Reproduce “Wasserstein Filtering: A Sample Selection Method for Robust Distribution Learning” 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
Yikai Xu, Zhao Chen, Jian Huang
Wasserstein Filtering introduces a sample selection framework to robustly recover the underlying clean distribution from adversarially contaminated data by identifying and removing samples that maximally distort the Wasserstein-1 distance of the empirical distribution. This model-agnostic preprocessing tool, developed by researchers from The Hong Kong Polytechnic University and Fudan University, achieves minimax optimal risk under a new Far Exclusion and Local Projection contamination model and demonstrated a 3-4 fold improvement in FID scores for robust generative learning.
Key insights
- Leveraging the Wasserstein-1 distance's sensitivity to the geometric characteristics of outliers allows WF to effectively identify and remove samples that cause significant distributional distortion, even if they are not traditionally considered 'extreme'.
- Framing robust distribution learning as a maximal distortion problem using optimal transport provides a model-agnostic sample selection mechanism, allowing for flexible preprocessing independent of the downstream learning task.
- The new FELP contamination model provides a strong theoretical basis for analyzing robust distribution estimation, demonstrating how WF can achieve minimax optimality by distinguishing 'far' and 'near' outliers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.