Tag: Neural networks

Google Discovers New Planet Which Proves Solar System Is Not Unique

The Kepler-90 star system has eight planets, like our own

Google has previously discovered lost tribes, missing ships and even a forgotten forest. But now it has also found two entire planets.

The technology giant used one its algorithms to sift through thousands of signals sent back to Earth by Nasa’s Kepler space telescope.

One of the new planets was found hiding in the Kepler-90 star system, which is around 2,200 light years away from Earth.

The discovery is important because it takes the number of planets in the star system up to eight, the same as our own Solar System. It is the first time that any system has been found to have as many planets ours.

Andrew Vanderburg, astronomer and Nasa Sagan Postdoctoral Fellow at The University of Texas, Austin, said: “The Kepler-90 star system is like a mini version of our solar system.

You have small planets inside and big planets outside, but everything is scrunched in much closer.

“There is a lot of unexplored real estate in Kepler-90 system and it would almost be surprising if there were not more planets in the system.”

The planet Kepler-90i, is a small rocky planet, which orbits so close to its star that the surface temperature is a ‘scorchingly hot’ 800F (426C). It orbits its own sun once every 14 days.

The Google team applied a neural network to scan weak signals discovered by the Kepler exoplanet-hunting telescope which had been missed by humans.

Kepler has already discovered more than 2,500 exoplanets and 1,000 more which are suspected.

The telescope spent four years scanning 150,000 stars looking for dips in their brightness which might suggest an orbiting planet was passing in front.

Although the observation mission ended in 2013, the spacecraft recorded so much data during its four year mission that scientists expect will be crunching the data for many years to come.

The new planet Kepler-90i is about 30 per cent larger than Earth and very hot.

Christopher Shallue, senior software engineer at Google AI in Mountain View, California, who made the discovery, said the algorithm was so simple that it only took two hours to train to spot exoplanets.

Test of the neural network correctly identified true planets and false positives 96 percent of the time. They have promised to release all of the code so that amateurs can train computers to hunt for their own exoplanets.

Machine learning will become increasingly important for keeping pace with all this data and will help us make more discoveries than ever before,” said Mr Shallue.

This is really exciting discovery and a successful proof of concept in using neural networks to find planets even in challenging situations where signals are very weak.

We plan to search all 150,000 stars, we hope using our technique we will be able to find lots of planets including planets like Earth.”

Please like, share and tweet this article.

Pass it on: Popular Science

Google’s Neural Network Is A Multi-Tasking Pro Can Tackle Eight Tasks At One Time

Neural networks have been trained to complete a number of different tasks including generating pickup lines, adding animation to video games, and guiding robots to grab objects.

But for the most part, these systems are limited to doing one task really well. Trying to train a neural network to do an additional task usually makes it much worse at its first.

However, Google just created a system that tackled eight tasks at one time and managed to do all of them pretty well.

The company’s multi-tasking machine learning system called MultiModal was able to learn how to detect objects in images, provide captions, recognize speech, translate between four pairs of languages as well as parse grammar and syntax. And it did all of that simultaneously.

The system was modeled after the human brain. Different components of a situation like visual and sound input are processed in different areas of the brain, but all of that information comes together so a person can comprehend it in its entirety and respond in whatever way is necessary.

Similarly, MultiModal has small sub-networks for audio, images and text that are connected to a central network.


The network’s performance wasn’t perfect and isn’t yet on par with those of networks that manage just one of these tasks alone. But there were some interesting outcomes.

The separate tasks didn’t hinder the performance of each other and in some cases they actually improved it.

In a blog post the company said, “It is not only possible to achieve good performance while training jointly on multiple tasks, but on tasks with limited quantities of data, the performance actually improves. To our surprise, this happens even if the tasks come from different domains that would appear to have little in common, e.g., an image recognition task can improve performance on a language task.”

MultiModal is still being developed and Google has open-sourced it as part of its Tensor2Tensor library.

Please like, share and tweet this article.

Pass it on: New Scientist