Research map / Architectures and efficiency / Attention and transformers
Sliced Recursive Transformer
Zhiqiang Shen, Zechun Liu, Eric Xing
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 non-linear projection layer (NLL) between recursions (ablation) | recursive transformer · ImageNet-1K | Top-1: 72.5 -> 74.7 | Improved |
| add learnable residual connection (LRC) (ablation) | Recursion + NLL · ImageNet-1K | Top-1: 74.7 -> 75.2 | Improved |
| add spatial pyramid stem (ablation) | Recursion + NLL · ImageNet-1K | Top-1: 74.7 -> 76.0 | Improved |
| remove class token (ablation) | Recursion + NLL · ImageNet-1K | Top-1: 74.7 -> 75.0 | Improved |
| use all components together (ablation) | Recursion + NLL · ImageNet-1K | Top-1: 74.7 -> 76.8 | Improved |
| soft distillation instead of GT + hard distillation (ablation) | SReT-TL · ImageNet-1K | Top-1: 77.5 -> 77.9 | Improved |
| use group self-attention with recursion instead of global attention without loop (ablation) | SReT-ExT · ImageNet-1K | Top-1: 73.6 -> 74.0 | Improved |
| use internal loops instead of external loop (ablation) | 12-layer recursive transformer · ImageNet-1K | Top-1: 67.0 -> 67.6 | Improved |
| use group self-attention approximation [8,8][4,4][1,1] (ablation) | SReT-TL · ImageNet-1K | Top-1: 76.78 -> 76.06 | Hurt |
| use fewer groups [8,2][4,1][1,1] (ablation) | SReT-TL · ImageNet-1K | Top-1: 76.06 -> 76.65 | Improved |
| add inverse permutation layer after sliced group attention (ablation) | SReT-TL group design · ImageNet-1K | Top-1: 75.41 -> 75.94 | Improved |
| add distillation | SReT-T · ImageNet-1K | Top-1: 76.0 -> 77.6 | Improved |
Closest papers by meaning
- Edge-RecViT: Efficient Vision Transformer via Semantic-Refined Dynamic Recursion (CVPR 2026 · similarity 0.89)
- Global Vision Transformer Pruning With Hessian-Aware Saliency (CVPR 2023 · similarity 0.89)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.88)
- LPViT: Low-Power Semi-structured Pruning for Vision Transformers (ECCV 2024 · similarity 0.88)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · similarity 0.87)
- RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers (ICML 2025 · 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.
