Abstract Details

Name: Yogesh Wadadekar
Affiliation: National Centre for Radio Astrophysics
Conference ID: ASI2021_525
Title : Teaching machines to study galaxies
Authors and Co-Authors : Shraddha Surana (Thoughtworks, Pune), Harsh Grover (BITS, Pilani), Ashwin Samudre (PICT, Pune), Omkar Bait (NCRA-TIFR), and Preetish K. Mishra (NCRA-TIFR)
Abstract Type : Oral
Abstract Category : Extragalactic Astronomy
Abstract : Large area surveys over the last two decades in the UVOIR and radio bands have led to an explosion in the availability of imaging data on galaxies. Sample sizes of a million are now common, making it difficult to classify, let alone study individual objects. In such a data-rich situation, machine learning algorithms based on neural networks have emerged as powerful tools for both classification and regression problems. A specific machine learning architecture called deep learning with many hidden layers in the network has proved particularly effective. Excellent implementations of deep learning techniques are also freely available. We have applied supervised deep learning algorithms to address three specific problems in extragalactic astronomy (1) predicting star-formation histories of galaxies using flux measurements in 21 broad bands from UV to far-infrared (2) predicting the bulge-to-total luminosity ratio of galaxies from $gri$ color composite images and (3) classifying radio galaxy images into subclasses - compact, FR-I, FR-II, and bent tail. In all these applications, our deep learning models are able to make robust predictions with low error rates and in real-time once the model has been trained. We highlight the challenges faced in terms of data size, availability, features, processing ability, and importantly, the interpretability of results for these specific problems. In the coming years, as the data tsunami in astronomy, grows exponentially with the commissioning of facilities like the SKA and the LSST, we will highlight how increased use of machine learning to understand the underlying physics in the information captured will maximise the scientific return.