Reproduce “Generalized Bayesian Inference for Scientific Simulators via Amortized Cost Estimation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Richard Gao, Michael Deistler, Jakob H. Macke

ACE learns simulator-to-observation distance costs and uses them for generalized Bayesian inference without simulator calls during MCMC. Across ten Allen neuron recordings, its predictive outputs exceeded five spikes in 66% of cases at β=25 and 96% at β=100, compared with 35.6% for NPE. All recordings had at least 12 spikes, so this permissive threshold does not show full trace agreement.

Key insights

  • Generalized Bayesian Inference (GBI), which replaces likelihoods with flexible cost functions, offers a robust framework for inference, especially when scientific models are misspecified or deterministic.
  • A regression neural network can be effectively trained to approximate and amortize the GBI cost function, making it computationally efficient for diverse observations.
  • The ability to directly optimize for predictive quality through a cost function, rather than exact likelihood, leads to more meaningful parameter estimates for real-world, complex systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.