Reproduce “CoreFlow: Low-Rank Matrix Generative Models” on OpenResearch

About this paper

Dongze Wu, Linglingzhi Zhu, Yao Xie

Researchers at Georgia Institute of Technology introduce CoreFlow, a two-stage generative model that learns shared low-rank geometry for matrix-valued data and then trains a continuous normalizing flow within this reduced-dimensional core space. CoreFlow demonstrates improved generative quality, spectral fidelity, and computational efficiency over ambient-space models, especially when handling high-dimensional, limited, or incomplete training data.

Key insights

• Many high-dimensional matrix distributions possess an intrinsic low-rank structure, implying a much smaller effective dimensionality than their ambient space.
• Decomposing matrix generation into learning shared low-rank subspaces and then modeling a low-dimensional core matrix significantly reduces generative complexity while preserving structural integrity.
• It is feasible to robustly learn these shared low-rank subspaces and consistent core representations even from incomplete training matrices via masked optimization and iterative imputation.

Summary by alphaXiv.