Research map / Architectures and efficiency / Attention and transformers
Edge-RecViT: Efficient Vision Transformer via Semantic-Refined Dynamic Recursion
YiZhou Li, Jinyi Xu, Mingyu Yin, Xianyi Zhao
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 the edge-aware ranker (EARR) to the fixed-depth shared model (ablation) | Edge-RecViT Base · ImageNet-1K | FLOPs (G): 35.2 -> 24.4 FLOPs, Top-1 82.1 -> 82.0 | Improved |
| share only middle layers instead of using fully independent layers (ablation) | Edge-RecViT Base · ImageNet-1K | Top-1 accuracy (%): 81.4 -> 82.0 (params 86M -> 23M) | Improved |
| share a single layer across the whole transformer instead of only the middle (ablation) | Edge-RecViT Base · ImageNet-1K | Top-1 accuracy (%): 82.0 -> 68.9 | Hurt |
| add STB-loss to ranker regularization (ablation) | Edge-RecViT Base · ImageNet-1K | Top-1 accuracy (%): 69.2 -> 82.0 | Improved |
| add DU-loss to ranker regularization (ablation) | Edge-RecViT Base · ImageNet-1K | Top-1 accuracy (%): 61.1 -> 82.0 (FLOPs 9.5 -> 24.39) | Improved |
| add both STB and DU regularization (ablation) | Edge-RecViT Base · ImageNet-1K | Top-1 accuracy (%): 55.7 -> 82.0 | Improved |
| Edge-RecViT versus DeiT on ImageNet-1K at Base | DeiT-B · ImageNet-1K | Top-1 accuracy (%): 81.8 -> 82.0 (86.0M -> 23.2M params, 35.2 -> 24.39 GFLOPs) | Improved |
| Edge-RecViT versus DeiT on ImageNet-1K at Tiny | DeiT-Ti · ImageNet-1K | Top-1 accuracy (%): 72.2 -> 72.4 (5.8M -> 1.6M params) | Improved |
| Edge-RecViT versus DeiT-Ti on CIFAR-100 | DeiT-Ti · CIFAR-100 | Top-1 accuracy (%): 86.61 -> 90.11 | Improved |
| Edge-RecViT versus DeiT-B on CIFAR-100 | DeiT-B · CIFAR-100 | Top-1 accuracy (%): 90.8 -> 90.78 | No clear effect |
Closest papers by meaning
- Vision-MoR: Scaling Vision Transformer via Patch-Level Mixture-of-Recursions (AAAI 2026 · similarity 0.91)
- Making Vision Transformers Efficient From a Token Sparsification View (CVPR 2023 · similarity 0.90)
- DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification (NeurIPS 2021 · similarity 0.90)
- Structured Hyperedge Adaptation for Parameter-Efficient Fine-Tuning of Vision Transformers (ECCV 2026 · similarity 0.89)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.88)
- RePaViT: Scalable Vision Transformer Acceleration via Structural Reparameterization on Feedforward Network Layers (ICML 2025 · 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.
