Fast, lightweight and semi-amortised simulation-based inference

Abstract: Bayesian inference for complex models with an intractable likelihood can be tackled using algorithms performing many calls to computer simulators. These approaches are collectively known as "simulation-based inference" (SBI). Recent SBI methods use neural-conditional-estimation, that is neural networks are employed to provide approximations to the likelihood function or the posterior distribution of model parameters. While neural-based posterior and likelihood estimation methods have produced exceptionally flexible inference strategies, these can be computationally intensive to run and have a non-negligible impact on energy expenditure and memory requirements. In this work, rather than using neural networks we propose more "frugal" strategies that display state-of-art inference quality, while being able to run with limited resources, being much faster to train and exhibiting a much smaller computational footprint. We investigate structured mixtures of probability distributions and design a new SBI method named Sequential Mixture Posterior and Likelihood Estimation (SeMPLE). SeMPLE learns closed-form approximations for both the posterior $p(θ|y)$ and the likelihood $p(y|θ)$ from the same training data, using Gaussian mixture models that can be efficiently learned. We show favorable results for a variety of stochastic models (including SDEs and Markov jump processes), also in presence of multimodal posteriors.

The talk will be approachable for the uninitiated audience, while novel results will be of interest for the experienced audience.

Joint work with Henrik Häggström, Pedro L. C. Rodrigues, Geoffroy Oudoumanessah and Florence Forbes, arxiv.org/abs/2403.07454

machine learningstatistics theory

Audience: advanced learners

( paper )

---

Gothenburg statistics seminar

Series comments: Gothenburg statistics seminar is open to the interested public, everybody is welcome. It usually takes place in MVL14 (http://maps.chalmers.se/#05137ad7-4d34-45e2-9d14-7f970517e2b60, see specific talk). Speakers are asked to prepare material for 35 minutes excluding questions from the audience.

---