Research map / Architectures and efficiency / Attention and transformers
Cumulative Spatial Knowledge Distillation for Vision Transformers
Borui Zhao, Renjie Song, Jiajun Liang
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 CSKD to DeiT distillation | DeiT-Ti · ImageNet-1k val | top-1 accuracy: 74.5 -> 76.3 | Improved |
| add CSKD to DeiT distillation | DeiT-B · ImageNet-1k val | top-1 accuracy: 83.4 -> 83.8 | Improved |
| train CSKD-Ti for 1000 epochs | DeiT-Ti · ImageNet-1k val (1000 epochs) | top-1 accuracy: 76.6 -> 78.1 | Improved |
| add CSKD to DeiT for transfer learning | DeiT-Ti · iNat19 | top-1 accuracy: 76.6 -> 77.7 | Improved |
| distill local responses from the CNN (ablation) | DeiT baseline · ImageNet-1k val | top-1 accuracy: 81.2 -> 81.6 | Improved |
| combine local and global responses by simple summation (ablation) | CSKD · ImageNet-1k val | top-1 accuracy: 81.6 -> 81.5 | No clear effect |
| add cumulative knowledge fusion (CKF) (ablation) | CSKD · ImageNet-1k val | top-1 accuracy: 81.5 -> 82.3 | Improved |
| use soft instead of hard distillation (ablation) | CSKD · ImageNet-1k val | top-1 accuracy: 82.3 -> 82.1 | No clear effect |
| change alpha decay strategy in CKF (ablation) | CSKD · ImageNet-1k val | top-1 accuracy: 82.3 vs 82.2 | No clear effect |
| vary CSKD loss weight between 0.5 and 2.0 (ablation) | CSKD · ImageNet-1k val | top-1 accuracy: 82.1, 82.3, 82.1 | No clear effect |
Closest papers by meaning
- DearKD: Data-Efficient Early Knowledge Distillation for Vision Transformers (CVPR 2022 · similarity 0.90)
- Fuse Before Transfer: Knowledge Fusion for Heterogeneous Distillation (ICCV 2025 · similarity 0.89)
- Vision Transformers provably learn spatial structure (NeurIPS 2022 · similarity 0.89)
- A Good Student is Cooperative and Reliable: CNN-Transformer Collaborative Learning for Semantic Segmentation (ICCV 2023 · similarity 0.89)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.89)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.87)
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.
