Research map / Architectures and efficiency / Attention and transformers
Self-Slimmed Vision Transformer
Zhuofan Zong, Kunchang Li, Guanglu Song, Yali Wang, Yu Qiao, Biao Leng, Yu Liu
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 token hard dropping (DynamicViT) with soft slimming TSM at FLOPs ratio 0.25 | SiT on DeiT-P · ImageNet | Top-1 accuracy: 65.6 -> 75.9 | Improved |
| replace DynamicViT with SiT at FLOPs ratio 0.5 | SiT on DeiT-P · ImageNet | Top-1 accuracy: 78.4 -> 79.4 | Improved |
| replace static aggregation (3x3 conv) with dynamic TSM (ablation) | SiT-Ti · ImageNet | Top-1 accuracy: 79.3 -> 80.1 | Improved |
| use learnable aggregation instead of 3x3 average pooling (ablation) | SiT-Ti · ImageNet | Top-1 accuracy: 77.4 -> 80.1 | Improved |
| use Token-MLP plus MLP recalibration (RTSM) instead of interpolation (ablation) | SiT-Ti · ImageNet | Top-1 accuracy: 78.3 -> 80.1 | Improved |
| add soft logits supervision (ablation) | SiT-Ti · ImageNet | Top-1 accuracy: 77.7 -> 79.0 (+1.3) | Improved |
| add block-to-block token supervision (ablation) | SiT-Ti · ImageNet | Top-1 accuracy: 79.0 -> 80.1 | Improved |
| add hard label supervision (ablation) | SiT-Ti · ImageNet | Top-1 accuracy: 80.1 -> 80.2 | Improved |
| add FRD (feature recalibration distillation) to TSM (ablation) | SiT-Ti · ImageNet | Top-1 accuracy: 77.7 -> 80.1 | Improved |
| use convolutional patch embedding instead of original DeiT patch embedding at FLOPs ratio 0.25 (ablation) | DeiT student · ImageNet | Top-1 accuracy: 75.3 -> 78.8 | Improved |
| use FRD combined with CNN distillation at FLOPs ratio 0.5 on DeiT_C (ablation) | DeiT_C student · ImageNet | Top-1 accuracy: 80.0 -> 80.6 | Improved |
| use token slimming (TSM) instead of scaling model depth (ablation) | SiT-Ti · ImageNet | Top-1 accuracy: 69.4 -> 77.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LeViT-256 — LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference (ICCV 2021) · 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
- Vision Transformer Slimming: Multi-Dimension Searching in Continuous Optimization Space (CVPR 2022 · similarity 0.90)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.88)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.88)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.87)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.87)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 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.
