Astrophysical Journal, Big Data, Machine Learning, NASA, stars

A recent article in The Astrophysical Journal reports that NASA has started to use machine learning techniques to understand the properties of stars more quickly. The article states, “New time-domain surveys have begun exploring everything from nearby extrasolar planets to the most distant known stellar explosions, and a veritable zoo of time-variable astrophysical phenomena in the space between (e.g., Borucki et al. 2010; Law et al. 2009; Gehrels et al. 2004). The volume of data and sheer breadth of inquiry of existing surveys will eventually be dwarfed by the Large Synoptic Survey Telescope (LSST; Ivezi? et al. 2008b), which will track the brightness variations of ~20 billion sources throughout the universe. The rapidly increasing rate at which we acquire and process observations for these surveys requires sophisticated algorithms capable of discovering and classifying new sources as well as, or better than, human experts.”

It continues, “Machine-learning methods provide a promising avenue for the necessary abstraction of the discovery and classification process.11The algorithms defining these methods are data-driven, built to learn relationships between observables and parameters of interest without relying on parametric physical models. The learning is achieved using objects with known properties (such as a variable star classification or a galaxy redshift), which is called the training set. Once a machine-learning model has been trained, it can be rapidly applied to new data providing predictions of the quantities of interest. As more data are obtained, and the quality and scope of the training set are improved, the machine can refine its knowledge and model of the data set, providing ever more accurate predictions. Furthermore, unlike humans, machine-learning models can nearly instantaneously and automatically produce predictions about new data via a fully scalable process.”

Read the full article here.