Research map / Architectures and efficiency / Attention and transformers
Scaling Vision Transformers
Check 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 |
|---|---|---|---|
| scale model size, data, and training compute together | ViT · ImageNet representation evaluation | error rate | Helped |
| increase model size | ViT pre-training · ImageNet 10-shot transfer | error rate: nearly 100 times more images | Helped |
| increase model size | ViT pre-training · ImageNet fine-tuning | error rate: about 20 | Helped |
| increase JFT dataset size from JFT-300M to JFT-3B (ablation) | ViT-B/32 and ViT-L/16 · ImageNet transfer | accuracy: about 1% | Helped |
| decouple weight decay and apply stronger decay to the final prediction head (ablation) | ViT-B/32 · ImageNet few-shot transfer | accuracy | Helped |
| replace the [class] token with global average pooling or multihead attention pooling (ablation) | ViT · ImageNet representation evaluation | accuracy | No clear effect |
| remove the final nonlinear projection before the prediction layer (ablation) | ViT pooling head · ImageNet representation evaluation | accuracy | No clear effect |
| store the first momentum in bfloat16 (ablation) | Adam optimizer · ViT training | training outcome | No clear effect |
| store the second momentum in half precision (ablation) | Adam optimizer · ViT training | training performance | Hurt |
| use the modified Adafactor optimizer (ablation) | large ViT models · ViT training | training performance | Helped |
| use reciprocal square-root schedule with cooldown instead of linear decay (ablation) | ViT pre-training · ImageNet validation | validation score | No clear effect |
| use cyclic learning-rate schedules (ablation) | ViT pre-training · training validation | validation score | Hurt |
Closest papers by meaning
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.90)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.89)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.89)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.88)
- On the Surprising Effectiveness of Attention Transfer for Vision Transformers (NeurIPS 2024 · similarity 0.88)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 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.