Reproduce “Multifidelity Simulation-based Inference for Computationally Expensive Simulators” on OpenResearch

About this paper

Anastasia N. Krouglova, Hayden R. Johnson, Basile Confavreux, Michael Deistler, Pedro J. Gonçalves

The paper presents multifidelity neural posterior estimation for inferring ion-channel parameters from a neuron’s voltage response when detailed simulations are costly. It pretrains a neural posterior estimator on a cheaper, simplified simulator, then fine-tunes it with high-fidelity simulations. In a layer-5 pyramidal-cell task, 50 high-fidelity simulations yielded similar accuracy to NPE’s 1,000-simulation result, but this task-specific finding does not guarantee savings.

Key insights

• Pre-training neural density estimators on low-fidelity simulations effectively reduces the high-fidelity data budget needed for accurate posterior inference in computationally expensive simulators.
• Successful transfer learning in multifidelity SBI requires both high mutual information and strong representational coherence between the low and high-fidelity models.
• The proposed methods maintain performance even when low- and high-fidelity models have different numbers of parameters, though extreme parameter space discrepancies can diminish transfer benefits.

Summary by alphaXiv.