Tested on the 18th release of data from the Sloan Digital Sky Survey, this unsupervised model competes with conventional methods: 0.46% outliers on good quality spectra.
With the growth of large spectroscopic surveys, measuring the distance of galaxies (via their redshift) and their properties becomes cumbersome. Classical methods rely on model spectra or reference catalogues, which do not represent the diversity of the objects observed. UnEZ takes the opposite approach: this physics-guided neural network learns without any redshift or line intensity labels.
A template built for each spectrum
An encoder predicts the amplitude and width of the emission and absorption lines, and then a fixed, deterministic decoder derives a spectrum-specific template, a shift estimate, a confidence curve, and a spectral model. Evaluated on Data Release 18 of the Sloan Digital Sky Survey, whose verified catalog serves as an independent judge, it obtains 0.459% outliers for a signal-to-noise ratio greater than 10, and 3.89% beyond 3, results comparable to modern algorithms.
Its predictions can even be used to make the templates needed by traditional methods, an approach tailored to the volumes of future surveys.
Read the full article on the original source
Go to the source →