Reproduce “Riemannian Laplace Approximation with the Fisher Metric” 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
Hanlin Yu, Marcelo Hartmann, Bernardo Williams, Mark Girolami, Arto Klami
Researchers from the University of Helsinki, University of Cambridge, and The Alan Turing Institute developed a refined Riemannian Laplace Approximation (RLA-F) that uses the Fisher Information Matrix as its metric, addressing the bias and underestimation issues found in prior RLA methods. This approach delivers more accurate posterior approximations and improved computational stability across various Bayesian inference tasks, often outperforming previous methods and achieving results comparable to MCMC.
Key insights
- The Monge metric, while computationally convenient for Christoffel symbol calculation, fundamentally distorts the parameter space, causing prior RLA to systematically underestimate uncertainty and be asymptotically biased.
- The Fisher Information Matrix provides a robust and statistically meaningful metric for the RLA framework, ensuring asymptotic exactness for Gaussian targets and even exactness for distributions diffeomorphic to Gaussians.
- Combining the Fisher metric with the Hausdorff MAP (which maximizes the posterior density under the Hausdorff measure) allows for more accurate and invariant posterior approximations, especially for complex and multimodal-like distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.