Reproduce “Distilling Long-tailed Datasets” 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
Zhenghao Zhao, Haoxuan Wang, Yuzhang Shang, Kai Wang, Yan Yan
Researchers at the University of Illinois Chicago introduce Distribution-agnostic Matching with Expert Decoupling, a method for long-tailed dataset distillation that generates small, balanced synthetic datasets from imbalanced sources. The approach achieves up to 10.6% higher accuracy on CIFAR-10-LT with an imbalance factor of 200 compared to prior methods and shows improved cross-architecture generalizability.
Key insights
- Distilling a long-tailed dataset into a *balanced* synthetic dataset is more effective than distilling it into another long-tailed synthetic dataset for improving overall model performance.
- Data imbalance, not just the total sample count, significantly hinders effective dataset distillation, highlighting the need for specialized methods for long-tailed distributions.
- Decoupling expert knowledge by training separate experts for representation and classification, combined with distribution-agnostic student training, is crucial for preserving tail-class information and achieving robust distillation from imbalanced data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.