Research map / Label-efficient and robust learning / Self-supervised and contrastive
Simple Filtering Improves Masked Autoencoders
Takumi Kobayashi
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 |
|---|---|---|---|
| replace random masking with 2x2 max-pool filtered masking (ablation) | MAE ViT-S · ImageNet-100 | accuracy (%): 68.99 -> 72.07 | Improved |
| replace random masking with 2x2 max-pool filtered masking (ablation) | MAE ViT · Cifar-10 | accuracy (%): 79.18 -> 82.16 | Improved |
| use 4x4 max-pool kernel instead of 2x2 (ablation) | MAE ViT-S · ImageNet-100 | accuracy (%): 72.07 -> 69.93 | Hurt |
| replace random masking with grid masking (ablation) | MAE ViT-S · ImageNet-100 | accuracy (%): 68.99 -> 58.82 | Hurt |
| replace random masking with Laplacian filtered masking (ablation) | MAE ViT-S · ImageNet-100 | accuracy (%): 68.99 -> 66.68 | Hurt |
| smooth reconstruction target with Gaussian sigma=4 (ablation) | MAE ViT-S random masking · ImageNet-100 | accuracy (%): 68.99 -> 72.10 | Improved |
| smooth reconstruction target with Gaussian sigma=4 on top of max-pool masking (ablation) | MAE ViT-S · ImageNet-100 | accuracy (%): 72.07 -> 73.29 | Improved |
| sharpen target (unsharp masking) instead of smoothing (ablation) | MAE ViT-S · ImageNet-100 | accuracy (%): 72.07 -> 64.51 | Hurt |
| Gaussian smoothing instead of down-scaling target (ablation) | MAE ViT-S max-pool · ImageNet-100 | accuracy (%): 72.45 -> 73.29 | Improved |
| smoothing with sigma=8 vs max-pool alone (ablation) | MAE ViT-S max-pool · ImageNet-100 | accuracy (%): 72.07 -> 72.07 | No clear effect |
| use filtering method instead of frequency losses (ablation) | MAE ViT-S · ImageNet-100 | accuracy (%): 70.94 -> 73.29 | Improved |
| apply method to MAE ViT-B at 1600 epochs | MAE ViT-B · ImageNet-1K linear probe | accuracy (%): 66.71 -> 72.55 | Improved |
Closest papers by meaning
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.93)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.92)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.91)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.91)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.90)
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.90)
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.
