Research map / Architectures and efficiency / CNN design
Do Vision Transformers See Like Convolutional Neural Networks?
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 |
|---|---|---|---|
| use self-attention to incorporate global information in early layers | ViT · JFT-300M image representations | similarity of early-layer features to ResNet representations | Hurt |
| train with insufficient data for early attention layers to learn local attention | ViT-L/16 and ViT-H/14 · ImageNet training | classification performance | Hurt |
| remove skip connections from a middle block (ablation) | ViT · image classification | performance: ~4\% | Hurt |
| remove skip connections from block i (ablation) | ViT · JFT-300M image representations | representation similarity across layers | Hurt |
| use global average pooling instead of a CLS token (ablation) | ViT · higher-layer spatial localization | spatial localization | Hurt |
| pretrain on JFT-300M instead of ImageNet | ViT-L/16 and ViT-H/14 · ImageNet 10-shot | linear-probe accuracy: ~30\% | Helped |
| reduce pretraining dataset size | ViT-B/32 and ViT-L/16 · JFT-300M image representations | higher-layer representation similarity to full-data training: 3\% | Hurt |
| train with CLS rather than GAP classification (ablation) | ViT-B/32 · higher-layer spatial localization | spatial localization | Helped |
Closest papers by meaning
- Vision Transformers provably learn spatial structure (NeurIPS 2022 · similarity 0.93)
- Intriguing Properties of Vision Transformers (NeurIPS 2021 · similarity 0.91)
- Bridging the Gap Between Vision Transformers and Convolutional Neural Networks on Small Datasets (NeurIPS 2022 · similarity 0.91)
- Scaling Vision Transformers (CVPR 2022 · similarity 0.91)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.90)
- Structured Initialization for Vision Transformers (NeurIPS 2025 · similarity 0.89)
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.