Informazioni sull’evento

02/09/2018

Attenzione: a seguito delle disposizioni per il contenimento della diffusione del coronavirus, i seminari potranno essere in modalità online e/o in presenza, si prega di consultare i singoli avvisi. In caso di modalità online il link per il collegamento verrà reso disponibile nei giorni precedenti la data del seminario. Qui, invece, la playlist dei remote talks già messi in rete »

Special Talk

Testing Machine Learning Methods for Classification and Physical Property Estimation

Andrew Humphrey (Instituto de Astrofísica e Ciências do Espaço (Porto))

Thursday 07/07/2022 @ 11:30, Sala IV piano Battiferro + remote

With the Euclid Space Telescope set to be launched in the very near future, preparation is underway for the classification of sources and the estimation of redshift and physical properties. The immense expected volume of images and photometry will require software tools for the characterisation of sources that are both accurate and highly scalable. With this motivation, we are exploring machine-learning methodologies for source classification, and for the estimation of redshifts and physical properties such as stellar mass, star-formation rate, etc., for eventual application to Euclid data. In this talk, I will present recent results from this work. Firstly, I will describe results from a Euclid pre-launch key project paper, in which we developed a new machine learning pipeline for the selection of quiescent galaxies from Euclid and LSST/UNIONS photometry. With this method, we obtain significant improvements over colour-colour or SED-fitting (LePhare) selection methods. Lessons learnt from the various tests and experiments will also be briefly described. Secondly, I will present a machine learning pipeline that uses CatBoost regressor chains to estimate galaxy redshifts and physical properties (mass, SFR, extinction, age) from broad band photometry. Key features of this method include its ability to take into account non-detections, and an emphasis on physically reasonable covariance between the predicted properties. Finally, I will briefly discuss results from a few other projects: (i) Redshift-aided classification of stars, galaxies and quasars (Cunha & Humphrey, 2022, arXiv:2204.02080); (ii) Improving photo-z and physical property ML predictions using unlabelled data; (iii) ML classifier performance prediction and population shift detection in the absence of ground truth.