Reproduce “Global Filter Networks for Image Classification” 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
Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, Jie Zhou
Researchers from Tsinghua University developed Global Filter Networks (GFNet) which utilize Fast Fourier Transforms for token mixing, achieving log-linear computational complexity (O(L log L)) in place of the quadratic complexity found in self-attention and spatial MLPs. The models demonstrate competitive or superior ImageNet classification accuracy, with GFNet-S achieving 80.0% top-1 accuracy compared to DeiT-S's 79.8%, and exhibit better scalability and robustness for high-resolution images.
Key insights
- Token mixing performed in the frequency domain via learnable filters can capture global spatial interactions with log-linear computational complexity, addressing the quadratic scaling of self-attention.
- Global Filter Networks (GFNet) offer a conceptually simple deep learning architecture that combines the low inductive bias and global context understanding of transformers with high computational efficiency.
- The inherent resolution-agnostic nature of FFT allows GFNet to adapt flexibly to varying input resolutions through filter interpolation, simplifying deployment across different image sizes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.