Reproduce “A Generalized Theory of Mixup for Structure-Preserving Synthetic Data” 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
Chungpa Lee, Jongho Im, Joseph H.T. Kim
Researchers from Yonsei University developed a generalized theoretical framework for mixup, demonstrating that standard mixup methods inherently reduce data variance and distort statistical structure. Their proposed Expanded Beta (EpBeta) distribution preserves the original data's mean, variance, and covariance, leading to more accurate statistical inference and preventing model collapse in multi-generational synthetic data pipelines.
Key insights
- Standard mixup methods with weights restricted to the [0,1] interval inherently reduce data variance unless weights are exclusively 0 or 1, fundamentally altering the original data's statistical structure.
- Precise theoretical conditions exist for preserving the mean, variance, and covariance of continuous variables, and for bounding the differences in conditional moments of continuous variables conditioned on categorical ones.
- By allowing mixup weights to be drawn from a distribution with support beyond [0,1], such as the proposed EpBeta, it is possible to generate synthetic data that preserves the key statistical properties of the original dataset.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.