Machine Learning Just Classified Over Half a Million Galaxies

Humanity is still a long way away from a fully artificial intelligence system. For now at least, AI is particularly good at some specialized tasks, such as classifying cats in videos.  Now it has a new skill set: identifying spiral patterns in galaxies.

As with all AI skills, this one started out with categorized data.  In this case, that data consisted of images of galaxies taken by the Subaru Telescope in Mauna Kea, Hawaii.  The telescope is run by the National Astronomical Observatory of Japan (NAOJ), and has identified upwards of 560,000 galaxies in images it has taken.

Only a small sub-set of those half a million were manually categorized by scientists at NAOJ.  The scientists then trained a deep-learning algorithm to identify galaxies that contained a spiral pattern, similar to the Milky Way.  When applied to a further sub-set of the half a million galaxies (known as a “test” set), the algorithm accurately classified 97.5% of the galaxies surveyed as either spiral or non-spiral.

The research team then applied the algorithm to the fully 560,000 galaxies identified in the data so far.  It classified about 80,000 of them as spiral, leaving about 480,000 as non-spiral galaxies.  Admittedly, there may be some galaxies that are actually spirals that were not identified as such by the algorithm, as they might only be visible edge-on from Earth’s vantage point.  In that case, even human classifiers would have a hard time correctly identifying a galaxy as a spiral.

The next step for the researchers is to train the deep learning algorithm to identify even more types and sub-types of galaxies.  But to do that, they will need even more well categorized data.  To help with that process, they have launched GALAXY CRUISE, a citizen science project where volunteers help to identify galaxies that are merging or colliding.  They will be following in the footsteps of another effort by scientists at the Sloan Digital Sky Survey, which used Galaxy Zoo, collection of citizen science projects, to train a AI algorithm to identify spiral vs non-spiral galaxies as well. After the manual classification is done, the team hopes to upgrade the AI algorithm and analyze all half a million galaxies again to see how many of them might be colliding.  Who knows, a few of those colliding galaxies might even look like cats.

Learn More:
EurekaAlert: Classifying galaxies with artificial intelligence
Physics Letters B: Classifying galaxies with AI and “people power”
Universe Today: Try your hand at identifying galaxies
Unite.ai: Astronomers Apply AI to Discover and Classify Galaxies