CosmoStat Logo_GalxyMark

Learning in Astrophysics day

Learning in Astrophysics day

Date: January the 26th, 2017

Venue:

Local information

CEA Saclay is around 23 km South of Paris. The astrophysics division (DAp) is located at the CEA site at Orme des Merisiers, which is around 1 km South of the main CEA campus. See http://www.cosmostat.org/link/how-to-get-to-sap/ for detailed information on how to arrive.


On January the 26th, 2017, we organize the third day on machine learning in astrophysics at DAp, CEA Saclay. 

Program:

All talks are taking place at DAp, Salle Galilée (Building 713)

10:00 - 10:45h. Marc Duranton  (CEA Saclay)
10:45 - 11:15h. Rémi Flamary (Université Nice-Sophia Antipolis)
11:15 - 11:45h. Christoph Ernst René Schäfer  (EPFL)

12:00 - 13:30h. Lunch

13:30 - 14:00h. Emille Ishida  (Laboratoire de Physique de Clermont)
14:00 - 14:30h. TBD                                                           
14:30 - 15:00h. Arthur Pajot (LIP6)
15:00 - 15:30h. Morgan Schmitz  (CEA Saclay - CosmoStat)

15:30 - 16:00h. Coffe break

16:00 - 17:00h. Round table

CosmoStat_Twitter_Profile

CosmoClub: 08-11-2017

Date: November 8th 2017

Speaker: François Orieux (L2S, Centrale-Supélec)

Title: Semi-unsupervised Bayesian Convex Image Restoration with Location Mixture of Gaussian


Abstract

Convex image restoration is a major field in inverse problems. The problem is often addressed by hand-tuning hyper-parameters. We propose an incremental contribution about a Bayesian approach where a convex field is constructed via Location Mixture of Gaussian and the estimator computed with a fast MCMC algorithm. Main contributions are a new field with several operator avoiding crosslike artifacts and a fallback sampling algorithm to prevent numerical errors. Results, in comparison to standard supervised results, have equivalent quality in a quasi-unsupervised approach and go with uncertainty quantification.

CosmoStat_Twitter_Profile

CosmoClub: 19-10-2017

Date: October 19th 2017

Speaker: François Lanusse (Carnegie Mellon University)

Title: Towards a better control of weak lensing systematics in the LSST era using Deep Learning


Abstract

In this talk, I will present several applications of the most recent advances made in the field of Deep Learning to some of the main sources of systematics affecting the weak lensing probe in next generation wide field cosmological surveys.
Most current shape measurement methods used for weak lensing will require precise calibration to reach their science requirements. This calibration step is typically performed through extensive galaxy image simulations. I will present an application of Deep Generative Models, based on variational inference, aiming at producing realistic galaxy images with complex morphologies which can be used as inputs of the image simulation pipeline instead of scarce and expensive space-based observations.
A second important source of systematics comes from uncertainties in photometric redshift estimation. Most current techniques rely solely on color measurements to build a redshift estimation. I will present an extension of these ideas but based on using multi-band galaxy images as input of the
supervised regression problem. The proposed architecture, based on Deep Residual Networks (resnet), is able to complement color information with morphology information in order to improve the accuracy of the redshift estimate.
Finally, another major potential source of systematics comes from intrinsic galaxy alignments (IA). While hydrodynamical simulations can reproduce to some extent these alignments, they are limited to fairly small cosmological volumes. However, to properly test IA marginalisation schemes at the scale of
LSST or Euclid, much larger simulation volumes are necessary. I will present an application Graph Convolutional Networks which allows us to model the IA signal on a graph of the cosmic web. This method allows us to populate simpler Dark Matter only N-body simulations of much larger volumes with realistic galaxy alignments.

CosmoStat_Twitter_Profile

CosmoClub: 06-10-2017

Date: October 6th 2017

Speaker: Daniela Saadeh (University of Nottingham)

Title: How isotropic is the Universe?


Abstract

A fundamental assumption of the standard model of cosmology is that the large-scale Universe is isotropic. Because of its centrality, it is essential to test this assumption. Breaking isotropy leads to Bianchi cosmologies, a set of solutions to the Einstein field equations of which only the subset describing rotating universes was previously tested against data.
In this talk, I present a general test of isotropy considering, for the first time, all the degrees of freedom of anisotropic expansion. We analyse cosmic microwave background data from Planck, carrying out the first joint analysis of temperature and polarization data for this purpose. We also show that improved constraints on anisotropy may be obtained by extending the likelihood to high ell.
For the vector mode (associated with rotating universes), we obtain a limit on the anisotropic expansion that is an order of magnitude tighter than previous Planck results using the CMB temperature only. We recover upper limits for all the other modes, with the weakest one arising from the regular tensor modes. We disfavour anisotropic expansion of the Universe with odds of 121,000:1 against.