Facebook uses giant AI backbone to power social network services

Firm uses AI for translation and image recognition

Facebook has revealed how much of its social network is powered by artificial intelligence (AI) and the extent to which the company is committed to exploring how far it can push smart technology.

Joaquin Quiñonero Candela, Facebook's director of engineering, explained that the firm has an AI backbone used by over 25 per cent of its engineers to improve how Facebook's billion-strong user base communicates.

"Powered by a massive 40 PFLOPS GPU cluster that teams use to train really large models with billions of parameters on huge datasets of trillions of examples, teams across the company run 50 times more AI experiments a day than a year ago, which means that research is going into production faster than ever," he said on Facebook's Code site.

Facebook is also researching where it can take its AI technology next, including the development of smart virtual assistant M for Messenger and delivering social network services unbeknownst to users.

Candela pointed out that 50 per cent of Facebook's users do not speak English or the languages of the people with whom they communicate.

An AI-based automatic translation system, built by Facebook's Applied Machine Learning team, allows these users to communicate without needing to learn a new language.

Candela said that AI-powered translation is necessary as off-the-shelf services are trained using general language structure rather than the more causal and varied nuances found in the languages spoken on Facebook. As such, normal translation services lack the granularity and scope that Facebook's AI can deliver.

"Facebook is all about human-to-human language. It's alive, people use new expressions all the time, they don't spell out their words, there are regional differences, and there are emojis," said Candela.

He also highlighted that the automatic image classifiers in Facebook are AI-supported to aid the process of finding photos and information such as when and where they were taken.

Candela said that Facebook is taking this further by looking at how images can be classified further using image segmentation and AI to help people with sight impediments find pictures and receive audio descriptions of the contents.

All this effectively plays into Facebook's overarching ambitions to create AI technology that can learn from experiences and memory rather than needing rigorous training built around pre-defined models.

AI development and Facebook's seemingly continuous activity to build out its social network and Messenger service as platforms highlight how the company has grown from a social network startup into a genuine technology giant.