Research map / Label-efficient and robust learning / Self-supervised and contrastive
SdAE: Self-Distillated Masked Autoencoder
Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, Qi Tian
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 teacher feature normalization (ablation) | SdAE with full-image teacher · ImageNet-1k fine-tuning (100 ep pretrain) | top-1 accuracy: 83.7 -> 84.0 | Improved |
| feed only masked tokens to the teacher instead of the full image (ablation) | SdAE · ImageNet-1k fine-tuning (100 ep pretrain) | top-1 accuracy: 84.0 -> 83.5 | Hurt |
| add multi-fold masking over masked-only teacher input (ablation) | SdAE · ImageNet-1k fine-tuning (100 ep pretrain) | top-1 accuracy: 83.5 -> 84.1 | Improved |
| replace full-image teacher input with multi-fold masking (ablation) | SdAE · ImageNet-1k fine-tuning (100 ep pretrain) | top-1 accuracy: 84.0 -> 84.1 | Improved |
| replace full-image teacher input with multi-fold masking (ablation) | SdAE · MAE-based training cost | training time per epoch (s): 826s -> 706s | Improved |
| feed the whole image instead of 36 patches to the teacher (ablation) | SdAE teacher crop · ImageNet-1k fine-tuning (100 ep pretrain) | top-1 accuracy: 82.81 -> 83.04 | Improved |
| replace MAE pixel targets with self-distilled latent targets | MAE · ImageNet-1k fine-tuning, ViT-B 300 ep | top-1 accuracy: 82.9 -> 84.1 | Improved |
| replace MAE pixel targets with self-distilled latent targets | MAE · ADE20K, 300 ep | mIoU: 45.8 -> 48.6 | Improved |
| replace MAE pixel targets with self-distilled latent targets | MAE · COCO detection, 300 ep | box AP: 45.4 -> 48.9 | Improved |
| pretrain longer, 300 to 800 epochs | SdAE · ImageNet-1k fine-tuning, ViT-B | top-1 accuracy: 84.1 -> 84.0 | Hurt |
| use a 2-layer decoder instead of 8 layers (ablation) | SdAE · ImageNet-1k fine-tuning (100 ep pretrain) | top-1 accuracy: 82.76 | No clear effect |
| update teacher with EMA every iteration instead of every epoch (ablation) | SdAE · ImageNet-1k fine-tuning | top-1 accuracy | Hurt |
Closest papers by meaning
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.90)
- Exploring Target Representations for Masked Autoencoders (ICLR 2024 · similarity 0.90)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.88)
- R-MAE: Regions Meet Masked Autoencoders (ICLR 2024 · similarity 0.88)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.88)
- DDAE: Towards Deep Dynamic Vision BERT Pretraining (AAAI 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.
