Reproduce “Towards Combating Frequency Simplicity-biased Learning for Domain Generalization” 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

Xilin He, Jingyu Hu, Qinliang Lin, Cheng Luo, Weicheng Xie, Siyang Song, Muhammad Haris Khan, Linlin Shen

Researchers from Shenzhen University and collaborators developed two adversarial frequency-domain data augmentation modules to combat deep neural networks' tendency to learn "frequency shortcuts" for Domain Generalization. Their method significantly reduces reliance on these shortcuts while achieving superior generalization performance across image classification and person re-identification tasks.

Key insights

  • Neural network learning behavior is inherently biased towards dominant frequencies within a dataset's statistical structure, allowing this bias to be manipulated through strategic frequency-domain data augmentation.
  • The observed generalization improvements from many prior data augmentation techniques can be a "hallucination," as they often lead to an increased reliance on non-semantic frequency shortcuts.
  • A holistic approach combining both low-frequency perturbation and adaptive frequency dropout is essential for effectively combating frequency shortcuts, as individual modules may inadvertently shift reliance to other frequency bands.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.