Facebook is using billions of publicly available Instagram photos to train its image recognition AI. The company stated this while showing off research about its image recognition AI at its annual developer conference, F8 today. This shouldn't be too surprising, seeing as Facebook is really only using publicly available images, and Facebook does collect all kinds of data from its users. The reason why it's using these images to train its AI is because users have curated these images with captions and loads of hashtags.
In the presentation, Facebook detailed how it took over 3.5 billion Instagram images, which spanned over 17,000 hashtags to train the AI. Which is something that even Facebook couldn't supervise closely. It also relied on hundreds of GPUs continuously running to parse the data. This left Facebook with some deep learning modes that beat the industry benchmarks. However, Facebook did run into some trouble with sorting hashtags, as a lot of Instagram users will add hashtags that are not relevant, to drive in more followers, likes and comments. So Facebook had to decipher which hashtags were relevant to the image as well. It wasn't just about training the AI, but accurately training it.
Facebook isn't the only one using images to train its AI, Google has been doing it as well. In fact, Google does it with your photos stored in Google Photos, which are not publicly available. Other companies are likely doing the same thing as well. An image recognition AI is going to become a pretty important feature in the coming months, and we'll likely hear plenty about it from Google at its developer conference next week in Mountain View. An image recognition AI could be a pretty useful thing to have for Facebook and Instagram. As it could ideally suggest hashtags for you to use when posting your images on either site.