Research map / Architectures and efficiency / Attention and transformers
The Principle of Diversity: Training Stronger Vision Transformers Calls for Reducing All Levels of Redundancy
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 all-level diversity regularization | ViT-Small · ImageNet-1k | Top-1 accuracy (%): 76.54 -> 78.60; ↑2.04 | Helped |
| Apply all-level diversity regularization | ViT-Base · ImageNet-1k | Top-1 accuracy (%): 77.90 -> 79.96; ↑2.06 | Helped |
| Apply all-level diversity regularization | DeiT-Small · ImageNet-1k | Top-1 accuracy (%): 79.78 -> 80.61; ↑0.83 | Helped |
| Apply all-level diversity regularization | DeiT-Base · ImageNet-1k | Top-1 accuracy (%): 80.98 -> 81.68; ↑0.70 | Helped |
| Apply all-level diversity regularization | Swin-Small · ImageNet-1k | Top-1 accuracy (%): 83.18 -> 83.33; ↑0.15 | Helped |
| Apply all-level diversity regularization | Swin-Base · ImageNet-1k | Top-1 accuracy (%): 83.40 -> 83.72; ↑0.32 | Helped |
| Apply all-level diversity regularization | DeiT-Small24 · ImageNet-1k | Top-1 accuracy (%): 80.03 -> 81.79; ↑1.76 | Helped |
| Add within-layer and cross-layer embedding diversity together (ablation) | DeiT-Small with mixing loss · ImageNet | Top-1 testing accuracy (%): 79.99 -> 80.43 | Helped |
| Add attention-map diversity to embedding diversity and mixing loss (ablation) | DeiT-Small · ImageNet | Top-1 testing accuracy (%): 80.43 -> 80.53 | Helped |
| Add weight diversity to embedding, attention, and mixing-loss regularization (ablation) | DeiT-Small · ImageNet | Top-1 testing accuracy (%): 80.53 -> 80.61 | Helped |
| Use contrastive rather than cosine attention regularization (ablation) | DeiT-Small · ImageNet | Top-1 testing accuracy (%): 79.95 -> 69.98 | Hurt |
| Use embedding diversity regularization (ablation) | DeiT-Small · ImageNet | Top-1 testing accuracy (%): 79.78 -> 80.20 | Helped |
Closest papers by meaning
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · similarity 0.90)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.89)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 2024 · similarity 0.89)
- Improving Adversarial Transferability on Vision Transformers via Forward Propagation Refinement (CVPR 2025 · similarity 0.89)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.89)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.88)
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.