Reproduce “Diffusion posterior sampling for simulation-based inference in tall data settings” on OpenResearch

About this paper

Julia Linhart, Gabriel Victorino Cardoso, Alexandre Gramfort, Sylvain Le Corff, Pedro L. C. Rodrigues

This work introduces a diffusion posterior sampling approach, primarily through the GAUSS algorithm, for simulation-based inference in "tall data" settings. The method overcomes the computational and stability issues of existing compositional methods by enabling deterministic score-based sampling, achieving up to 2.5 times faster inference and improved accuracy across various benchmarks, including a complex neural mass model.

Key insights

• Explicitly modeling the diffusion process for the tall data posterior enables the use of deterministic score-based samplers, offering a more stable and efficient alternative to Langevin dynamics in compositional SBI.
• Approximating backward transition kernels using a second-order Gaussian framework effectively handles intractable correction terms that previously necessitated iterative MCMC methods.
• The principled incorporation of precision weighting in the composed score contributes to superior accuracy and robustness for aggregated observations, especially as the number of observations increases.

Summary by alphaXiv.