Research map / Architectures and efficiency / Pruning and efficient tuning
AdaptFormer: Adapting Vision Transformers for Scalable Visual Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| parallel AdaptMLP adapters instead of full fine-tuning | ViT-B/16 · SSv2 | top-1 accuracy: 53.97 -> 59.02 | Helped |
| AdaptFormer instead of VPT | ViT-B/16 · SSv2 | top-1 accuracy: 43.73 -> 59.02 | Helped |
| AdaptFormer instead of full fine-tuning | ViT-B/16 · Food-101 | top-1 accuracy: 90.09 -> 87.61 | Hurt |
| AdaptFormer instead of full fine-tuning | ViT-B/16 · NUS-WIDE | mAP: 61.26 -> 59.07 | Hurt |
| AdaptFormer instead of full fine-tuning with image-pretrained model | ViT-B/16 · SSv2 | top-1 accuracy: 41.50 -> 46.06 | Helped |
| increase middle dimension 1 -> 64 (ablation) | AdaptFormer · SSv2 val | top-1 accuracy: 50.03 -> 59.02 | Helped |
| increase middle dimension 64 -> 256 (ablation) | AdaptFormer · SSv2 val | top-1 accuracy: 59.02 -> 58.87 | No clear effect |
| insert adapters in all 12 layers instead of layers 7-12 (ablation) | AdaptFormer · SSv2 val | top-1 accuracy: 57.99 -> 59.02 | Helped |
| place adapters in top layers 7-12 instead of bottom 1-6 (ablation) | AdaptFormer · SSv2 val | top-1 accuracy: 50.48 -> 57.99 | Helped |
| parallel instead of sequential adapter insertion (ablation) | AdaptFormer · SSv2 val | top-1 accuracy: 58.17 -> 59.02 | Helped |
| scaling factor 0.10 instead of 0.01 (ablation) | AdaptFormer · SSv2 val | top-1 accuracy: 53.44 -> 59.02 | Helped |
| more VPT prompt tokens (>=8) | VPT · HMDB51 | top-1 accuracy | Hurt |
Closest papers by meaning
- Memory Efficient Transformer Adapter for Dense Predictions (ICLR 2025 · similarity 0.89)
- Dynamic Tuning Towards Parameter and Inference Efficiency for ViT Adaptation (NeurIPS 2024 · similarity 0.88)
- Efficient Adaptation of Pre-trained Vision Transformer via Householder Transformation (NeurIPS 2024 · similarity 0.87)
- Adapters Strike Back (CVPR 2024 · similarity 0.87)
- Time- Memory- and Parameter-Efficient Visual Adaptation (CVPR 2024 · similarity 0.87)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale 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.