Fast and Credible Inference with Truncated Marginal Neural Ratio Estimation
Alex Cole, University of AmsterdamAbstract:
Across fields, scientific models are computationally implemented via parametric stochastic simulators. However, solving the “inverse problem” and constraining model parameters from data is a challenge in this context. Recently, the field of simulation-based inference has made great strides thanks to deep learning methods. I will outline a new method in simulation-based inference called Truncated Marginal Neural Ratio Estimation (TMNRE). TMNRE is (i) simulation-efficient, actively identifying the relevant regime of parameter space without sacrificing amortization (ii) scalable to high-dimensional data and model parameter spaces (iii) trustworthy, in the sense that statistical consistency tests beyond those available to e.g. MCMC can be rapidly performed. I will show examples of these benefits in the context of cosmological inference. I will also describe our development of a user-friendly and general package for TMNRE called swyft.
Implementation of TMNRE available at https://github.com/undark-lab/swyft. Talk based on https://arxiv.org/abs/2011.13951 (NeurIPS ML4PS ’20), https://arxiv.org/abs/2107.01214 (NeurIPS ’21), https://arxiv.org/abs/2111.08030.