Research map / Architectures and efficiency / Attention and transformers
A Closer Look at Self-Supervised Lightweight Vision Transformers
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 MAE self-supervised pre-training on IN1K | ViT-Tiny · ImageNet-1k val | top-1 accuracy: 75.8 -> 78.0 | Helped |
| Add MoCo-v3 self-supervised pre-training on IN1K | ViT-Tiny · ImageNet-1k val | top-1 accuracy: 75.8 -> 76.8 | Helped |
| Add MAE pre-training and fine-tune for 300 epochs | ViT-Tiny · ImageNet-1k val | top-1 accuracy: 76.2 -> 78.5 (+2.3) | Helped |
| Add MAE pre-training and fine-tune for 1000 epochs | ViT-Tiny · ImageNet-1k val | top-1 accuracy: 77.8 -> 79.0 (+1.2) | Helped |
| Increase MAE pre-training data from IN1K to IN21K (ablation) | ViT-Tiny · ImageNet-1k val | top-1 accuracy: 78.0 -> 78.0 (+0.0) | No clear effect |
| Reduce MAE pre-training data to 1% IN1K (ablation) | ViT-Tiny · ImageNet-1k val | top-1 accuracy: 78.0 -> 77.9 (-0.1) | No clear effect |
| Reduce MAE pre-training data to 10% IN1K (ablation) | ViT-Tiny · ImageNet-1k val | top-1 accuracy: 78.0 -> 78.0 (+0.0) | No clear effect |
| Use long-tailed IN1K pre-training data (ablation) | ViT-Tiny with MAE · ImageNet-1k val | top-1 accuracy: 78.0 -> 77.9 (-0.1) | No clear effect |
| Add attention-map distillation from MAE-Base during MAE pre-training | ViT-Tiny · ImageNet-1k val | top-1 accuracy: 78.0 -> 78.4 (+0.4) | Helped |
| Add attention-map distillation from MAE-Base during MAE pre-training | MAE-Tiny · Flowers, Aircraft, and COCO object detection | top-1 accuracy / AP: 85.8 -> 95.2 (+9.4); 64.6 -> 79.2 (+14.6); 39.9 -> 42.3 (+2.4) | Helped |
| Distill attention at the last transformer block rather than lower blocks (ablation) | MAE-Tiny using MAE-Base teacher · ImageNet-1k | top-1 accuracy | Helped |
| Use global average pooling instead of class-token pooling (ablation) | MoCo-v3 ViT-Tiny fine-tuning · ImageNet-1k val | top-1 accuracy: 73.7% vs. 76.8% | Helped |
Closest papers by meaning
- Efficient Self-supervised Vision Transformers for Representation Learning (ICLR 2022 · similarity 0.91)
- Boosting Vanilla Lightweight Vision Transformers via Re-parameterization (ICLR 2024 · similarity 0.90)
- A Theoretical Analysis of Self-Supervised Learning for Vision Transformers (ICLR 2025 · similarity 0.90)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.90)
- Distilling Self-Supervised Vision Transformers for Weakly-Supervised Few-Shot Classification & Segmentation (CVPR 2023 · similarity 0.90)
- On the Surprising Effectiveness of Attention Transfer for Vision Transformers (NeurIPS 2024 · similarity 0.88)
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.