Research map / Architectures and efficiency / Attention and transformers
Stratified Knowledge-Density Super-Network for Scalable Vision Transformers
Longhua Li, Lei Qi, Xin Geng
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 |
|---|---|---|---|
| Use WPAC instead of Taylor FO criterion for attention pruning | DeiT-S · ImageNet-1k, 2:4 size, no fine-tuning | top-1 accuracy: 39.4 -> 72.9 | Improved |
| Use WPAC instead of Taylor FO criterion for attention pruning | DeiT-B · ImageNet-1k, 1:4 size, no fine-tuning | top-1 accuracy: 6.0 -> 41.8 | Improved |
| Add importance weighting to PCA over all tokens (versus uniform weight 0.01) (ablation) | WPAC DeiT-B · ImageNet-1k, 50% dims | top-1 accuracy: 76.1 -> 76.9 | Improved |
| Use all tokens with importance weighting instead of class token only (ablation) | WPAC DeiT-B · ImageNet-1k, 50% dims | top-1 accuracy: 65.1 -> 76.9 | Improved |
| Increase randomly sampled tokens from 1 to 2 (ablation) | WPAC DeiT-S · ImageNet-1k, 50% dims | top-1 accuracy: 64.5 -> 71.0 | Improved |
| Use 5 random tokens instead of 2 on DeiT-B (ablation) | WPAC DeiT-B · ImageNet-1k, 50% dims | top-1 accuracy: 75.7 -> Ill-Cond. | Hurt |
| Use unweighted PCA over all tokens (weight 1.0) (ablation) | WPAC · ImageNet-1k, 50% dims | top-1 accuracy: 76.9 -> Ill-Cond. | Hurt |
| Train with PIAD instead of layerdrop (ablation) | SKD DeiT-S · ImageNet-1k, 4:12 sub-network | top-1 accuracy: 39.7 -> 70.6 | Improved |
| Train with PIAD instead of weighted channel dropout (ablation) | SKD DeiT-S · ImageNet-1k, 4:12 sub-network | top-1 accuracy: 34.2 -> 70.6 | Improved |
| Train with PIAD instead of channel dropout (ablation) | SKD DeiT-S · ImageNet-1k, 10:12 sub-network | top-1 accuracy: 46.5 -> 79.0 | Improved |
| Train with PIAD instead of plain WPAC baseline (ablation) | SKD DeiT-Ti · ImageNet-1k, 4:12 sub-network | top-1 accuracy: 1.4 -> 61.4 | Improved |
| Use SKD instead of WAVE for scaled models without fine-tuning | DeiT-S · ImageNet-1k, 8:12 size | top-1 accuracy: 74.1 -> 78.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- WAVE — WAVE: Weight Templates for Adaptive Initialization of Variable-sized Models (CVPR 2025) · better on 6 of 6 shared results
- LPViT — LPViT: Low-Power Semi-structured Pruning for Vision Transformers (ECCV 2024) · better on 1 of 1 shared result
- X-Pruner — X-Pruner: eXplainable Pruning for Vision Transformers (CVPR 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.86)
- Efficient Data Driven Mixture-of-Expert Extraction from Trained Networks (CVPR 2025 · similarity 0.86)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.86)
- Hawaii: Hierarchical Visual Knowledge Transfer for Efficient Vision-Language Models (NeurIPS 2025 · similarity 0.85)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 2024 · similarity 0.85)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · 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.
