Computer Program Learns to Sort Galaxies Like a Human
A picture of the Abell Cluster taken using the Hubble Space Telescope shows the diversity in galaxy types, including a giant elliptical galaxy at the center of the cluster, a beautiful spiral in the bottom right-hand corner and many smaller systems displaying a wide range of shapes, sizes and colors.
Credit: NASA/ESA and the Hubble Heritage Team (STScI/AURA)

A computer algorithm modeled after the human brain has learned how to recognize different galaxy types ranging from spiral to elliptical, and can now help flesh-and-blood stargazers with the daunting task of classifying billions of galaxies.

The machine-learning codes have proven reliable enough to agree with human classifications of galaxies 90 percent of the time, according to scientists at University College London and the University of Cambridge in the UK.

That should help astronomers keep up with a deluge of galaxy imagery from observational projects such as the Sloan Digital Sky Survey and the Galaxy Zoo. The billions of galaxies in the known universe include a wide range of shapes such as spiral, elliptical, barred and irregular.

"Next generation telescopes now under construction will image hundreds of millions and even billions of galaxies over the coming decade," said Manda Banerji, an astronomer at the University of Cambridge. "The numbers are overwhelming and every image cannot viably be studied by the human eye."

Classifying the types of galaxies represents the first step toward understanding the origin and evolution of galaxies.

More than 250,000 people have helped astronomers classify 60 million galaxies in the online Galaxy Zoo project. Astronomers then used the Galaxy Zoo classifications to train their computer algorithm, known as an artificial neural network, as part of the process for learning to recognize galaxy types.

The artificial neural network can analyze the complex relationships between different variables such as shape, size and color of astrophysical objects, and then come up with the appropriate galaxy type. That process mimics the biological neural network found in living creatures.

Astronomers first trained the computer algorithm on 75,000 astrophysical objects from the Sloan Digital Sky Survey included in the Galaxy Zoo project, before testing its abilities in classifying 1 million objects. They also fiddled with the weighted parameters that the algorithm used so that they could finally achieve the 90 percent success rate.

Some limitations exist. For instance, the algorithm had a problem with misclassifying red spirals and blue ellipticals, because too few examples of those objects exist for the algorithm to get a good read.

The study, published in the journal Monthly Notices of the Royal Astronomical Society, also did not include intermediary objects such as galaxy mergers that defy easy classification.

Still, even a limited auto-classification method should prove useful for astronomers drowning in galaxy imagery and data. One recent celestial census even found that astronomers had been undercounting the number of distant galaxies by about 90 percent.

"While human eyes are very efficient in recognizing patterns, clever computational techniques that can reproduce this behavior are essential as we begin to push the boundaries of our observable Universe and detect more distant galaxies," said Ofer Lahav, an astrophysicist at the University College London. "This study is an important step in that direction."