SBEB method enables Bayesian inference without explicit likelihoods
New technique combines nonparametric EB with simulation-based inference for better accuracy.
Deep Dive
Shen, Cai, Zhang, and Blei introduce Simulation-Based Empirical Bayes (SBEB), a method that performs empirical Bayes inference when likelihoods are only available through simulators. SBEB uses observed data, simulator samples, and an amortized inference network to iteratively refine the EB prior. Tests on scientific simulators and real-world data show SBEB improves accuracy over standard SBI with a fixed prior.
Key Points
- SBEB connects nonparametric Empirical Bayes to simulation-based inference without requiring tractable likelihoods.
- It uses an amortized inference network to iteratively refine the EB prior using observed data and simulator samples.
- On scientific simulators and real-world data, SBEB achieves higher accuracy than standard SBI with a fixed prior.
Why It Matters
Makes powerful Bayesian inference feasible where only simulators exist, improving accuracy in complex scientific domains.