Reproduce “Architecture-Agnostic Masked Image Modeling -- From ViT back to CNN” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Siyuan Li, Di Wu, Fang Wu, Zelin Zang, Stan.Z.Li
An architecture-agnostic Masked Image Modeling (A²MIM) framework unifies self-supervised pre-training for both Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs). It demonstrates that MIM works by fostering middle-order interactions among image patches, achieving competitive performance on ImageNet-1K and strong transfer learning capabilities across downstream tasks.
Key insights
- Masked Image Modeling (MIM) improves model robustness to input occlusions for both ViTs and CNNs, confirming it fosters robust feature extraction.
- MIM's effectiveness stems from teaching models to learn "better middle-order interactions" among image patches, critical for extracting complex features like shape and edge.
- Explicitly encouraging the learning of medium-frequency components in the Fourier domain can enhance the capture of these middle-order interactions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.