Congratulations to Compass PhD student, Dan Ward, who along with his supervisors has had their paper Robust Neural Posterior Estimation and Statistical Model Criticism accepted for publication in NeurIPS 2022.
Simulators are idealistic approximations of reality, meaning that there is often a disparity between simulations and observed data. We propose a novel method that allows identification of these discrepancies and facilitates robust fitting of simulators to observed data. To do this, we extend an existing approach for fitting simulators, neural posterior estimation (NPE) (Papamakarios, 2016; Lueckmann, 2017; Greenberg, 2019), with an assumed error model to explicitly model the “gap” between simulations and observed data. We name the approach robust neural posterior estimation (RNPE). Empirical results are used to demonstrate utility of this approach.
This work, which has been accepted for publication in NeurIPS 2022, was carried out as a collaboration between the University of Bristol and Improbable (https://www.improbable.io/), authored by Daniel Ward (Compass PhD student, University of Bristol), Patrick Cannon (Improbable), Mark Beaumont (Biological Sciences, University of Bristol), Matteo Fasiolo (Mathematics, University of Bristol) and Sebastian Schmon (Improbable; Mathematical Sciences, Durham University).
Papamakarios, G. and Murray, I., 2016. Fast ε-free inference of simulation models with Bayesian conditional density estimation. Advances in neural information processing systems, 29.
Lueckmann, J.M., Goncalves, P.J., Bassetto, G., Öcal, K., Nonnenmacher, M. and Macke, J.H., 2017. Flexible statistical inference for mechanistic models of neural dynamics. Advances in neural information processing systems, 30.
Greenberg, D., Nonnenmacher, M. and Macke, J., 2019, May. Automatic posterior transformation for likelihood-free inference. In International Conference on Machine Learning (pp. 2404-2414). PMLR.