Machine learning in classifying galaxies. Seconds instead of months of arduous work

Machine learning algorithms are consistently gaining importance as tools for scientific and research progress, also in astronomy. This has recently been confirmed by researchers from the University of Western Australia, who have developed a program that performs in a few seconds the tasks previously carried out by a whole staff of people for months. During this time, the software can assign up to thousands of galaxies to the appropriate category.

“Galaxies come in all shapes and sizes,” comments Mitchell Cavanagh, co-author of a recent publication on the subject in the Monthly Notices of the Royal Astronomical Society (MNRAS). “Classifying galaxy shapes is an important step in understanding their formation and evolution. It can even help you get to know the nature of the universe, ”she adds.

Thanks to increasingly dynamic research, scientists are collecting data about galaxies in amounts that they are unable to analyze. “We are talking about several million galaxies in the next few years. Sometimes amateur scientists help to classify them in projects such as the Galaxy Zoo, but it still takes a lot of time “- emphasizes the researcher.

This is what convolutional neural networks (CNNs) can be used for. Today they are used almost everywhere – astronomers explain – in medical research, stock market analysis or customer behavior analysis. They are also present in astronomy, e.g. they have already been used to classify galaxies, but only in a simple system – whether the galaxy is spiral or not.

Meanwhile, the new program checks whether the galaxies are spiral, elliptical, lenticular or irregular in shape, and it works with greater accuracy than the previous ones.

The great advantage of neural networks is the speed of their operation. Images of galaxies that would take months for humans to analyze can be classified in minutes. Using a classic graphics card, we can easily analyze 1,400 galaxies in less than 3 seconds.

However, the computer will not necessarily be more accurate than humans, the researchers note. The reason is that algorithms learn from information developed by humans. However, they managed to achieve an accuracy of 80%, and in the case of spiral and elliptical galaxies – 97%.

Scientists are already preparing to classify 100 million galaxies located at different distances from Earth and in different structures (groups, clusters, etc.).

The developed tool can also be adapted to other domains that require the analysis of large amounts of data. “CNNs will play an increasingly important role in information analysis, especially in areas such as astronomy, which have to deal with big data challenges,” says Cavanagh.

The content of the article itself and a brief description of its main threads are available for viewing on the website.