Amortised Multifidelity Neural Posterior Estimation via Stochastic Residual Correction

Neural posterior estimation (NPE) enables amortised Bayesian inference for models that are only available through simulation, but its practical applicability can be limited by the computational cost of high-fidelity simulators. Recent work in multifidelity NPE (MF-NPE) has shown that inexpensive low-fidelity simulations can be combined with limited high-fidelity simulations to reduce the computational cost of simulation-based inference. We consider the setting of coupled stochastic simulators, where low- and high-fidelity simulations share common stochastic variation, and propose SRC-MF-NPE, a two-stage stochastic residual correction approach that first learns the low-to-high-fidelity residual from coupled simulations and then uses the learned correction to generate high-fidelity training data for amortised NPE. Our preliminary results show that SRC-MF-NPE achieves competitive posterior inference compared with existing MF-NPE methods under a fixed computational budget.