Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows

About

We present Sequential Neural Likelihood (SNL), a new method for Bayesian inference in simulator models, where the likelihood is intractable but simulating data from the model is possible. SNL trains an autoregressive flow on simulated data in order to learn a model of the likelihood in the region of high posterior density. A sequential training procedure guides simulations and reduces simulation cost by orders of magnitude. We show that SNL is more robust, more accurate and requires less tuning than related neural-based methods, and we discuss diagnostics for assessing calibration, convergence and goodness-of-fit.

George Papamakarios, David C. Sterratt, Iain Murray · 2018

Related benchmarks

Task	Dataset	Result
Simulation-Based Inference	Hierarchical Two Moons	l-C2ST0.241	55
Simulation-Based Inference	Hierarchical Gaussian Linear	l-C2ST0.00104	55
Simulation-Based Inference	Hierarchical Gaussian Linear Uniform	l-C2ST0.00174	55
Simulation-Based Inference	Hierarchical Gaussian Mixture	l-C2ST0.0587	55
Simulation-Based Inference	Hierarchical SLCP	l-C2ST23.6	54
Simulation-Based Inference	Hierarchical SIR	l-C2ST0.000573	53
Bayesian Inference	Bayesian Inverse Problems Task 5 1.0	C2ST66.8	17
Bayesian Inference	Bayesian Inverse Problems Task 4 1.0	C2ST73.1	17
Bayesian Inference	Bayesian Inverse Problems Task 1 1.0	C2ST51.9	17
Bayesian Inference	Bayesian Inverse Problems Task 3 1.0	C2ST69.9	17