Serving more than a billion people a day, Facebook has its work cut out for it when providing customized news feeds. That is where the social network giant takes advantage of deep learning to serve up the most relevant news to its vast user base.
Facebook is challenged with finding the best personalized content, Andrew Tulloch, Facebook software engineer, said at the company’s recent @scale conference in Silicon Valley. “Over the past year, more and more, we’ve been applying deep learning techniques to a bunch of these underlying machine learning models that power what stories you see.”
Applying such concepts as neural networks, deep learning is used in production in event prediction, machine translation models, natural language understanding, and computer vision services. Event prediction, in particular, is one of the largest machine learning problems at Facebook, which must serve the top couple of stories out of thousands of possibilities for users, all in a few hundred milliseconds. “Predicting relevance in and of itself is a very challenging problem in general and relies on understanding multiple content modalities like text, pixels from images and video, and the social context,” Tulloch said.
The company must also deal with content posted in more than 100 languages daily, thus complicating classic machine learning, Tulloch said. Text must be understood at a deep level for proper ranking and display. In its deep learning efforts, Facebook has gone with its DeepText text understanding engine, which reads and understands users’ posts and has been open-sourced in part.
In addition, Facebook must account for visual content. “The real challenge is to understand the content of photos and videos from just the pixels because that’s all a computer sees,” Tulloch noted. High-level understanding of content helps Facebook surface visual memories. But deep learning has pushed the state of the art forward in computer vision tasks, Tulloch said, including with classifying videos.
Also deployed is convolution, which takes images and tries to apply filters to identify patterns, to help with high-level semantic understanding, said Yangqing Jia. Facebook has worked to optimize convolution. Still, deep learning is a very generic technique in general, Tulloch said. A lot of approaches to it transfer cleanly across domains.