Detecção de supernovas via redes neurais

Abstract

The work sought to present a general idea about what astronomical events known as supernovas are, the reason to investigate them and a way to carry out this movement using Machine Learning and Deep Learning heuristics is known as convolutional neural networks (CNN). It is verified how a CNN can be used to identify supernovae and how CNNs that have rotational invariance properties are able to produce better results with regard to the detection of these astronomical events. It was observed how the CNN methodology with rotational invariance, called CAP and proposed by [Reyes et al. 2018], achieves more refined results when compared to the Deep-HiTs modeling (which preceded the CAP methodology). Still, the reasons for such improvement in the results were investigated, verifying that it is exactly the fact that the network of layers of operations inferred under the rotated data are capable of performing the processing and correctly predicting the data of a sample data. recorded that corroborates an object verified for a supernova hypothesis (candidate) or not (artifact is observed for a supernova object). Finally, a way to compare the results of such methodologies through statistical tests was illustrated.