AstroDeep

Astronomical surveys planned for the coming years will produce data that present analysis challenges not only because of their scale (hundreds of petabytes), but also by the complexity of the measurement challenges on very deep images (for instance subpercent-level measurement of colors or shapes on blended objects). Machine learning techniques appear very promising: once trained, they are very fast, and excel at extracting features from images. Preliminary results using modified variational auto-encoders as deblenders are very encouraging. We aim to develop machine learning techniques that can be applied directly on multi-bandpass, multi-instrument individual images to address the key observation challenges without going through the traditional steps of image stacking, explicit deblending, catalog generation, that lose information at each stage. We will target as main objectives the measurement of weak gravitational lensing (through a local average shape of objects), and of photometric redshifts (producing a bayesian p(z) probability for each object of a blend). To achieve these goals, we will develop new deblending techniques, and will use and improve bayesian deep learning techniques in order to achieve a proper and consistent handling of uncertainties.Those techniques will help leverage the observation capabilities of future surveys like LSST, Euclid and WFIRST, and will allow a joint analysis of their data.