Google open sources machine learning tech it used to find new planets

Google is open sourcing its code for processing data from NASA's Kepler space telescope and training neural network models, in the search for exoplanets.

NASA's coding moonshot that launched Apollo
Building a slide deck, pitch, or presentation? Here are the big takeaways:
  • Google is open sourcing its code for processing data from NASA's Kepler space telescope, training its neural network model, and making predictions about planet signals, allowing others to search for exoplanets.
  • Google discovered two exoplanets using a neural network analyzing data from the Kepler space telescope.

Google will open source the machine learning technology that allowed it to discover new exoplanets, the tech giant announced in a Thursday blog post.

In December, Google announced that it had found two exoplanets by training a neural network to analyze data from NASA's Kepler space telescope and identify signals that could be coming from planets, our sister site ZDNet reported at the time.

This success suggests that machine learning could be used both for discovering exoplanets, as well as a number of other scientific disciplines, including healthcare and quantum chemistry, Chris Shallue, a senior software engineer on the Google Brain Team, wrote in the blog post.

SEE: IT pro's guide to working smarter with Linux (Tech Pro Research)

This code may be useful for those developing similar models for other NASA missions, including Kepler's second mission K2, and the upcoming Transiting Exoplanet Survey Satellite mission, the post noted.

In the post, Google also explained the basics of how the model works. When searching for planets in Kepler data, scientists use automated software to first detect signals that could potentially be caused by planets. They then must examine the signals manually to determine if they match a planet, or are a false positive.

While there are cutoffs to the automated detections to try and avoid those false positives, it's still an issue: To date, more than 30,000 detected Kepler signals have been manually examined, but only about 2,500 of those were found to be planets, the post said. And planets with signals that fall below the cutoff may be left out of the search.

SEE: Exomedicine arrives: How labs in space could pave the way for healthcare breakthroughs on Earth (TechRepublic cover story) | download the PDF version

Google's Brain Team developed a neural network to help search the low signal-to-noise detections for planets. It trained its model on 15,000 of the signals that had already been manually examined, to help the network learn the difference between actual planets and false positives.

The team then tested the effectiveness of its model by searching for new planets in a set of 670 stars, and allowed the search to include weaker signals that would be passed over by astronomers previously. It discovered two planets: Kepler-90 i and Kepler-80 g.

Google will continue working in this space, the post noted, and will make improvements to its model, helping it become more skilled at rejecting false positives.

"Our work here is far from done," Shallue wrote in the post. "We've only searched 670 stars out of 200,000 observed by Kepler -- who knows what we might find when we turn our technique to the entire dataset."

Google has also been pushing its machine learning into new areas on Earth that may impact the enterprise, ZDNet noted, including partnerships with Rolls Royce to work on autonomous ships, as well as adding machine learning capabilities to its Sheets app.

Also see

Image: iStockphoto/ClaudioVentrella