Research map / Label-efficient and robust learning / Self-supervised and contrastive
Masked Siamese Networks for Label-Efficient Learning
Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas
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 a random mask to the anchor view (ablation) | MSN · ImageNet-1K 1% labels (ViT-B/16) | top-1 accuracy: 49.3 -> 52.3 | Improved |
| use only focal mask instead of no mask (ablation) | MSN · ImageNet-1K 1% labels (ViT-B/16) | top-1 accuracy: 49.3 -> 39.3 | Hurt |
| add focal mask views to random mask views (ablation) | MSN · ImageNet-1K 1% labels (ViT-B/16) | top-1 accuracy: 52.3 -> 59.8 | Improved |
| add ColorJitter to the anchor view relative to sharing the target view (ablation) | MSN · ImageNet-1K 1% labels (ViT-B/16) | top-1 accuracy: 7.0 -> 48.7 | Improved |
| add independent crop, flip and blur to the anchor view (ablation) | MSN · ImageNet-1K 1% labels (ViT-B/16) | top-1 accuracy: 48.7 -> 52.3 | Improved |
| increase random masking ratio from 0.15 to 0.5 for ViT-S/16 (ablation) | MSN · ImageNet-1K 1% labels (ViT-S/16) | top-1 accuracy: 66.3 -> 64.8 | Hurt |
| increase random masking ratio from 0.15 to 0.3 for ViT-B/16 (ablation) | MSN · ImageNet-1K 1% labels (ViT-B/16) | top-1 accuracy: 68.8 -> 69.6 | Improved |
| increase random masking ratio from 0.5 to 0.7 for ViT-L/16 (ablation) | MSN · ImageNet-1K 1% labels (ViT-L/16) | top-1 accuracy: 70.1 -> 69.4 | Hurt |
| increase masking ratio from 0.0 to 0.7 on ViT-L/7 (ablation) | MSN · pre-training efficiency | throughput (imgs/s): 415 -> 600 | Improved |
| replace Sinkhorn normalization with tuned me-max weight 5.0 (ablation) | MSN ViT-S/16 · ImageNet-1K 1% labels (ViT-S/16) | top-1 accuracy: 66.4 -> 67.2 | Improved |
| remove Sinkhorn normalization without retuning me-max weight (ablation) | MSN ViT-S/16 · ImageNet-1K 1% labels (ViT-S/16) | top-1 accuracy: 66.4 -> 60.8 | Hurt |
| halve number of prototypes from 1024 to 512 (ablation) | MSN ViT-B/16 · ImageNet-1K 1% labels | top-1 accuracy: 69.5 -> 67.6 | Hurt |
Closest papers by meaning
- Siamese Image Modeling for Self-Supervised Vision Representation Learning (CVPR 2023 · similarity 0.91)
- Siamese Masked Autoencoders (NeurIPS 2023 · similarity 0.89)
- Exploring Simple Siamese Representation Learning (CVPR 2021 · similarity 0.88)
- Masked Autoencoders Are Scalable Vision Learners (CVPR 2022 · similarity 0.87)
- Exploring Target Representations for Masked Autoencoders (ICLR 2024 · similarity 0.85)
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.85)
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.
