Reproduce “EMS Coreset: An Efficient Expectation-Maximization Algorithm for Sinkhorn Coreset” 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
Haoyun Yin, Chuanhui Liu, Xiao Wang
EMS Coreset recasts weighted Sinkhorn coreset construction as a variational EM procedure, replacing repeated transport-plan optimization with soft assignments and closed-form updates for representative locations and weights. On Gaussian, MNIST, and Fashion-MNIST benchmarks, it achieved distributional distances close to gradient-based WCSL while reducing runtime by roughly two to three orders of magnitude, and it also improved SHAP background approximation on Census and Diabetes data.
Key insights
- Learning coreset weights alongside locations lets the discrete measure account for data-density differences that uniform k-means cannot capture.
- The regularization parameter controls a fidelity–stability trade-off: larger values produce smoother, more perturbation-resistant assignments, while smaller values approach hard allocation and are faster.
- EMS provides block-coordinate objective monotonicity for its M-step, but the overall nonconvex procedure is not guaranteed to reach a global optimum.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.