A recent report shows that Facebook trained its image recognition machine by feeding it 3.5 billion photos from Instagram. The artificial intelligence algorithms can now more easily detect certain objects in the photos.
The social media behemoth unveiled the project this week at its F8 developer conference.
Facebook’s chief technology officer Mike Schroepfer unveiled that the company’s computers are now more accurate when trying to detect specific objects in users’ visual input. One of the greatest challenges the image recognition tech had to face was lack of proper labeling of photos by users.
An image recognition machine needs to first learn what an apple looks like before it can recognize the fruit in other photos.
Since Facebook owns Instagram, it is no surprise that the company used the photo-sharing service’s gigantic image database to improve the platform’s image recognition capabilities.
On Wednesday, Schroepfer told coders that the company selected 3.5 billion images from Instagram to produce “state of the art results.” Most of the photos had hashtags or user descriptions.
Facebook’s Image Recognition Algorithms Getting Better
Another Facebook executive Manohar Paluri acknowledged in an interview prior to this year’s F8 that image recognition algorithms were not making progress because many users attached “noisy hashtags” to Instagram’s images.
For instance, a “noisy hashtag” is naming a dog in an image a husky when that dog is a different breed.
The noise is all over the place,
the Facebook executive said about Instagram’s photo database.
The company has managed to reduce the noise by crosschecking the search results with WordNet, a massive word database for the English language. WordNet allowed Facebook researchers to group hashtags and lower the noise.
As a result, the image recognition system can now tell between a brown bear and other types of bears or between different weather conditions in Instagram images.
Image Source: PxHere