Bayesian Light Source Separator (BLISS)

BLISS is a Bayesian procedure for deblending light sources. BLISS provides:

  • Accurate estimation of parameters in blended field.

  • Calibrated uncertainties through fitting an approximate Bayesian posterior.

  • Scalability of Bayesian inference to entire astronomical surveys.

BLISS uses state-of-the-art methods in variational inference including:

  • Amortized inference, in which a neural network maps telescope images to an approximate Bayesian posterior on parameters of interest.

  • Variational auto-encoders (VAEs) to fit a flexible model for galaxy morphology.

  • Wake-sleep algorithm to jointly fit the approximate posterior and model parameters such as the PSF and the galaxy VAE.

Latest updates

Galaxies

  • BLISS now includes a galaxy model based on a Variational AutoEncoder that was trained on CATSIM bulge+disk galaxies.

  • We are working on testing galaxy detection functionality and developing galaxy shape measurement.

Stars

Table of Contents