Scalable Diffusion SBI for Compositional Inference under Simulator Misspecification

TL;DR

This paper introduces scalable diffusion-based methods for compositional inference under simulator misspecification. The authors develop Hierarchical Blockwise Diffusion Sampling and path-regularized fine-tuning to improve inference accuracy. These methods enable variable observation sets and groupings without retraining.

Why It Matters

The proposed methods address challenges in simulation-based inference, particularly when dealing with heterogeneous observations and misspecified simulators. This has practical significance in fields like systems biology, where accurate modeling of complex systems is crucial. The methods can improve the accuracy of mechanistic models, leading to better decision-making.

Key Findings

• Hierarchical Blockwise Diffusion Sampling (HBDS) for compositional inference
• Path-regularized fine-tuning for addressing simulator misspecification
• Improved posterior-predictive accuracy in a mechanistic Bone Morphogenetic Protein signaling model