We present a new, fully generative model for constructing astronomicalcatalogs from optical telescope image sets. Each pixel intensity is treated asa random variable with parameters that depend on the latent properties of starsand galaxies. These latent properties are themselves modeled as random. Wecompare two procedures for posterior inference. One procedure is based onMarkov chain Monte Carlo (MCMC) while the other is based on variationalinference (VI). The MCMC procedure excels at quantifying uncertainty, while theVI procedure is 1000 times faster. On a supercomputer, the VI procedureefficiently uses 665,000 CPU cores to construct an astronomical catalog from 50terabytes of images in 14.6 minutes, demonstrating the scaling characteristicsnecessary to construct catalogs for upcoming astronomical surveys.
translated by 谷歌翻译