures and efficiency / RNNs and normalization
SP-Net: Slowly Progressing Dynamic Inference Networks
Huanyu Wang, Wenhu Zhang, Shihao Su, Hui Wang, Zhenwei Miao, Xin Zhan, Xi Li
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 the dynamic auxiliary module (momentum teacher with path and logits distillation) (ablation) | vanilla dynamic inference, ResNet-110 · CIFAR-10 | top-1 accuracy (%): 93.83 -> 94.40 | Improved |
| add the dynamic auxiliary module (ablation) | vanilla dynamic inference, ResNet-110 · CIFAR-100 | top-1 accuracy (%): 70.17 -> 73.63 | Improved |
| add multi-stage classification (ablation) | SP-Net with auxiliary module, ResNet-110 · CIFAR-10 | top-1 accuracy (%): 94.40 -> 95.04 | Improved |
| add multi-stage classification (ablation) | SP-Net with auxiliary module, ResNet-110 · CIFAR-100 | top-1 accuracy (%): 73.63 -> 74.91 | Improved |
| add feature evolution direction regularization (ablation) | SP-Net A+B, ResNet-110 · CIFAR-10 | top-1 accuracy (%): 95.04 -> 95.22 | Improved |
| add feature evolution direction regularization (ablation) | SP-Net A+B, ResNet-110 · CIFAR-100 | top-1 accuracy (%): 74.91 -> 75.14 | Improved |
| raise momentum coefficient from 0.7 to 0.8 (ablation) | SP-Net, ResNet-110 · CIFAR-10 | top-1 accuracy (%): 95.22 -> 94.82 | Hurt |
| lower momentum coefficient from 0.7 to 0.5 (ablation) | SP-Net, ResNet-110 · CIFAR-10 | top-1 accuracy (%): 95.22 -> 94.97 | Hurt |
| regularize feature evolution across different resolutions instead of all adjacent blocks (ablation) | SP-Net, ResNet-110 · CIFAR-10 | top-1 accuracy (%): 94.43 -> 95.22 | Improved |
| regularize feature evolution across resolutions instead of adjacent blocks in same resolution (ablation) | SP-Net, ResNet-110 · CIFAR-100 | top-1 accuracy (%): 74.05 -> 74.86 | Improved |
| use SP-Net instead of vanilla dynamic baseline at target executing rate 0.2 | dynamic ResNet-110 · CIFAR-10 | top-1 accuracy (%): 90.48 with 0.11 GMACCs -> 93.75 at the computational cost of 0.14 GMACCs | Improved |
| use SP-Net instead of vanilla dynamic baseline at target executing rate 0.4 | dynamic ResNet-110 · CIFAR-10 | top-1 accuracy (%): 93.52 -> 94.47 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RDI-Net — RDI-Net: Relational Dynamic Inference Networks (ICCV 2021) · better on 2 of 2 shared results
- CoDiNet — CoDiNet: Path Distribution Modeling With Consistency and Diversity for Dynamic Routing (TPAMI 2021) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- RDI-Net: Relational Dynamic Inference Networks (ICCV 2021 · similarity 0.89)
- SplineNets: Continuous Neural Decision Graphs (NeurIPS 2018 · similarity 0.85)
- DISC: Dynamic Decomposition Improves LLM Inference Scaling (NeurIPS 2025 · similarity 0.85)
- Dr.LLM: Dynamic Layer Routing in LLMs (ICLR 2026 · similarity 0.84)
- LightHGNN: Distilling Hypergraph Neural Networks into MLPs for 100x Faster Inference (ICLR 2024 · similarity 0.84)
- Scaling Inference-Efficient Language Models (ICML 2025 · 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.
