Reproduce “Sebra: Debiasing Through Self-Guided Bias Ranking” 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
Adarsh Kappiyath, Abhra Chaudhuri, Ajay Jaiswal, Ziquan Liu, Yunpeng Li, Xiatian Zhu, Lu Yin
SEBRA, developed by researchers from the University of Surrey, Fujitsu Research, University of Texas, and Queen Mary University of London, automatically identifies and ranks data samples based on their spuriosity without explicit bias annotations, leveraging training dynamics. This ranking is then used for robust model training through contrastive learning, leading to up to a 10% improvement on benchmark datasets and enabling the discovery of previously unknown biases.
Key insights
- Machine learning models tend to learn spurious features and correlations earlier in the training process due to their simplicity and consistency in the data.
- A "Hardness-Spuriosity Symmetry" exists, where samples that are 'harder' to learn contain less spurious correlation, and vice-versa, allowing spuriosity to be inferred from learning dynamics.
- Providing a fine-grained ranking of samples by their spuriosity, rather than just discrete biased/unbiased labels, is more effective for debiasing through contrastive learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.