Reproduce “Stochastic Localization via Iterative Posterior Sampling” 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
Louis Grenioux, Maxence Noble, Marylou Gabrié, Alain Oliviero Durmus
This research introduces Stochastic Localization via Iterative Posterior Sampling (SLIPS), a learning-free algorithm for efficiently sampling from high-dimensional, multi-modal probability distributions. SLIPS leverages a generalized stochastic localization framework combined with MCMC-based denoiser estimation, showcasing superior accuracy in estimating mode weights and capturing local distribution properties compared to existing methods on challenging benchmarks.
Key insights
- The 'duality of log-concavity' reveals that the marginal observation distribution is log-concave at small times, facilitating initialization, while the posterior distribution is log-concave at larger times, enabling efficient MCMC denoiser estimation.
- Flexible 'denoising schedules' within the generalized SL framework allow for systematic control over the localization rate, significantly influencing the efficiency and accuracy of sampling.
- Utilizing MCMC for the estimation of the SL denoiser provides a learning-free, unbiased, and practical approach for density-based sampling, addressing limitations of neural network-based methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.