Dissertation defense (December 10, 2020): Raphael Correia de Souza Fialho

Abstract: In Astronomy it has become common to use machine learning algorithms during the process of capturing and analyzing astronomical events. Due to the current amount of data captured by telescopes and antennas in astronomical surveys, these data are usually stored, cataloged and transformed for further analysis and studies. A particular type of analysis done on these data is the prediction of the photometric redshift, a measure that is related to how far an object (e.g., galaxy or quasar) is in relation to a given reference point. A relevant feature of the datasets used for investigating photometric redshift is that each entry presents not only the measurements made for a given object, but also an error value corresponding to each measurement. In this dissertation we study the construction of prediction models for photometric redshift using machine learning algorithms. We focus on the use of artificial neural networks. Our goal is to investigate how these models behave in scenarios where information about measurement errors is considered or ignored during the learning stage. In particular, we propose a training technique that aims to take advantage of the error values. We performed comparative computational experiments to evaluate the effectiveness of the proposed technique.