Research map / Architectures and efficiency / Pruning and efficient tuning
Beyond Static Allocation: Dynamic Sensitivity-Aware Fine-Tuning for Vision Transformers
Yuanyang Cao, Xichun Liu, Fuwei Zhang, Shangqi Deng, Ziyang Ren, Jianji Wang
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 dynamic reconfiguration over static allocation (ablation) | DAF · VTAB-1k | mean accuracy (%): 75.8 -> 76.4 | Improved |
| add context-aware sensitivity analysis (with LoRA in forward pass) instead of pristine backbone analysis (ablation) | DAF · VTAB-1k | mean accuracy (%): 74.6 -> 76.4 | Improved |
| add Rebuild-and-Refocus (freeze outdated, reallocate) instead of accumulating parameters (ablation) | DAF · VTAB-1k | mean accuracy (%): 75.3 -> 76.4 | Improved |
| use Taylor-based |g*w| sensitivity instead of random selection (ablation) | DAF · VTAB-1k | mean accuracy (%): 73.5 -> 76.4 | Improved |
| use |g*w| instead of gradient magnitude |g| (ablation) | DAF · VTAB-1k | mean accuracy (%): 75.0 -> 76.4 | Improved |
| reconfiguration interval of 10 epochs versus 5, 25 or 50 (ablation) | DAF · VTAB-1k | mean accuracy (%): 76.0 / 76.4 / 76.1 / 75.9 for 5/10/25/50 epochs | Mixed |
| increase parameter budget tau beyond 0.2 (ablation) | DAF · VTAB-1k | mean accuracy (%): 76.4 -> 76.2 (0.3), 76.1 (0.4) | Hurt |
| decrease parameter budget tau from 0.2 to 0.1 (ablation) | DAF · VTAB-1k | mean accuracy (%): 76.4 -> 75.7 | Hurt |
| vary number of sensitivity batches M from 8 to 32 (ablation) | DAF · VTAB-1k | mean accuracy (%): within approximately 0.3% | No clear effect |
| apply DAF to MAE and MoCo v3 self-supervised ViT backbones | DAF · VTAB-1k (MAE) | mean accuracy (%): 69.2 -> 69.9 vs SPT-Adapter | Improved |
| apply DAF to dense prediction (detection) versus LoRA | Swin-B Cascade Mask R-CNN · COCO val2017 | AP box: 50.4 -> 53.5 | Improved |
| apply DAF to semantic segmentation versus LoRA | Swin-L UperNet · ADE20K | mIoU: 50.3 -> 52.0 | 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.
- Res-Tuning — Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone (NeurIPS 2023) · better on 2 of 2 shared results
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
- Sensitivity-Aware Efficient Fine-Tuning via Compact Dynamic-Rank Adaptation (CVPR 2025 · similarity 0.90)
- Dynamic Tuning Towards Parameter and Inference Efficiency for ViT Adaptation (NeurIPS 2024 · similarity 0.89)
- Low-Rank Rescaled Vision Transformer Fine-Tuning: A Residual Design Approach (CVPR 2024 · similarity 0.88)
- Efficient Adaptation of Pre-trained Vision Transformer via Householder Transformation (NeurIPS 2024 · similarity 0.88)
- TADFormer: Task-Adaptive Dynamic TransFormer for Efficient Multi-Task Learning (CVPR 2025 · similarity 0.88)
- Be Confident in What You Know: Bayesian Parameter Efficient Fine-Tuning of Vision Foundation Models (NeurIPS 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.
