Facebook is opening up many of the AI tools it uses to drive its online services.
Most of these tools seek to take better advantage of artificial intelligence algorithms that Facebook and other researchers have already published in academic journals, and the hope is that this newly open sourced code can save outsiders quite a bit of time as they build their own AI services, involving everything from speech and image recognition to natural language processing. The algorithms alone aren’t always enough.
“Someone has to go and implement the algorithm in a program, and that’s not trivial in general,” says Facebook artificial intelligence researcher and software engineer Soumith Chintala. “You have to have a lot of skill to implement it efficiently.”
Chintala says the open source project could help research labs and startups that don’t have a lot of resources and wind-up spending most of their time just implementing existing algorithms instead of doing new research. In that sense, Facebook will benefit too. “Even though we don’t collaborate day-to-day with that world, it could provide a general catalyst to the community and that will benefit us indirectly,” he says.
The tools came out of the Facebook Artificial Intelligence Research lab, a project started within Facebook about a year ago to research a subfield of artificial intelligence called “deep learning,” which seeks to model certain behaviors of the brain in order to create software that can learn and make predictions. With Facebook, Google, and Microsoft leading the way, deep learning is poised to hone so many of the online services we used on a daily basis.
Facebook already uses deep learning to filter your Facebook feed, making intelligence guesses as to which items you’ll find most interesting, and to recognize faces in the photos you upload. But eventually, the company expects to create digital assistants that can, for example, stop you from posting drunk selfies in the middle of the night.
What Facebook released today is a set of modules for Torch, an open source computing framework for working with deep learning widely used in academia and by companies like Google and Twitter. Torch already includes several deep learning algorithms, but Chintala says Facebook’s are far faster and more efficient. That will allow researchers to tackle much larger problems than ever before, he says. For example, one team of researchers Facebook has already worked with were able to create a photo recognition tool that can tell what physical poses—standing, sitting, lying down, etc.—characterized people in photos.
“We benchmarked our code, and these are the fastest open source implementations out there,” he says. “People didn’t explore certain areas because they didn’t think it was possible and now they are.”