In the creases of deepest space,
the sky is ablaze with light:
fat galaxies, thin galaxies,
barred galaxies, ring galaxies,
new galaxies, old galaxies,
hot galaxies, cold galaxies.
How many galaxies can you see?
Bulging galaxies pregnant with
the dreams of a million million stars,
their tidal tails flapping against
the cooling breeze
of an interstellar wind.
We try to count them all,
keeping track with grains of sand
that run through fingers like
fissures in time.
Till turning the hourglass,
we train machines
to spot spirals
hidden in the brightness.
So many patterns
across the cosmic sea;
how many galaxies could there be?
This poem is inspired by recent research, which has applied artificial intelligence (AI) to finding and classifying galaxies.
The Subaru Telescope is a telescope at the Mauna Kea Observatory in Hawaii, operated by the National Astronomical Observatory of Japan. The lens of the telescope is 8.2 m in diameter, and its large field of view and strong light-gathering power means that it is well suited for making observations of deep-sky objects, i.e. celestial objects (such as galaxies and star clusters) that exist outside of our solar system. Thanks to its high sensitivity, as many as 560,000 galaxies have been detected in the images that have been generated by the Subaru Telescope. In order to better understand the similarities and differences of these galaxies, and how they formed, they need to be classified according to their structural properties (or morphology); for example, are they elliptical or spiral in shape; do they have a bar or ring like structures; do they possess a bulge? Traditionally this has been done by visual classification, as the human brain is much better at recognising patterns than a computer. However, classifying such a huge number of galaxies presents something of a challenge.
Automated processing techniques that are trained using AI have been successfully used for the extraction and judgment of specific features in other applications, such as autonomous vehicles, security cameras, and molecular discovery. In this new study, researchers are utilising such an approach to find and classify those galaxies that are characterised as having a spiral formation. By using training data prepared by humans, AI has been shown to successfully classify spiral galaxies in the Subaru Telescope images with an accuracy of 98%, identifying spirals in about 80,000 of the observed galaxies. Now that this technique has been proven to be effective, it can potentially be extended to classify other types of galaxies. However, such an approach will still rely on an initial training set that is produced by humans, primarily via a citizen-science project ‘Galaxy Cruise’, where participants examine galaxy images taken with the Subaru Telescope to manually search for distinguishing features.