We propose a new Bayesian Spectral Energy Distribution (SED) fitting tool alleviating computational cost of both the physical modelling and the statistical procedure while retaining the modularity and interpretability of traditional SED fitting.

Bayesian SED fitting is a powerful tool to consistently measure physical properties of galaxies’ stellar population, nebular gas, dust masses and Active Galactic Nucleus, along with the corresponding estimation uncertainties. Physical models have been proposed to simulate each component of galaxies electromagnetic emissions, and numerical codes have been developed to combine those models and compare the resulting simulations to observed SEDs to statistically estimate the models’ parameters. The statistical procedure usually relies either on the computation of massive mock catalogs along a pre-defined parameter grid (CIGALE) or on Markov Chain Monte Carlo schemes. However, the computational cost of the procedure is important: some physical models are computationally intensive, the number of simulations required to make the simulation grid explodes quickly with the number of estimated parameters, and MCMC procedures are notoriously expensive.

We use artificial neural networks as emulators of the expensive physical models and propose a new tempered adaptive multiple importance sampling scheme to replace the grid sampling. The neural network approximators speed up part of the simulation process by a factor $10^3$ to $10^5$ while our importance sampling scheme drastically reduces the number of required simulations. The current rethinking of the methodology also allows to analyze the combined information brought by spectroscopic and photometric surveys (especially from the VLT MOONS project).

The spatial distribution of matter at z=0 depicts a complex pattern commonly referred to as the Cosmic Web in which massive nodes are linked together by elongated bridges found at the intersection of thin mildly-dense walls, themselves surrounding large empty voids.
In this presentation, we introduce T-ReX, a framework allowing the extraction of the filamentary part of the pattern as a graph passing ''in the middle'' of the distribution with robustness to heteroscedastic noise and outliers.
Based on a regularised version of a mixture model, the algorithm approximates the underlying one-dimensional manifold by a graph structure acting like a topological prior on a set of Gaussian clusters paving the galaxy distribution. As a consequence of the unsupervised probabilistic modelling, the learnt graph adapts itself to the local size of the features and provides a natural association of the galaxies to the several branches of the graph.
The overall procedure is meant to converge towards a local maximum of the log-posterior and exploits well-established, fast and simple algorithms making it suitable for large-scale applications.

Amidst an era of rapid technological progress where large quantities of data are produced daily at an ever-increasing rate, unsupervised learning methods have proven to be an extremely useful tool, and have been successfully applied in various scientific fields, ranging from medical science to economics. In this work, we have applied an unsupervised classification method named Fisher-EM on a large sample of optical galaxy spectra simulated using CIGALE.

Short for "Code Investigating GALaxy Emission", CIGALE is an algorithm that can either fit measured spectra or generate ones based on a set of input physical properties of the source galaxy such as the fraction of burst stars, the extinction or the star formation history. Here, it was used to generate a total of 11 475 optical galaxy spectra containing 494 wavelengths each, ranging from 380 nm to 737 nm.

This simulated sample was used to investigate the reliability and repeatability of the Fisher-EM method for unsupervised classification of galaxy spectra. The effect of noise on the process was also studied. Fisher-EM is based on the well-known Expectation-Maximization (EM) algorithm, but adds a third step to it -- the F-step --, where the data is projected onto a discriminant subspace determined by the algorithm itself. The classes are defined in the discriminant subspace by the algorithm as Gaussian probability densities -- it uses a Gaussian mixture model. The fitting process thus also occurs in this discriminant subspace.

The best-fit number of clusters was determined using Integrated Completed Likelihood criterion (ICL). Through spectrum analysis, the method was shown to be able to detect and gather in classes simulated galaxies issued from similar input parameters. Specifically, metallicity, star formation history and burst fraction appear to be the most discriminative parameters for the method. The Fisher-EM analysis of the CIGALE sample has shown only slight variations of goodness of fit over multiple computations, meaning that while the method may not be able to find the absolute best solution, it does consistently provide a very good approximation to it. The effect of the noise was found to be somewhat dependent on the details within the spectra rather than on the signal-to-noise ratio.

Remy Koskas & Jean-Michel Alimi

Cutting-edge instrumentation such as the Euclid satellite and the Rubin Observatory are being prepared in order to observe in the next few years billions of galaxies with exquisite quality, spanning half of the history of the Universe. Surprisingly however, many questions concerning the nature and chronology of processes shaping galaxy evolution are currently still unanswered, which might in fact underline some shortcomings of our conventional approaches.  In order to maximise the scientific return of  exquisite upcoming photometric surveys, our techniques to extract meaningful information on galaxy assembly from their colours must be revisited. In this talk and based on our recent works using the Horizon-AGN Virtual Observatory (Dubois+14, Laigle+19, Davidzon+19), I will discuss how self-organizing maps applied on galaxy colours could be used to improve our understanding on galaxy mass assembly. 

Les techniques de Machine Learning sont de plus en plus utilisées en astrophysique, comme dans la reste de la société. Je présenterai des exemples d'utilisation de ces techniques dans le domaine de l'étude du milieu interstellaire, en particulier pour la determination de la composition chimique des nébuleuses planétaires ou des regions HII. Les méthodes présentées font principalement appel à des regresseurs de type XGBoost ou Réseau de Neurones. Quelques applications et résultats préliminaires seront discutés.

Gravitational lensing is one of the most useful and spectacular consequences of General Relativity which can result in observing multiple lensed images of a distant quasar. It is the most accurate technique to weigh galaxies, study their dark matter content/mass profile over cosmic history, and one of the few techniques that allow the measurement of the Hubble constant H0 with <4% accuracy. At present, only ~40 quadruply-imaged quasars are known (Ducourant et al. 2018), due to the small separation of components (<1"). The Gaia space mission offers for the first time an all-sky survey with an exceptional spatial resolution of ~ 0.2" up to magnitude G~20.7 mag. Its final Data Release (DR) is expected to contain, out of two billion sources, ~250 quadruply imaged quasars (quads) including compact (separations <0.8") configurations and many more doubly-imaged quasars. In this presentation, I will tell you about my work to blindly identify them all using a deep learning algorithm.

Mergers of galaxies are extremely violent which may trigger starbursts or black hole accretion. The issue to address their effects from an observational perspective is that, as galaxy mergers last several Gyr, we can only observe them at a particular time which limits our ability to know the fate of a galaxy pair. To overcome this issue, we use the Horizon-AGN simulation to build a catalog of a billion of galaxy pair for which we know the fate. We then test three selection processes aiming at identifying true merging pairs. We find that simple threshold gives similar results as more complex neural network as the problem is particularly unbalanced and degenerate. We end up with simple criterion that can easily be applied to real catalogs.

We present a novel methodology to address ill-posed inverse problems, by providing a description of the posterior distribution instead of a point estimate solution. Our approach combines Neural Score Matching for learning a prior distribution from physical simulations, and an Annealed Hamiltonian Monte-Carlo technique to sample the full high-dimensional posterior of our problem.

In the astrophysical problem we address, by measuring the lensing effect on a large number of galaxies, it is possible to reconstruct maps of the Dark Matter distribution on the sky. However, presence of missing data and noise dominated measurement makes the inverse problem non-invertible.

We propose to reformulate the problem in a Bayesian framework, where the target becomes the posterior distribution of mass given the galaxies shape observations. The likelihood factor, describing how light-rays are bent by gravity, how measurements are affected by noise, and accounting for missing observational data, is fully described by a physical model. The prior factor is learned over simulations using a recent class of Deep Generative Models based on Neural Score Matching and takes into account theoretical knowledge. We are thus able to obtain samples from the full Bayesian posterior of the problem and can provide Dark Matter map reconstruction alongside uncertainty quantifications.

We present an application of this methodology on the first deep-learning-assisted Dark Matter map reconstruction of the Hubble Space Telescope COSMOS field.

I would like to present the new innovative methods based on machine learning technics that I developed to obtain precise distances for distant stars using spectroscopic data and/or broad band photometry. The distances thus derived by these new methods have already proved to be extremely useful in discovering new galactic structures, and to complement Gaia data. One of these method will be used to measure distance to all the stars observed by the incoming WEAVE spectroscopic survey, while the other will be directly applicable on the next LSST data. </span>

In this talk, I present BlendHunter, a proof-of-concept for a deep transfer learning based approach for the automated and robust identification of blended sources in galaxy survey data. We take the VGG-16 network with pre-trained convolutional layers and train the fully connected layers on parametric models of COSMOS images. We test the efficacy of the transfer learning by taking the weights learned on the parametric models and using them to identify blends in more realistic CFIS-like images. We compare the performance of this method to SEP (a Python implementation of SExtractor) as function of noise level and the separation between sources. We find that BlendHunter outperforms SEP about 15% in terms of classification accuracy for close blends (<10 pixel separation between sources) regardless of the noise level used for training. Additionally, the method provides consistent results to SEP for distant blends (>=10 pixel separation between sources) provided the network is trained on data with a relatively close noise standard deviation to the target images.