Research map / Label-efficient and robust learning / Self-supervised and contrastive
POA: Pre-training Once for Models of All Sizes
Yingying Zhang, Xin Guo, Jiangwei Lao, Lei Yu, Lixiang Ru, Jian Wang, Guo Ye, HUIMEI HE, Jingdong Chen, Ming Yang
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 elastic student branch with once-for-all sub-network extraction (ViT-B from ViT-L teacher) | POA · ImageNet-1K k-NN, ViT-B/16 | top-1 accuracy (%): 77.1 -> 80.9 (vs iBOT) | Improved |
| add same-view distillation loss L_ES2 from intact to elastic student (ablation) | POA · ImageNet-1K k-NN, ViT-S/16 | top-1 accuracy (%): 72.8 -> 76.2 (w/o vs with, no MPH) | Improved |
| add cross-view elastic loss L_ES1 (ablation) | POA · ImageNet-1K k-NN, ViT-S/16 | top-1 accuracy (%): 75.1 -> 76.2 | Improved |
| add multiple projection heads (MPH) with different prototype counts (ablation) | POA · ImageNet-1K k-NN, ViT-S/16 | top-1 accuracy (%): 76.2 -> 76.8 | Improved |
| extract sub-networks from POA teacher instead of SEED distillation from a pre-trained ViT-L (600 epochs) | ViT-S/16 · ImageNet-1K k-NN | top-1 accuracy (%): 74.0 -> 76.8 | Improved |
| drop intact student and use elastic teacher (POA-V1) (ablation) | POA · ImageNet-1K k-NN, ViT-L/16 | top-1 accuracy (%): 82.3 -> 79.2 | Hurt |
| drop intact student, EMA teacher from elastic students (POA-V2) (ablation) | POA · ImageNet-1K k-NN, ViT-L/16 | top-1 accuracy (%): 82.3 -> 80.9 | Hurt |
| add an extra elastic teacher branch (POA-V3) (ablation) | POA · ImageNet-1K k-NN, ViT-S/16 | top-1 accuracy (%): 76.8 -> 76.9 | No clear effect |
| use POA-pretrained ViT-S/16 as backbone instead of iBOT | Cascade Mask R-CNN · COCO detection | AP^b: 49.4 -> 50.6 | Improved |
| use POA-pretrained ViT-B/16 instead of iBOT with linear segmentation head | ADE20K linear segmentation · ADE20K | mIoU: 38.3 -> 40.3 | Improved |
| use extracted ResNet-50 vs ReLICv2 trained 4000 epochs | POA · ImageNet-1K linear probing, ResNet-50 | top-1 accuracy (%): 77.1 -> 76.9 | No clear effect |
| add elastic student branch to ResNet pre-training | POA with ResNet backbone · pre-training stability | training stability | Improved |
Closest papers by meaning
- The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining (ICCV 2023 · similarity 0.88)
- Chain-of-Models Pre-Training: Rethinking Training Acceleration of Vision Foundation Models (CVPR 2026 · similarity 0.88)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.87)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.86)
- Harnessing small projectors and multiple views for efficient vision pretraining (NeurIPS 2024 · similarity 0.86)
- PELA: Learning Parameter-Efficient Models with Low-Rank Approximation (CVPR 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.
