Research map / Architectures and efficiency / Attention and transformers
ViTAE: Vision Transformer Advanced by Exploring Intrinsic Inductive Bias
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 |
|---|---|---|---|
| add parallel convolution module (PCM) in normal cells with pre-fusion and BN (ablation) | T2T-ViT-7 · ImageNet-1k val (100 ep) | top-1: 68.7 -> 69.9 | Helped |
| add BN in PCM (pre-fusion) (ablation) | ViTAE NC · ImageNet-1k val (100 ep) | top-1: 69.1 -> 69.9 | Helped |
| use larger dilation rates [1,2,3,4,5] in all RCs instead of [1,2,3] (ablation) | ViTAE RC · ImageNet-1k val (100 ep) | top-1: 69.9 -> 68.9 | Hurt |
| add PCM to reduction cells (ablation) | RC with dilation · ImageNet-1k val (100 ep) | top-1: 69.8 -> 71.7 | Helped |
| combine RCs and NCs (ablation) | RC-only ViTAE · ImageNet-1k val (100 ep) | top-1: 71.7 -> 72.6 | Helped |
| add extra BN to make PCM three identical conv layers (ablation) | ViTAE · ImageNet-1k val (100 ep) | top-1: 72.6 -> 69.6 | Hurt |
| replace PCM with PRM in normal cells (ablation) | ViTAE · ImageNet-1k val (100 ep) | top-1: 72.6 -> 71.9 | Hurt |
| replace SiLU with GeLU in PCM (ablation) | ViTAE · ImageNet-1k val (100 ep) | top-1: 72.6 -> 72.3 | Hurt |
| remove position embedding (ablation) | ViTAE-T · ImageNet-1k val | top-1: 75.3 -> 75.3 | No clear effect |
| ViTAE-T vs T2T-ViT-7 baseline (300 ep) | T2T-ViT-7 · ImageNet-1k val | top-1: 71.7 -> 75.3 | Helped |
| stage-wise design (ablation) | ViTAE-S · ImageNet-1k val | top-1: 82.0 -> 82.2 | Helped |
| ViTAE backbone instead of Swin-T | Mask RCNN · COCO detection | box mAP: 43.7 -> 44.6 | Helped |
Closest papers by meaning
- Co-Advise: Cross Inductive Bias Distillation (CVPR 2022 · similarity 0.88)
- ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · similarity 0.87)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.87)
- Bridging the Gap Between Vision Transformers and Convolutional Neural Networks on Small Datasets (NeurIPS 2022 · similarity 0.87)
- Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · similarity 0.87)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.86)
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.