Facebook is building artificial intelligence (AI) algorithms that can help build AI algorithms by automating much of the trial and error that goes into their testing.
"We wanted to build a machine-learning assembly line that all engineers at Facebook could use," says Facebook engineer Hussein Mehanna, whose team built a tool known as Flow. With Flow, engineers can build, test, and execute machine-learning algorithms on a huge scale, enabling the testing of a limitless stream of AI concepts across Facebook's data center network.
Mehanna says the company uses Flow to train and test approximately 300,000 machine-learning models every month. He notes this has made it possible for Facebook to launch several new AI models onto its social network every week, whereas it used to deploy a new model onto the network about every 60 days.
Mehanna says Facebook intends eventually to make Flow open source so the rest of the world, including companies such as Twitter and Uber, can use it.
Another tool from his team, AutoML, runs atop Flow to automatically "clean" data needed to train algorithms so they are ready for testing without human intervention. AutoML can apply the outcomes of tests on machine-learning models to train another model that can optimize the training of other models.
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA