Research map / Architectures and efficiency / Attention and transformers
Revisiting Vision Transformer from the View of Path Ensemble
Shuning Chang, Pichao Wang, Hao Luo, Fan Wang, Mike Zheng Shou
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 |
|---|---|---|---|
| prune short paths (keep p7-p12) in the ensemble prediction (ablation) | DeiT-S · ImageNet-1K val | top-1 accuracy (%): 79.8 -> 80.2 | Improved |
| add EnsembleScale to re-weight paths (ablation) | DeiT-S · ImageNet-1K val | top-1 accuracy (%): 79.8 -> 80.3 | Improved |
| add EnsembleScale to 18-layer DeiT-S (ablation) | DeiT-S (18 layers) · ImageNet-1K val | top-1 accuracy (%): 80.1 -> 81.2 | Improved |
| add EnsembleScale to 24-layer DeiT-S (ablation) | DeiT-S (24 layers) · ImageNet-1K val | top-1 accuracy (%): 78.9 -> 81.6 | Improved |
| apply EnsembleScale and path pruning to Swin-T (ablation) | Swin-T · ImageNet-1K val | top-1 accuracy (%): 81.3 -> 81.5 | Improved |
| add dynamic early exit using first seven paths | DeiT-S · ImageNet-1K val | FLOPs and accuracy: 4.58G -> 3.42G, accuracy 79.8 -> 79.8 | Improved |
| add EnsembleScale before fine-tuning on CIFAR (ablation) | DeiT-S · CIFAR-100 | accuracy (%): 85.9 -> 87.0 | Improved |
| add prediction-logit then hidden-state self-distillation between paths (ablation) | DeiT-S · ImageNet-1K val | top-1 accuracy (%): 79.8 -> 80.2 -> 80.6 | Improved |
| combine self-distillation with path combination (ablation) | DeiT-S · ImageNet-1K val | top-1 accuracy (%): 80.6 -> 81.0 | Improved |
| add hidden-state distillation and path combination to Swin-B prediction-logit distillation (ablation) | Swin-B · ImageNet-1K val | top-1 accuracy (%): 84.3 -> 84.2 | Hurt |
| use a larger teacher-student distance (Delta=4 vs 2) (ablation) | self-distillation on DeiT-S · ImageNet-1K val | top-1 accuracy (%): 81.0 -> 80.4 | Hurt |
| include very short paths in distillation (start from p0 instead of p5) (ablation) | self-distillation on DeiT-S · ImageNet-1K val | top-1 accuracy (%): 81.0 -> 80.6 | Hurt |
Closest papers by meaning
- A Simple Single-Scale Vision Transformer for Object Detection and Instance Segmentation (ECCV 2022 · similarity 0.88)
- Pruning Self-Attentions Into Convolutional Layers in Single Path (TPAMI 2024 · similarity 0.88)
- Towards Robust Vision Transformers: Path Dependency Analysis and a Simple Two-Stage Adversarial Training (CVPR 2026 · similarity 0.87)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.86)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.86)
- Slicing Vision Transformer for Flexible Inference (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.
