Facebook is in a privileged position to understand human behavior and, well, to profit from it. But the fact that it sees so much data, so many photos, so many posts and so many likes, is both a blessing and a curse. Many companies would kill for all that data, but it all means nothing if it can't be understood.
Today, even the most sophisticated computers do a very poor job at understanding human behavior. Facebook, though it's not the only one, aims to change that.
It has put together a dedicated group, the AI team, of eight people who will focus on deep learning to tackle the biggest problems facing computing, especially when big data is involved.
MIT's Technology Review
has some details on this team and what they plan to achieve. Deep learning systems are smarter than regular machine learning systems in that they require much less hand-holding and can make sense of much more garbled data.
In regular machine learning, the quality of the results depend on what the program is instructed to look for, the quality of the filters and rules, but also on the quality of the data, i.e. how well it's labeled. This means that it requires a lot of human input and, as a result, can only solve a small set of problems.
With deep learning, the system doesn't need too much guidance, it learns for itself what is relevant just by analyzing all the data it's fed, just like humans.
For example, Google's researchers were able to create a neural network that figured out what a cat looked like and even provided a recreation just by watching stills from YouTube videos, even though it was never "told" what a cat looks like in the first place.
It did this by analyzing the objects in all the videos and categorizing them. With enough data and processing power, it was able to recognize cats in videos and label all of them as cats, or rather, the object it associated with cats.
Facebook actually has some catching up to do in the field, but the applications of a similar system are quite obvious. For example, Facebook could start tagging photos by itself, it already does that with human faces. It could expand this to recognize nature scenes, parties and so on and make decisions based on what the photos are about.
The same goes for the News Feed. Facebook's biggest challenge is figuring out what to show you in the news feed on any given day. The average user gets about 1,500 updates from friends and pages every single day, but Facebook narrows that down to 30 to 60 entries.
But as more people join Facebook and spend more time on the site, picking what to show gets harder. A deep learning system would be able to figure out what you like and what others like, make correlations and make better decisions on what to show you.