Research map / Architectures and efficiency / Pruning and efficient tuning
Stimulative Training++: Go Beyond the Performance Limits of Residual Networks
Peng Ye, Tong He, Shengji Tang, Baopu Li, Tao Chen, Lei Bai, Wanli Ouyang
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 stimulative training (subnet sampling + KL) (ablation) | ResNet50 common training · ImageNet-1k | top-1 accuracy (%): 76.70 -> 77.56 | Improved |
| add KL- loss with 6 sampled subnets (ablation) | ResNet50 ST · ImageNet-1k | top-1 accuracy (%): 77.56 -> 78.53 | Improved |
| add random smaller subnet inputs (SITrans) (ablation) | ResNet50 ST · ImageNet-1k | top-1 accuracy (%): 77.56 -> 78.56 | Improved |
| add SITrans to ST with KL- and Snet6 (ablation) | ResNet50 ST · ImageNet-1k | top-1 accuracy (%): 78.53 -> 79.83 | Improved |
| adjust val resize 240 and train crop 200 (ablation) | ResNet50 ST++ · ImageNet-1k | top-1 accuracy (%): 79.83 -> 80.04 | Improved |
| add inter-stage sampling rule (InterSSR) (ablation) | ResNet50 ST++ · ImageNet-1k | top-1 accuracy (%): 80.04 -> 80.50 | Improved |
| add uniform augment and extend to 400 epochs (ablation) | ResNet50 ST++ (A1) · ImageNet-1k | top-1 accuracy (%): 80.50 -> 80.97 | Improved |
| use larger subnet inputs (up to 272 instead of 224) (ablation) | ResNet50 ST · ImageNet-1k | top-1 accuracy (%): 78.562 -> 78.058 | Hurt |
| apply jitter and flip transforms to subnet inputs instead of downscaling (ablation) | ResNet50 ST · ImageNet-1k | top-1 accuracy (%): 77.564 -> 76.566 | Hurt |
| sampling space [1,2,4,1] instead of [3,4,6,3] (ablation) | ResNet50 ST++ · ImageNet-1k | top-1 accuracy (%): 80.038 -> 80.496 | Improved |
| ordered residual sampling instead of stochastic residual sampling (ablation) | MobileNetV3 · CIFAR-100 | accuracy: +0.58% | Improved |
| finetune with subnet sampling (ST*) instead of plain finetuning of ST++ weights | Faster R-CNN R50-ST · COCO (AdamW) | Bbox mAP: 40.0 -> 41.9 | Improved |
Closest papers by meaning
- Stimulative Training of Residual Networks: A Social Psychology Perspective of Loafing (NeurIPS 2022 · similarity 0.97)
- Boosting Residual Networks with Group Knowledge (AAAI 2024 · similarity 0.89)
- Residual Alignment: Uncovering the Mechanisms of Residual Networks (NeurIPS 2023 · similarity 0.87)
- Residual Networks Behave Like Ensembles of Relatively Shallow Networks (NeurIPS 2016 · similarity 0.86)
- A Mean Field Analysis Of Deep ResNet And Beyond: Towards Provably Optimization Via Overparameterization From Depth (ICML 2020 · similarity 0.86)
- Efficient stagewise pretraining via progressive subnetworks (ICLR 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.
