James Pearson
Gather.town id
MLA20
Poster Title
Strong Lensing with Bayesian Neural Networks
Institution
The University of Nottingham
Abstract (short summary)
Strong galaxy-galaxy gravitational lensing is the distortion of the paths of light rays from a background galaxy into arcs or rings as viewed from Earth, caused by the gravitational field of an intervening foreground lens galaxy. The vast quantity of strong lenses expected by future large-scale surveys necessitates the development of automated methods to efficiently model their mass profiles. For this purpose, we train an approximate Bayesian convolutional neural network (CNN) to predict mass profile parameters and associated uncertainties, and compare its accuracy to that of conventional parametric modelling. In addition, we present a method for combining the CNN with conventional modelling in an automated fashion, where the CNN provides initial priors on the latter's parameters. These methods are tested on a range of increasingly complex lensing systems, from standard smooth parametric mass and light profiles to images containing hydrodynamical EAGLE galaxies, Hubble Ultra Deep Field source galaxies and the inclusion of foreground mass structures. While the CNN is undoubtedly the fastest modelling method, the combination of the two increases the speed of conventional modelling, especially when priors include CNN-predicted uncertainties. This, combined with significantly improved accuracy, highlights the benefits one can obtain through combining neural networks with conventional techniques in order to achieve an efficient automated modelling approach.
Plain text (extended) Summary
Strong Lensing with Bayesian Neural Networks
James Pearson, Jacob Maresca, Nan Li, Simon Dye
Pearson, et al. (2021). doi:10.1093/mnras/stab1547
Strong Galaxy-Scale Gravitational Lensing
This is the distortion of the paths of light rays from a background galaxy into arcs or rings as viewed from Earth, caused by the gravitational field of an intervening foreground galaxy (the ‘lens’).
Lensing provides a way of investigating the properties of distant galaxies and the early Universe, from constraining the lensing galaxy’s dark matter to providing a magnified view of the high-redshift source galaxy. But this requires accurate modelling of the lens' mass profile, usually through slow but accurate parametric techniques to determine parameters of the mass profile (Einstein radius, orientation & axis ratio).
Project Overview
Upcoming surveys such as Euclid and LSST will generate tens of thousands of images containing lensing systems, so an efficient automated modelling method is needed to cope with such a large data set. We utilise machine learning, training an approximate Bayesian convolutional neural network (CNN) to estimate lens mass profile parameters for increasingly complex Euclid-style images and comparing this to conventional parameter-fitting techniques.
Conventional Lens Modelling – PyAutoLens
PyAutoLens is an example of conventional modelling, where an automated process adjusts parameters of a mass profile to best fit the observed image. However, this requires manually-set initial ‘guess’ values (priors) and a large amount of time and computing power.
Convolutional Neural Networks (CNNs)
Just as the brain is made up of interconnected neurons, a neural network consists of interconnected layers of nodes. CNNs are a subset of neural networks that have grid-like layers that apply filters to extract information. CNNs can be improved through training on tens of thousands of images, which must be simulated as not enough images of real lenses exist.
Comparing & combining with conventional fitting
The CNN was trained on 100,000 complex images generated to resemble expected observations by Euclid (VIS band). As well as parameter values, it’s approximate Bayesian formalism allowed the CNN to predict uncertainties on these values.
We compared the CNN to PyAutoLens for increasingly complex test sets generated using the software PICS, from smooth light & mass profiles (first test set) to images with EAGLE simulation lenses, real HUDF sources and extra line-of-sight structures (last test set). We also combined the two techniques, using CNN predicted values & uncertainties as priors for PyAutoLens (PyAL).
Accuracy
(Graphs: Mass model parameters are accurately predicted by the CNN for the first test set, but accuracies of all methods drop when testing on the last test set. The combination method incorporating CNN-predicted uncertainties reduces errors by 37-44% for the former and 15-34% for the latter.)
Overall, CNN errors were 19 ± 22% lower than PyAutoLens’ blind modelling. The combination method instead achieved 27 ± 11% lower errors, reduced further to 37 ± 11% when incorporating CNN-predicted uncertainties into the priors.
Modelling Times
(Graphs: The time taken to model lenses is reduced by incorporating CNN predictions. CNN times are not included as they are less than a second!)
