Reproduce “LAVA: Data Valuation without Pre-Specified Learning Algorithms” 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
Hoang Anh Just, Feiyang Kang, Jiachen T. Wang, Yi Zeng, Myeongseob Ko, Ming Jin, Ruoxi Jia
LAVA proposes a learning-agnostic framework for data valuation that quantifies data contribution using class-wise Wasserstein distance. This approach achieves orders of magnitude faster computation and demonstrates superior efficacy in detecting various types of low-quality data compared to learning-dependent methods.
Key insights
- A theoretically grounded class-wise Wasserstein distance can serve as an effective proxy for a model's validation performance, independent of the specific learning algorithm used.
- Individual data point values, crucial for tasks like data selection and filtering, can be extracted efficiently from the dual solutions of optimal transport, bypassing expensive retrainings.
- A distribution-based, learning-agnostic approach to data valuation can surprisingly outperform learning-dependent methods in various data quality detection tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.