Research map / Label-efficient and robust learning / Self-supervised and contrastive
Learning with Unmasked Tokens Drives Stronger Vision Learners
Taekyung Kim, Sanghyuk Chun, Byeongho Heo, Dongyoon Han
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 broader contextualization loss on unmasked tokens to MAE | MAE ViT-B/16 · ImageNet-1K fine-tuning | top-1 accuracy: 83.7 -> 84.2 | Improved |
| add LUT to MAE at 400 epochs (ablation) | MAE ViT-B/16 · ImageNet-1K ft (400 epochs) | top-1 accuracy: 82.8 -> 83.5 | Improved |
| also steer masked tokens in addition to unmasked (ablation) | LUT · ImageNet-1K ft (400 epochs) | top-1 accuracy: 83.5 -> 83.0 | Hurt |
| use pooled visual tokens instead of cls-token as context (ablation) | LUT · ImageNet-1K ft (400 epochs) | top-1 accuracy: 83.2 -> 83.5 | Improved |
| add token-wise supervision on top of aggregated contextualization (ablation) | LUT · ImageNet-1K ft (400 epochs) | top-1 accuracy: 83.5 -> 83.1 | Hurt |
| use cosine distance instead of infoNCE or smoothed-l1 (ablation) | LUT · ImageNet-1K ft (400 epochs) | top-1 accuracy: 83.0 -> 83.5 | Improved |
| mask the target encoder input (ablation) | LUT · ImageNet-1K ft (400 epochs) | top-1 accuracy: 83.5 -> 82.5 | Hurt |
| use simple resized crop instead of random resized crop in LUT (ablation) | LUT · ImageNet-1K ft (400 epochs) | top-1 accuracy: 83.4 -> 83.5 | Improved |
| use simple resized crop instead of random resized crop in MAE (ablation) | MAE · ImageNet-1K ft (400 epochs) | top-1 accuracy: 82.8 -> 82.5 | Hurt |
| increase loss weight alpha from 0.25 to 1.0 (ablation) | LUT · ImageNet-1K ft (400 epochs) | top-1 accuracy: 83.5 -> 82.9 | Hurt |
| train with contextualization loss only, without MIM loss (ablation) | LUT · ImageNet-1K ft (100 epochs) | accuracy: 82.1 -> 82.0 | No clear effect |
| apply LUT to SimMIM | SimMIM · ImageNet-1K ft (SimMIM, 100 epochs) | accuracy: 81.6 -> 81.8 | Improved |
Closest papers by meaning
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.90)
- On the Role of Discrete Tokenization in Visual Representation Learning (ICLR 2024 · similarity 0.90)
- Morphing Tokens Draw Strong Masked Image Models (ICLR 2025 · similarity 0.89)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.89)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.89)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.89)
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.
