Andrew Williamson
Gather.town id
MLA07
Poster Title
Lightning-Fast Gravitational-Wave Parameter Inference Through Neural Amortization
Institution
University of Portsmouth
Abstract (short summary)
Gravitational waves from compact binaries measured by the LIGO and Virgo detectors are routinely analyzed using Markov Chain Monte Carlo sampling algorithms. Because the evaluation of the likelihood function requires evaluating millions of waveform models that link between signal shapes and the source parameters, running Markov chains until convergence is typically expensive and requires days of computation. In this work, we provide a proof of concept that demonstrates how the latest advances in neural simulation-based inference can speed up the inference time by up to three orders of magnitude -- from days to minutes -- without impairing the performance. Our approach is based on a convolutional neural network modeling the likelihood-to-evidence ratio and entirely amortizes the computation of the posterior. We find that our model correctly estimates credible intervals for the parameters of simulated gravitational waves.
Plain text (extended) Summary
The inference of gravitational wave source parameters is typically achieved via stochastic sampling methods, making it a slow and computationally expensive process. As the rate of detections grows, the associated burden could represent a significant obstacle to maximising science outcomes. We demonstrate that neural simulation-based inference can provide a factor of roughly 1000 speed-up. We consider precessing binary black hole systems, generating our dataset by adding model waveforms to Gaussian noise to mimick Advanced LIGO data. We high-pass filter and whiten this data before passing it to the neurtal network. We employ neural amortisation, approximating the likelihood-to-evidence ratio by training a classifier on: data and the source parameters used to generate the data (positive classification); and data with source parameters different from those used to generate the data (negative). A network may be trained on any subset of parameters of interest, with all others treated as nuisance parameters that are marginalised out. Thus the training procedure bears the computational cost, producing neural networks that produce marginal posteriors as output in 1 minute or less. Testing shows these results to be robust and comparable to slower stochastic methods. Work is ongoing to further develop this into a more efficient software package that can be deployed on real data.
URL
andrew.williamson@port.ac.uk
Poster file