Research map / Architectures and efficiency / Neural architecture search
ViTAS: Vision Transformer Architecture Search
Xiu Su, Shan You, Jiyang Xie, Mingkai Zheng, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, Chang Xu
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 |
|---|---|---|---|
| add cyclic weight sharing instead of ordinal sharing (ablation) | ViTAS superformer · ImageNet-1k (Twins space, 1.4G) | top-1 accuracy: 77.2 -> 77.9 | Improved |
| add cyclic weight sharing instead of bilateral sharing (ablation) | ViTAS superformer · ImageNet-1k (Twins space, 1.4G) | top-1 accuracy: 77.6 -> 77.9 | Improved |
| add weak augmentation to superformer training with cyclic sharing (ablation) | ViTAS · ImageNet-1k (Twins space, 1.4G) | top-1 accuracy: 77.9 -> 79.4 | Improved |
| add weak augmentation to ordinal sharing (ablation) | AutoSlim-style ordinal superformer · ImageNet-1k (Twins space, 1.4G) | top-1 accuracy: 77.2 -> 78.1 | Improved |
| add weak augmentation to bilateral sharing (ablation) | BCNet-style bilateral superformer · ImageNet-1k (Twins space, 1.4G) | top-1 accuracy: 77.6 -> 78.5 | Improved |
| progressively remove augmentation and regularization (RandAug, Mixup, CutMix, jitter, erasing, stoch depth, repeated aug, weight decay) from superformer (ablation) | ViTAS superformer · ImageNet-1k search (Twins 1.4G) | search accuracy: 59.4% -> 67.7% | Improved |
| progressively remove augmentation and regularization from superformer (ablation) | ViTAS superformer · ImageNet-1k retraining (Twins 1.4G) | retrain top-1: 77.9% -> 79.4% | Improved |
| add identity shifting to the superformer (ablation) | ViTAS superformer · ImageNet-100 | final training loss: 2.4 -> 1.5 | Improved |
| add identity shifting; retrain searched architecture from scratch (ablation) | ViTAS · ImageNet-100 | accuracy: 88.3% -> 90.4% | Improved |
| search with private class tokens (ablation) | ViTAS-DeiT-A · ImageNet-1k (DeiT space) | top-1 accuracy: 75.1 -> 75.6 | Improved |
| search with private class tokens (ablation) | ViTAS-DeiT-B · ImageNet-1k (DeiT space) | top-1 accuracy: 79.9 -> 80.2 | Improved |
| use cyclic weight sharing in ResNet50 dimension search instead of bilateral (BCNet) (ablation) | ResNet50 search · ImageNet-1k (1.4G) | top-1 accuracy: 75.4 -> 75.8 | Improved |
Closest papers by meaning
- Vision Transformer Neural Architecture Search for Out-of-Distribution Generalization: Benchmark and Insights (NeurIPS 2024 · similarity 0.88)
- Vision Transformers with Self-Distilled Registers (NeurIPS 2025 · similarity 0.88)
- Auto-Prox: Training-Free Vision Transformer Architecture Search via Automatic Proxy Discovery (AAAI 2024 · similarity 0.88)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.87)
- A General and Efficient Training for Transformer via Token Expansion (CVPR 2024 · similarity 0.87)
- DA-VPT: Semantic-Guided Visual Prompt Tuning for Vision Transformers (CVPR 2025 · 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.
