Deep learning algorithms for morphological classification of galaxies

Galaxies exhibit a wide variety of morphologies which are strongly related to their star formation histories. Having large samples of morphologically classified galaxies is fundamental to understand their formation and evolution. I will present recent results on morphological classifications for SDSS and DES surveys obtained with Deep Learning (DL) algorithms using convolutional neural networks (CNN). Supervised DL algorithms are fast, accurate and efficient but they rely on large training sets (~5000 ) of pre-labelled galaxies. I will show how transfer learning (i.e., the ability of CNNs to export knowledge acquired from an existing survey to a new dataset) helps to reduce by almost one order of magnitude the necessary training sample for morphological classification. Another important caveat is that visually classified galaxies are usually very bright. We model fainter objects by simulating what the brighter objects with well-determined classifications would look like if they were at higher redshifts. The CNNs reach 97% accuracy to mr ~ 21.5, suggesting that they are able to recover features hidden to the human eye. Where a comparison is possible, our classifications correlate very well with Sérsic index, ellipticity and spectral type, even for the fainter galaxies. We provide classifications for ~27 million galaxies, the largest multi-band catalogue of automated galaxy morphologies to date.

Zoom details:
Please contact: jtous@fqa.ub.edu