Innovation

Facebook's machine learning director shares tips for building a successful AI platform

At the Applied Artificial Intelligence Conference for tech innovators, Facebook's director of Core Machine Learning explained how the social media giant developed its own machine learning system.

Image: Andrew Grosser

It's no longer up for debate that AI is set to have a major impact on most businesses, if it isn't already—and any company that wants to stay ahead must figure out how to integrate the new technology into its structure. But how is a successful AI platform built?

On Wednesday, at the Applied Artificial Intelligence Conference in San Francisco, hosted by BootstrapLabs, Hussein Mehanna, the director of core machine learning at Facebook, tackled this question.

In Mehanna's session, he explained how Facebook developed its own machine learning platform, and how Facebook employees are using it.

In 2012, Mehanna said, Facebook's AI platform was "a snowball of complexity"—a system that slowed progress down significantly.

"We had to do something," he said.

Mehanna described the development of FBLearner Flow, a machine learning platform that could take data, produce machine learning models, feed the information back to FBlearner predictor, and integrate the information back into the system. The information is then used in Facebook products like Search, Ads and News Feed.

So, how did the new system change machine learning at Facebook?

SEE: Why Facebook wants to use AI to track your conversations online (TechRepublic)

A quarter of the engineering workforce at Facebook leverages the system. "We made AI available to our engineers, without them having to learn more about AI," Mehanna said. "We want to make AI as simple to use as a possible, so that everybody can leverage it to build better products at Facebook."

So what's the "recipe" for Facebook's success in developing the platform? Here are Mehanna's three takeaways:

  1. Design is critical. "You need to nail the architecture of the platform, otherwise it may not live," said Mehanna. Specifically, you must be algorithm-agnostic. When a platform is optimized for a few set of algorithms, it's open to disruption, said Mehanna. "Being algorithm-agnostic is an advantage."
  2. Prepare for disruption. No matter what you do, Mehanna said, there will be far more people working on AI elsewhere than in your own organization. So how to prepare? Start by disrupting yourself.
  3. Know your users. If you want your employees to use the platform, you should know who they are, which can help you figure out what direction to go. The mistake most companies make, Mehanna said, is that they assume one kind of user: The algorithm-designer type. But, Mehanna realized, far more people could leverage AI if the algorithm was user-friendly. There's a second type to use the platform—the general user, who values "usability over power." Facebook saw that the work of every expert was being used by 13 non-experts. So, experts could then easily build workflows, include them in the platform, and other users could use their work with a minimum amount of overhead.

The system, it could be argued, has access to some of the largest data sets on earth. Each day, the system processes 2 trillion training examples—from behavior on the social site. There are 500,000 models trained a month.

What does Facebook do with all that information? "We see how people are communicating, we see how they are having conversations," said Alan Packer, the director of Facebook's Language Technology Group, at EmTechDIGITAL 2016 on Monday.

"We know you."

Also see

About Hope Reese

Hope Reese is a Staff Writer for TechRepublic. She covers the intersection of technology and society, examining the people and ideas that transform how we live today.

Editor's Picks