GATSBI: Generative Adversarial Training for Simulation-Based Inference

Download

Abstract

Simulation-based inference (SBI) enables Bayesian inference for models where evaluating the likelihood is intractable but simulating from the model is possible. Many SBI methods approximate the posterior distribution using neural density estimators, but training these estimators requires a large number of simulations. We introduce GATSBI (Generative Adversarial Training for Simulation-Based Inference), which leverages generative adversarial networks (GANs) to improve the sample efficiency of posterior estimation. GATSBI trains a discriminator to distinguish between samples from the true posterior and samples from a learned generator, while simultaneously training the generator to approximate the posterior. This adversarial training objective provides a more direct learning signal compared to maximum likelihood estimation, leading to improved performance with fewer simulations. We demonstrate GATSBI’s effectiveness on benchmark problems from neuroscience and epidemiology, showing that it achieves comparable or better accuracy than existing methods while requiring significantly fewer simulations.

Citation

@inproceedings{ramesh2022gatsbi,
  title={GATSBI: Generative Adversarial Training for Simulation-Based Inference},
  author={Ramesh, Poornima and Lueckmann, Jan-Matthis and Boelts, Jan and Tejero-Cantero, {\'A}lvaro and Greenberg, David S and Macke, Jakob H},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2022}
}