Research map / Architectures and efficiency / Pruning and efficient tuning
Scaling & Shifting Your Features: A New Baseline for Efficient Model Tuning
Dongze Lian, Zhou Daquan, Jiashi Feng, Xinchao 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 |
|---|---|---|---|
| Scale and shift features of a frozen pretrained model | ViT-B/16 · FGVC (mean) | Top-1 accuracy: 88.54 -> 90.72 | Improved |
| Scale and shift features of a frozen pretrained model | ViT-B/16 · VTAB-1k (mean) | Top-1 accuracy: 65.57 -> 73.10 | Improved |
| Use feature scaling and shifting instead of full fine-tuning | ViT-B/16 · ImageNet-1K | Top-1 accuracy: 83.58 -> 83.10 | Hurt |
| Scale and shift features of a frozen pretrained model | ViT-B/16 fine-tuned on ImageNet-1K · ImageNet-A | Top-1 accuracy: 34.49 -> 45.88 | Improved |
| Scale and shift features of a frozen pretrained model | ViT-B/16 fine-tuned on ImageNet-1K · ImageNet-R | Top-1 accuracy: 51.29 -> 56.77 | Improved |
| Scale and shift features of a frozen pretrained model | ViT-B/16 fine-tuned on ImageNet-1K · ImageNet-C | mCE: 46.47 -> 41.47 | Improved |
| Use SSF with Swin-T instead of full fine-tuning | Swin-T · COCO val2017 (Mask R-CNN) | AP^b: 43.7 -> 34.9 | Hurt |
| Use SSF with Swin-T instead of full fine-tuning | Swin-T · ADE20K (UPerNet) | mIoU: 44.5 -> 38.9 | Hurt |
| Add SSF-ADA in more transformer layers (ablation) | SSF on ViT-B/16 · CIFAR-100 | Top-1 accuracy: 88.70 -> 92.69 -> 93.30 -> 93.60 -> 93.99 | Improved |
| Add SSF-ADA after MLP operations (ablation) | SSF on ViT-B/16 · CIFAR-100 | Top-1 accuracy: 93.46 -> 93.99 | Improved |
| Add SSF-ADA after attention operations (ablation) | SSF on ViT-B/16 · CIFAR-100 | Top-1 accuracy: 93.69 -> 93.99 | Improved |
| Initialize scale factors around one and shift factors around zero rather than randomly around zero (ablation) | SSF on ViT-B/16 · CIFAR-100 | Top-1 accuracy: 90.11 -> 93.99 | Improved |
Closest papers by meaning
- Tuning Stable Rank Shrinkage: Aiming at the Overlooked Structural Risk in Fine-tuning (CVPR 2024 · similarity 0.86)
- Mitigating Parameter Interference in Model Merging via Sharpness-Aware Fine-Tuning (ICLR 2025 · similarity 0.86)
- Domain-Aware Fine-Tuning: Enhancing Neural Network Adaptability (AAAI 2024 · similarity 0.85)
- SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors (NeurIPS 2024 · similarity 0.85)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · similarity 0.85)
- TTE: Two Tokens Are Enough to Improve Parameter-Efficient Tuning (AAAI 2025 · similarity 0.85)
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.
