ARTIFICIAL INTELLIGENCE GOOGLE determine where the picture was taken

If you are afraid of the surveillance, that the news is unlikely you will please. Google Inc. is developing a new neural network PlaNet, which will be able to determine the place where almost every photograph was taken. In the process of learning artificial intelligence as the prompts using geotags, but in the future it budetpo shoulder defining the place, in the photographs, without such "cheat sheets".

For training the neural network Google engineers used a collection of 90 million images tagged geotagged. The results of the first phase of testing quite promising, because PlaNet already learned to identify the country with a probability of 28.4%, and the continent - with a probability of 48%. In 3.6% of cases, even artificial intelligence determines the street where the picture, and in 10% of cases was made - the city.

The author of PlaNet - Tobias Weiland. To implement his ideas, he shared the planet at 26,000 sectors, the size of which depends on how many photos done in a particular area. 

That is why large and densely populated city are covered with smaller sector than the desert and uninhabited. The oceans, the North and South poles of the Earth program are not considered at all. During testing the neural network developers fed her more than 2.3 million photos from Flickr service, parallel protocol leading the successful operation of artificial intelligence.

The results, of course, is far from ideal, but the network is constantly continuing to learn. To make the system even more intelligent, the developers have made it a party Geoguessr games available at this link. During the game, users show photos from the service Google Street View, and he must guess the approximate location on the map where these pictures were taken. When PlaNet competed online against 10 experienced traveler, she won in 28 rounds of 50. Pretty bad for a beginner.

The most surprising is the fact that the whole system PlaNet takes only 377 MB of memory in your computer, which makes it quite possible in the future, its use on mobile devices. What? Not only the security services to use neural networks for their needs