Tokenised Flow Matching for Hierarchical Simulation Based Inference

2026-04-22 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

Simulation Based Inference (SBI) faces a significant bottleneck due to the high cost of simulator evaluations, particularly in hierarchical models with shared global parameters and exchangeable site-level parameters. Existing hierarchical SBI methods improve efficiency by factorizing the posterior but still require multi-site simulations per training sample. A new approach, likelihood factorisation (LF), trains from single-site simulations by learning a per-site neural simulator surrogate and then assembling synthetic multi-site observations for amortized inference of the full hierarchical posterior. Building on this, Tokenised Flow Matching for Posterior Estimation (TFMPE) is proposed, which uses a tokenised flow matching technique to support function-valued observations via LF. The authors introduce a new benchmark for hierarchical SBI and validate TFMPE on it, as well as on infectious disease and computational fluid dynamics models, demonstrating well-calibrated posteriors and reduced computational cost.

Key takeaway

For AI Scientists and Machine Learning Engineers working with hierarchical Simulation Based Inference, consider adopting Tokenised Flow Matching for Posterior Estimation (TFMPE) to significantly reduce computational costs. Your teams can achieve well-calibrated posteriors even with complex, function-valued observations by leveraging likelihood factorisation and single-site simulations, thereby accelerating model development and deployment in resource-intensive domains like infectious disease modeling or computational fluid dynamics.

Key insights

Likelihood factorisation and tokenised flow matching significantly reduce simulation costs in hierarchical SBI.

Principles

• Exploit hierarchical structure for efficiency
• Factorize likelihood for single-site training

Method

TFMPE learns per-site neural surrogates via likelihood factorisation, then uses tokenised flow matching to support function-valued observations for hierarchical posterior estimation.

In practice

• Apply LF for single-site simulation training
• Use TFMPE for function-valued observations

Topics

• Simulation Based Inference
• Hierarchical Models
• Likelihood Factorisation
• Tokenised Flow Matching
• Posterior Estimation

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Context-Aware Hierarchical Bayesian Modeling of IVF Laboratory Environmental Conditions — Context-aware temporal features from IVF lab environments, modeled hierarchically, significantly improve pregnancy rate prediction and are transferable.
• Probabilistic Signature Inversion: Learning Conditional Distributions from Truncated Signatures — Truncated signature inversion is reframed as learning a path's conditional distribution using flow matching, quantifying irreducible uncertainty.
• Flow Matching for Efficient and Scalable Data Assimilation — EnFF uses flow matching to create a training-free, scalable data assimilation framework with improved cost-accuracy.
• Robust Simulation-Based Inference Under Missing Data via Neural Processes — RISE jointly learns data imputation and parameter inference in SBI to robustly handle missing data and prevent biased posteriors.
• Neural Galerkin Normalizing Flows for Bayesian Inference of Diffusions with Inaccessible Boundaries — Neural Galerkin Normalizing Flows efficiently learn diffusion transition densities for Bayesian inference, bypassing complex analytical challenges.
• Topological Flow Matching — Topological Flow Matching enhances generative models by integrating domain topology into flow matching for structured data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.