A Bayesian Neural Network Approach to identify Stars and AGNs observed by XMM Newton

TL;DR

This paper introduces a Bayesian Neural Network approach for classifying stars and AGNs in XMM Newton data, effectively reducing overconfidence and improving prediction accuracy over traditional methods.

Contribution

It presents the first application of Bayesian Neural Networks in observational astronomy for classifying celestial objects, enhancing confidence estimation and classification accuracy.

Findings

Successfully classified over 62,807 AGNs and 88,107 stars with high confidence.

Reduced misclassification of objects like CVs, Pulsars, ULX, and LMX.

The approach significantly lowers the error rate compared to frequentist methods.

Abstract

In today's era, a tremendous amount of data is generated by different observatories and manual classification of data is something which is practically impossible. Hence, to classify and categorize the objects there are multiple machine and deep learning techniques used. However, these predictions are overconfident and won't be able to identify if the data actually belongs to the trained class. To solve this major problem of overconfidence, in this study we propose a novel Bayesian Neural Network which randomly samples weights from a distribution as opposed to the fixed weight vector considered in the frequentist approach. The study involves the classification of Stars and AGNs observed by XMM Newton. However, for testing purposes, we consider CV, Pulsars, ULX, and LMX along with Stars and AGNs which the algorithm refuses to predict with higher accuracy as opposed to the frequentist approaches wherein these objects are predicted as either Stars or AGNs. The proposed algorithm is one of the first instances wherein the use of Bayesian Neural Networks is done in observational astronomy. Additionally, we also make our algorithm to identify stars and AGNs in the whole XMM-Newton DR11 catalogue. The algorithm almost identifies 62807 data points as AGNs and 88107 data points as Stars with enough confidence. In all other cases, the algorithm refuses to make predictions due to high uncertainty and hence reduces the error rate.