Research map / Architectures and efficiency / Pruning and efficient tuning
Parameter-Efficient and Memory-Efficient Tuning for Vision Transformer: A Disentangled Approach
Taolin Zhang, Jiawang Bai, Zhihe Lu, Dongze Lian, genping wang, Xinchao Wang, Shu-Tao Xia
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 VPT-Deep with SynQT disentangled query-only tuning | ViT-B/16 ImageNet-21K · VTAB-1K | training memory (GB): 8.1 -> 3.4 | Improved |
| replace VPT-Deep with SynQT | ViT-B/16 ImageNet-21K · VTAB-1K | average Top-1 accuracy: 72.0 -> 72.9 | Improved |
| replace VQT with SynQT | ViT-B/16 ImageNet-21K · VTAB-1K | average Top-1 accuracy: 68.3 -> 72.9 | Improved |
| apply memory constraint by dropping structures near input; SynQT versus AdaptFormer | ViT-B/16 ImageNet-21K · VTAB-1K (5GB constraint) | average Top-1 accuracy: 65.8 -> 72.9 | Improved |
| apply SynQT to different pre-training backbones | MAE, DINO, supervised ImageNet-1K ViT-B/16 · VTAB-1K | average Top-1 accuracy: MAE 67.1 -> 68.9; DINO 69.9 -> 72.3; supervised 70.6 -> 71.4 | Improved |
| apply SynQT to Swin-B and ConvNeXt-B | Swin-B, ConvNeXt-B · VTAB-1K | average Top-1 accuracy: Swin 71.6 -> 73.8; ConvNeXt 68.7 -> 69.5 | Improved |
| add trainable prompt to QSM input (ablation) | SynQT · VTAB-1K | average Top-1 accuracy: 70.7 -> 72.9 | Improved |
| feed previous block output into QSM (ablation) | SynQT · VTAB-1K | average Top-1 accuracy: 71.4 -> 72.9 | Improved |
| add DropFeat regularization in classification head (ablation) | SynQT · VTAB-1K | average Top-1 accuracy: 72.4 -> 72.9 | Improved |
| use conditional feature aggregation instead of simple average (ablation) | SynQT head · VTAB-1K | average Top-1 accuracy: 72.7 -> 72.9 | Improved |
| use shared instead of independent projection in head (ablation) | SynQT head · VTAB-1K | average Top-1 accuracy: 72.6 -> 72.9 | Improved |
| increase QSM hidden size beyond 48 (ablation) | SynQT · VTAB-1K | average Top-1 accuracy: 72.9 -> 72.7 (width 96) | No clear effect |
Closest papers by meaning
- DTL: Disentangled Transfer Learning for Visual Recognition (AAAI 2024 · similarity 0.91)
- PELA: Learning Parameter-Efficient Models with Low-Rank Approximation (CVPR 2024 · similarity 0.88)
- Efficient Adaptation of Pre-trained Vision Transformer via Householder Transformation (NeurIPS 2024 · similarity 0.88)
- Low-Rank Rescaled Vision Transformer Fine-Tuning: A Residual Design Approach (CVPR 2024 · similarity 0.88)
- Transforming Vision Transformer: Towards Efficient Multi-Task Asynchronous Learner (NeurIPS 2024 · similarity 0.88)
- TADFormer: Task-Adaptive Dynamic TransFormer for Efficient Multi-Task Learning (CVPR 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.