Incorporating CNN predictions makes PyAutoLens more consistent, and compared to PyAL (blind) modelling speed is increased by a mean factor of 1.73x and 1.19x for the combination with and without incorporation of CNN-predicted uncertainties.
Summary
The CNN can accurately measure mass profile parameters for Euclid-style images much more rapidly than conventional modelling. CNN accuracy equals or exceeds an automated PyAutoLens while the combination method significantly improves upon both, especially when including CNN-predicted uncertainties which additionally increase PyAutoLens’ modelling speed. Hence, combining CNNs with conventional approaches is a promising method that for automated lens modelling can potentially outperform either one separately.
James Pearson, Jacob Maresca, Nan Li, Simon Dye
Pearson, et al. (2021). doi:10.1093/mnras/stab1547
Strong Galaxy-Scale Gravitational Lensing
This is the distortion of the paths of light rays from a background galaxy into arcs or rings as viewed from Earth, caused by the gravitational field of an intervening foreground galaxy (the ‘lens’).
Lensing provides a way of investigating the properties of distant galaxies and the early Universe, from constraining the lensing galaxy’s dark matter to providing a magnified view of the high-redshift source galaxy. But this requires accurate modelling of the lens' mass profile, usually through slow but accurate parametric techniques to determine parameters of the mass profile (Einstein radius, orientation & axis ratio).
Project Overview
Upcoming surveys such as Euclid and LSST will generate tens of thousands of images containing lensing systems, so an efficient automated modelling method is needed to cope with such a large data set. We utilise machine learning, training an approximate Bayesian convolutional neural network (CNN) to estimate lens mass profile parameters for increasingly complex Euclid-style images and comparing this to conventional parameter-fitting techniques.
Conventional Lens Modelling – PyAutoLens
PyAutoLens is an example of conventional modelling, where an automated process adjusts parameters of a mass profile to best fit the observed image. However, this requires manually-set initial ‘guess’ values (priors) and a large amount of time and computing power.
Convolutional Neural Networks (CNNs)
Just as the brain is made up of interconnected neurons, a neural network consists of interconnected layers of nodes. CNNs are a subset of neural networks that have grid-like layers that apply filters to extract information. CNNs can be improved through training on tens of thousands of images, which must be simulated as not enough images of real lenses exist.
Comparing & combining with conventional fitting
The CNN was trained on 100,000 complex images generated to resemble expected observations by Euclid (VIS band). As well as parameter values, it’s approximate Bayesian formalism allowed the CNN to predict uncertainties on these values.
We compared the CNN to PyAutoLens for increasingly complex test sets generated using the software PICS, from smooth light & mass profiles (first test set) to images with EAGLE simulation lenses, real HUDF sources and extra line-of-sight structures (last test set). We also combined the two techniques, using CNN predicted values & uncertainties as priors for PyAutoLens (PyAL).
Accuracy
(Graphs: Mass model parameters are accurately predicted by the CNN for the first test set, but accuracies of all methods drop when testing on the last test set. The combination method incorporating CNN-predicted uncertainties reduces errors by 37-44% for the former and 15-34% for the latter.)
Overall, CNN errors were 19 ± 22% lower than PyAutoLens’ blind modelling. The combination method instead achieved 27 ± 11% lower errors, reduced further to 37 ± 11% when incorporating CNN-predicted uncertainties into the priors.
Modelling Times
(Graphs: The time taken to model lenses is reduced by incorporating CNN predictions. CNN times are not included as they are less than a second!)
Incorporating CNN predictions makes PyAutoLens more consistent, and compared to PyAL (blind) modelling speed is increased by a mean factor of 1.73x and 1.19x for the combination with and without incorporation of CNN-predicted uncertainties.
Summary
The CNN can accurately measure mass profile parameters for Euclid-style images much more rapidly than conventional modelling. CNN accuracy equals or exceeds an automated PyAutoLens while the combination method significantly improves upon both, especially when including CNN-predicted uncertainties which additionally increase PyAutoLens’ modelling speed. Hence, combining CNNs with conventional approaches is a promising method that for automated lens modelling can potentially outperform either one separately.
URL
james.pearson@nottingham.ac.uk
Poster file