Research map / Architectures and efficiency / Attention and transformers
Efficient Visual Transformer by Learnable Token Merging
Yancheng Wang, Yingzhen Yang
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 |
|---|---|---|---|
| replace transformer blocks with LTM blocks (learnable token merging) | MobileViT-S · ImageNet-1k val (from scratch) | top-1 accuracy: 78.4 -> 79.7 | Improved |
| replace transformer blocks with LTM blocks | Swin-T · ImageNet-1k val (from scratch) | top-1 accuracy: 81.3 -> 81.8 | Improved |
| replace transformer blocks with LTM blocks | ViT-B · ImageNet-1k val (from scratch) | top-1 accuracy: 83.7 -> 83.9 | Improved |
| LTM instead of ToMe/LTMP token merging | MobileViT-S · ImageNet-1k val (from scratch) | top-1 accuracy: 76.7 -> 79.7 | Improved |
| LTM instead of LTMP token merging, fine-tuned 50 epochs | Swin-B · ImageNet-1k val (fine-tuned 50 epochs) | top-1 accuracy: 82.55 -> 83.64 | Improved |
| LTM instead of LTMP token merging, fine-tuned 50 epochs | Swin-T · ImageNet-1k val (fine-tuned 50 epochs) | top-1 accuracy: 80.30 -> 81.38 | Improved |
| LTM backbone for SSDLite detection | MobileViT-S · COCO detection (SSDLite) | mAP: 27.7 -> 28.4 | Improved |
| LTM backbone for Mask R-CNN | EfficientViT-B1 · COCO instance segmentation (Mask R-CNN) | mAP box: 33.5 -> 34.3 | Improved |
| add token merging with learnable mask (LTM) to reduce IB loss (ablation) | MobileViT-S · ImageNet-1k val (from scratch, IB loss) | IB loss: -0.00432 -> -0.01725 | Improved |
| add token merging with learnable mask (LTM) vs ToMe (ablation) | ViT-B · ImageNet-1k val (from scratch, IB loss) | IB bound: 0.04583 -> 0.03094 | Improved |
| lower compression ratio from 0.95 to 0.65 (ablation) | LTM-ViT-B · ImageNet-1k val (from scratch, compression ratio ablation) | top-1 accuracy: 84.43 -> 83.75 | Hurt |
| lower compression ratio from 0.65 to 0.40 (ablation) | LTM-ViT-B · ImageNet-1k val (from scratch, compression ratio ablation) | top-1 accuracy: 83.75 -> 82.23 | Hurt |
Closest papers by meaning
- Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers (NeurIPS 2024 · similarity 0.88)
- A General and Efficient Training for Transformer via Token Expansion (CVPR 2024 · similarity 0.88)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.87)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.87)
- vid-TLDR: Training Free Token Merging for Light-weight Video Transformer (CVPR 2024 · similarity 0.87)
- Visual Transformer with Differentiable Channel Selection: An Information Bottleneck Inspired Approach (ICML 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.
