Research map / Architectures and efficiency / Attention and transformers
Vision Transformer Finetuning Benefits from Non-Smooth Components
Ambroise Odonnat, Laetitia Chapel, Romain Tavenard, Ievgen Redko
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 |
|---|---|---|---|
| finetune attention (MHA) instead of LN1 (ablation) | ViT-Base SGD · 11 benchmarks avg | top-1 accuracy (%): 89.8 -> 90.8 | Improved |
| finetune FC1 instead of LN1 (ablation) | ViT-Base SGD · 11 benchmarks avg | top-1 accuracy (%): 89.8 -> 90.7 | Improved |
| finetune FC2 instead of LN1 (ablation) | ViT-Base SGD · 11 benchmarks avg | top-1 accuracy (%): 89.8 -> 90.3 | Improved |
| finetune MHA instead of LN2 (ablation) | ViT-Base SGD · 11 benchmarks avg | top-1 accuracy (%): 89.9 -> 90.8 | Improved |
| finetune MHA instead of FC1 (ablation) | ViT-Base SGD · 11 benchmarks avg | top-1 accuracy (%): 90.7 -> 90.8 | No clear effect |
| finetune MHA instead of LN1 (ablation) | ViT-Base Adam · 4 datasets avg | top-1 accuracy (%): 79.2 -> 82.2 | Improved |
| finetune MHA instead of LN2 (ablation) | ViT-Base Adam · 4 datasets avg | top-1 accuracy (%): 79.4 -> 82.2 | Improved |
| finetune MHA instead of FC1 (ablation) | ViT-Base Adam · 4 datasets avg | top-1 accuracy (%): 81.4 -> 82.2 | Improved |
| finetune MHA instead of LN1 (ablation) | ViT-Base low-data · 4 datasets avg | top-1 accuracy (%): 93.0 -> 94.4 | Improved |
| finetune MHA instead of FC1 (ablation) | ViT-Base low-data · 4 datasets avg | top-1 accuracy (%): 93.8 -> 94.4 | Improved |
| finetune LayerNorms only instead of LoRA (ablation) | ViT-Base Adam · Cifar100 | top-1 accuracy (%): 88.1 -> 89.4 | Mixed |
| finetune FC1 instead of MHA (ablation) | ViT-Base Adam · Cifar100 | top-1 accuracy (%): 91.0 -> 91.3 | No clear effect |
Closest papers by meaning
- Unveil Benign Overfitting for Transformer in Vision: Training Dynamics, Convergence, and Generalization (NeurIPS 2024 · similarity 0.85)
- Transformers Learn Low Sensitivity Functions: Investigations and Implications (ICLR 2025 · similarity 0.85)
- Adapting Pretrained ViTs with Convolution Injector for Visuo-Motor Control (ICML 2024 · similarity 0.85)
- On the Surprising Effectiveness of Attention Transfer for Vision Transformers (NeurIPS 2024 · similarity 0.85)
- Improving Interpretation Faithfulness for Vision Transformers (ICML 2024 · similarity 0.85)
- Low-Rank Rescaled Vision Transformer Fine-Tuning: A Residual Design Approach (CVPR 2024 · 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.
