Reproduce “Kernel Stein Discrepancy Descent” 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
Anna Korba, Pierre-Cyril Aubin-Frankowski, Szymon Majewski, Pierre Ablin
KSD Descent turns sampling from an unnormalized target into a finite-particle optimization problem. It minimizes a computable score-based discrepancy using the target’s score without its normalization constant or target samples. On a two-dimensional Gaussian task, its particles gather around the mode similarly to SVGD. However, symmetry can trap particles, and on Bayesian ICA it performs barely better than random output.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.