Tokenised Flow Matching Improves Hierarchical Simulation Efficiency
The recently introduced Tokenised Flow Matching for Posterior Estimation (TFMPE) technique lowers simulation expenses in hierarchical simulation-based inference by utilizing single-site simulations through likelihood factorisation. This method develops a neural surrogate for each site, subsequently creating synthetic multi-site observations to streamline inference across the complete hierarchical posterior. Additionally, TFMPE accommodates function-valued observations. To facilitate thorough assessment, the authors present a benchmark for hierarchical SBI. This research has been made available on arXiv.
Key facts
- TFMPE uses tokenised flow matching for posterior estimation.
- Likelihood factorisation enables training from single-site simulations.
- A per-site neural surrogate of the simulator is learned.
- Synthetic multi-site observations are assembled for amortised inference.
- The method supports function-valued observations.
- A benchmark for hierarchical SBI is introduced.
- The paper is available on arXiv with ID 2604.20723.
- The approach aims to reduce the cost of simulator evaluations.
Entities
Institutions
- arXiv