DEEP Learning For Large Deep Imaging Programs – DEEPDIP

DEEPDIP: Deep learning for Deep Imaging Programs

In future large surveys like LSST or Euclid, high precision photometric redshift measurements and light curve classification will play a central role. We propose to revisit the current methods used to analyze these data by developing deep learning techniques that will directly use the multi-band images at the pixel level to measure a redshift or to perform a classification.

Improving the accuracy of comological analysis by exploiting the statistical power of photometric surveys

Understanding the origin of the accelerated expansion of the Universe, the formation of the large-scale structure and the galaxies embedded in it are the quests of modern cosmology. Huge observational and theoretical efforts are being made to address these fundamental questions. In this proposal, we focus on two major observational probes: the expansion of the universe through the Hubble diagram of type IA supernovae and the growth of cosmic structures. For the former, one challenge is to identify and classify SNIa signatures among a huge number of observational time series; another challenge, shared by the latter probe, is to obtain reliable redshifts for as many galaxies as possible. Our goal is to address these two challenges by exploiting the statistical power of photometric surveys, which will allow us to use and improve the prohibitively time-consuming spectroscopic approach.