Research map / Architectures and efficiency / Attention and transformers
From Linearity to Non-Linearity: How Masked Autoencoders Capture Spatial Correlations
Anthony Bisulco, Rahul Ramesh, Randall Balestriero, Pratik Chaudhari
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 |
|---|---|---|---|
| increase masking ratio from 0.1 to 0.9 (ablation) | ViT MAE · CIFAR-10 pretraining time | training time (hours): 10 hours -> 4 hours | Improved |
| use a 1-layer decoder instead of the best decoder depth (ablation) | ViT MAE · CIFAR-10 fine-tuning | accuracy (%): 95.26% accuracy, only 0.30% worse than best model | No clear effect |
| fine-tune only the last Transformer block (freeze the rest) (ablation) | ViT MAE · CIFAR-10 fine-tuning | accuracy (%): within 2% of full fine-tuning accuracy, nearly 2x speedup | No clear effect |
| add MAE pretraining versus supervised-only training (ablation) | ViT · CIFAR-10 | accuracy (%): 83-84% plateau without pretraining | Improved |
| increase number of encoder layers (ablation) | ViT MAE · CIFAR-10 linear probe | linear-probe accuracy | Improved |
| increase number of decoder layers beyond 2-4 (ablation) | ViT MAE · CIFAR-10 linear probe | linear-probe accuracy: optimal decoder contains 2-4 layers | No clear effect |
| increase masking ratio (ablation) | ViT MAE · CIFAR-10 linear probe | linear-probe accuracy | No clear effect |
| increase patch size (ablation) | ViT MAE · CIFAR-10 linear probe | linear-probe accuracy | Hurt |
Closest papers by meaning
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.89)
- R-MAE: Regions Meet Masked Autoencoders (ICLR 2024 · similarity 0.88)
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.88)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.88)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.87)
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.87)
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.
