Training the Facebook AI took 22 days, scanning 3.5Billion photos!

Facbook AI

Facebook spent a significant amount of time to train their AI in order to accurately recognize and categorize photos on their platforms. 22 days and 3.5Billion Instagram photos later, the Facebook AI is now ready.

So how ready is the Facebook AI? Well, it can now tell the difference between various types of dosas, different weather conditions, bird species and many more types of images from fruits to specific landscapes.

Also Read: Facebook introduces Crisis Response tools update; new Groups tab

It wasn’t smooth sailing for the company as they tried to jump through hoops to train the Facebook AI due to some completely unrelated and inaccurate hashtags on Instagram pictures. Manohar Paluri, Facebook’s Head of Applied Computer Vision explained just how difficult it was to navigate through inaccurate Instagram descriptions and hashtags accompanying the pictures.

Facebook used an English based linguistic database called WordNet after the computers had analyzed over a billion photos, further grouping popular hashtags together to effectively reduce the noise.

One of the positive results of understanding and recognizing the vast number of images on social media is that it can help blind users understand them through audio descriptions. It is also being incorporated into most of Facebook’s products and the company claims to be learning much more from the data.