Estimating redshifts via self-organizing maps and an early-stopping approach

Modern photometric and spectroscopic surveys enable the analysis of millions of sources, such as galaxies and active galactic nuclei, using a multiwavelength approach to study their physical characteristics. In order to exploit the scientific output from such extensive volumes of data, the analysis can be conducted with innovative methods such as those employing machine learning techniques and artificial intelligence algorithms. Here, we investigate the use of a specific type of neural network, a self-organizing map (SOM) with an early stopping feature implemented, capable of catching similarities among features in the input data. We consider a sample of galaxies from the COSMOS survey with photometric observations in the Y, J, H , and K s VISTA bands (and associated colors) and show that, by training the SOM and minimizing the over fitting problem, the neural network is able to identify clusters of data depending on the intrinsic redshift of the sources. We then study the behavior of the SOM to estimate the redshift of each input galaxy and address the performance of the algorithm by evaluating the purity and completeness of the output as well as the number of expected outliers.