Research map / Architectures and efficiency / RNNs and normalization
Learning to Weight Samples for Dynamic Early-Exiting Networks
Yizeng Han, Yifan Pu, Zihang Lai, Chaofei Wang, Shiji Song, Junfeng Cao, Wenhui Huang, Chao Deng, Gao Huang
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 learned sample weighting via weight prediction network | 7-exit MSDNet · CIFAR-100 anytime prediction, exit 1 | accuracy: 61.07 -> 62.47 | Improved |
| add learned sample weighting | 7-exit MSDNet · CIFAR-100 anytime prediction, exit 7 | accuracy: 74.50 -> 75.81 | Improved |
| add learned sample weighting | 6-exit RANet · CIFAR-100 anytime prediction, exit 1 | accuracy: 63.41 -> 65.33 | Improved |
| add learned sample weighting | 6-exit RANet · CIFAR-100 anytime prediction, exit 6 | accuracy: 71.78 -> 72.45 | Improved |
| add sample weighting on top of IMTA | MSDNet + IMTA · CIFAR-100 anytime prediction, exit 7 | accuracy: 75.19 -> 76.31 | Improved |
| add sample weighting to 5-exit MSDNet | 5-exit MSDNet · CIFAR-100 dynamic inference at ~15M Mul-Adds | top-1 accuracy: increased by about 1.3% | Improved |
| add sample weighting to RANet | RANet · CIFAR-100 at ~15M Mul-Adds | top-1 accuracy: ~1.4% | Improved |
| add sample weighting to MSDNet | MSDNet · ImageNet latency | computation time to reach 75.4% accuracy: 70% of original time (1.4x speed-up) | Improved |
| replace learned weighting by vanilla selection (loss only on allocated training samples) (ablation) | 5-exit MSDNet · CIFAR-100 | accuracy: drastic drop in accuracy | Hurt |
| replace learned weights with hand-designed fixed weights 0.6 to 1.4 (ablation) | 5-exit MSDNet · CIFAR-100 | accuracy: gain limited | No clear effect |
| freeze WPN instead of joint optimization (ablation) | 5-exit MSDNet · CIFAR-100 | accuracy: joint optimization outperforms frozen WPN | Hurt |
| use meta objective on samples that actually exit instead of whole meta set (ablation) | 5-exit MSDNet · CIFAR-100 | accuracy: higher accuracy at larger budgets | Improved |
Closest papers by meaning
- Jointly-Learned Exit and Inference for a Dynamic Neural Network (ICLR 2024 · similarity 0.88)
- How to Train Your Multi-Exit Model? Analyzing the Impact of Training Strategies (ICML 2025 · similarity 0.88)
- Initializing Models with Larger Ones (ICLR 2024 · similarity 0.87)
- Beyond Greedy Exits: Improved Early Exit Decisions for Risk Control and Reliability (NeurIPS 2025 · similarity 0.86)
- BEEM: Boosting Performance of Early Exit DNNs using Multi-Exit Classifiers as Experts (ICLR 2025 · similarity 0.86)
- Why Do We Need Weight Decay in Modern Deep Learning? (NeurIPS 2024 · similarity 0.86)
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.
