Bai and Zhi model a point's value as a context function that depends on which other points are already selected, rather than as a fixed scalar.
Top-k data scores under-cover latent modes; class balancing or coverage fixes selection
Bai and Zhi prove fixed per-point valuation is provably suboptimal under tight budgets, and show when label balancing or geometry-based coverage recovers the gap.
Academic
Yingjie Bai · Weiming Zhi
University of Sydney
Research Digest··3 min read
The authors analyze the common practice of selecting training data by top-k per-point value scores such as Data Shapley, influence, and leverage.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.