Estimation of Photometric Redshifts. II. Identification of Out-of-Distribution Data with Neural Networks

TL;DR

This paper introduces a three-stage neural network training method for accurate photometric redshift estimation of galaxies and effective detection of out-of-distribution objects, leveraging both labeled and unlabeled data for improved real-world applicability.

Contribution

The study presents a novel three-stage training approach combining supervised and unsupervised learning for simultaneous redshift estimation and OOD detection, utilizing unlabeled data.

Findings

Achieves over 98% accuracy in identifying LOOD objects.

Produces photometric redshifts closely matching spectroscopic redshifts for ID samples.

Effectively filters OOD-like unlabeled objects with reasonable redshift estimates.

Abstract

In this study, we propose a three-stage training approach of neural networks for both photometric redshift estimation of galaxies and detection of out-of-distribution (OOD) objects. Our approach comprises supervised and unsupervised learning, which enables using unlabeled (UL) data for OOD detection in training the networks. Employing the UL data, which is the dataset most similar to the real-world data, ensures a reliable usage of the trained model in practice. We quantitatively assess the model performance of photometric redshift estimation and OOD detection using in-distribution (ID) galaxies and labeled OOD (LOOD) samples such as stars and quasars. Our model successfully produces photometric redshifts matched with spectroscopic redshifts for the ID samples and identifies well the LOOD objects with more than 98% accuracy. Although quantitative assessment with the UL samples is impracticable due to the lack of labels and spectroscopic redshifts, we also find that our model successfully estimates reasonable photometric redshifts for ID-like UL samples and filter OOD-like UL objects. The code for the model implementation is available at https://github.com/GooLee0123/MBRNN_OOD.