CONNECT: A neural network based framework for emulating cosmological observables and cosmological parameter inference

TL;DR

CONNECT is a neural network framework that emulates cosmological model computations, drastically reducing the computational cost of parameter inference by requiring fewer training evaluations and enabling faster likelihood calculations.

Contribution

It introduces a novel training algorithm for neural networks that reduces the number of expensive model evaluations needed for cosmological inference.

Findings

Posteriors differ from traditional methods by less than 0.1 standard deviations.

Training reduces model evaluations by two orders of magnitude.

Inference speed is significantly increased, dominated by likelihood calculations.

Abstract

Bayesian parameter inference is an essential tool in modern cosmology, and typically requires the calculation of $10^5$--$10^6$ theoretical models for each inference of model parameters for a given dataset combination. Computing these models by solving the linearised Einstein-Boltzmann system usually takes tens of CPU core-seconds per model, making the entire process very computationally expensive. In this paper we present \textsc{connect}, a neural network framework emulating \textsc{class} computations as an easy-to-use plug-in for the popular sampler \textsc{MontePython}. \textsc{connect} uses an iteratively trained neural network which emulates the observables usually computed by \textsc{class}. The training data is generated using \textsc{class}, but using a novel algorithm for generating favourable points in parameter space for training data, the required number of \textsc{class}-evaluations can be reduced by two orders of magnitude compared to a traditional inference run. Once \textsc{connect} has been trained for a given model, no additional training is required for different dataset combinations, making \textsc{connect} many orders of magnitude faster than \textsc{class} (and making the inference process entirely dominated by the speed of the likelihood calculation). For the models investigated in this paper we find that cosmological parameter inference run with \textsc{connect} produces posteriors which differ from the posteriors derived using \textsc{class} by typically less than $0.01$--$0.1$ standard deviations for all parameters. We also stress that the training data can be produced in parallel, making efficient use of all available compute resources. The \textsc{connect} code is publicly available for download at \url{https://github.com/AarhusCosmology}.