Research map / Architectures and efficiency / CNN design
Deep Feature Surgery: Towards Accurate and Efficient Multi-Exit Networks
Cheng Gong, Yao Chen, Qiuyang Luo, Ye Lu, Tao Li, Yuzhi Zhang, Yufei Sun, Le Zhang
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 |
|---|---|---|---|
| apply Deep Feature Surgery to multi-exit ResNet18 | ResNet18 multi-exit · CIFAR-100 Exit1 | top-1 accuracy: 67.85 -> 74.17 | Improved |
| apply Deep Feature Surgery to multi-exit ResNet18 ensemble | ResNet18 multi-exit · CIFAR-100 Ensemble | top-1 accuracy: 79.67 -> 80.89 | Improved |
| apply Deep Feature Surgery to MSDNet | MSDNet · CIFAR-100 all exits | top-1 accuracy: 72.46 vs 68.78 (Exit1); 79.56 vs 77.47 (Exit7) | Improved |
| apply Deep Feature Surgery to MSDNet on ImageNet | MSDNet · ImageNet average of 5 exits | top-1 accuracy: 67.58 -> 69.05 | Improved |
| DFS versus Mate-GF at the deepest ImageNet exit | MSDNet · ImageNet Exit5 | top-1 accuracy: 73.27 vs 72.88 | Hurt |
| lower partition coefficient beta (more exit-specific features) (ablation) | DFS on VGG7-64 · CIFAR-100 early exits | top-1 accuracy: 17.88 -> 22.81 (Exit1, beta 0.9 -> 0.1) | Improved |
| higher partition coefficient beta (more shared features) (ablation) | DFS on VGG7-64 · CIFAR-100 late exits | top-1 accuracy: 65.46 -> 73.06 (Exit6, beta 0.1 -> 0.9) | Improved |
| feature partitioning reduces backward operations in training | DFS · VGG7-64 CPU training | training FPS: up to +14.43% vs DSN, +50.00% vs BYOT | Improved |
| DFS training on GPU with VGG16 | DFS · VGG16 GPU training | training FPS: up to +14.99% vs DSN, +22.23% vs BYOT | Improved |
| DFS in budgeted batch classification versus Meta-GF | MSDNet · CIFAR-100 budgeted batch | FLOPs for same accuracy: about 2x fewer FLOPs | Improved |
Closest papers by meaning
- Multi-Exit Semantic Segmentation Networks (ECCV 2022 · similarity 0.84)
- BEEM: Boosting Performance of Early Exit DNNs using Multi-Exit Classifiers as Experts (ICLR 2025 · similarity 0.84)
- How to Train Your Multi-Exit Model? Analyzing the Impact of Training Strategies (ICML 2025 · similarity 0.83)
- Beyond Skip Connection: Pooling and Unpooling Design for Elimination Singularities (AAAI 2025 · similarity 0.82)
- ZipIt! Merging Models from Different Tasks without Training (ICLR 2024 · similarity 0.82)
- B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable (NeurIPS 2024 · similarity 0.82)
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.
