Reproduce “Bridged SBI: Correcting Biased Low-Fidelity Posteriors for Cost-Efficient High-Fidelity Inference” on OpenResearch

About this paper

Gahee Kim, Yuki Kadokawa, Sandro M. Alcantara Tacora, Taro Abe, Daisuke Endo, Genki Yamauchi, Takeshi Hashimoto, Takamitsu Matsubara

Bridged SBI presents a framework to perform cost-efficient, high-fidelity parameter inference in computationally expensive simulators by explicitly correcting biases from low-fidelity models. It learns a residual bridge to transport biased low-fidelity posterior samples towards high-fidelity consistent regions, improving accuracy with limited high-fidelity simulations.

Key insights

• Directly relying on biased low-fidelity posteriors in sequential multi-fidelity SBI can lead to irreducible bias and miscoverage of high-fidelity consistent regions.
• Explicitly learning a local residual bridge can effectively correct the systematic discrepancy between low-fidelity and high-fidelity simulators, even when LF posteriors are significantly biased.
• The probabilistic bridge allows for recovery of high-fidelity posterior mass that was underrepresented by the low-fidelity posterior, enabling more accurate inference with limited high-fidelity data.

Summary by alphaXiv.