At Facebook, machine learning provides a wide range of functions to control many aspects of the user experience, including ranking posts, understanding content, object recognition, and augmented and virtual reality tracking, voice and text translation. Although machine learning models are currently being trained on the adapted data center infrastructure, Facebook is bringing machine learning inference to the edge. It improves the user experience by reducing latency (reasoning time) and reduces reliance on network connections. In addition, this supports many other applications of deep learning that have essential features that are only available at the edge. This article uses a data-driven approach to describe the opportunities and design challenges for Facebook to implement local machine learning on smartphones and other edge platforms.
Teaching AI systems to understand what happens in videos like humans is one of the most daunting challenges in machine learning and the most significant potential breakthrough. Today, Facebook announced a new plan which will cost up to 34 billion dollars, hoping to give it an advantage in follow-up work: training its artificial intelligence on public videos of Facebook users.
Obtaining training data is one of the most significant competitive advantages of artificial intelligence. By collecting this resource from millions of users, technology giants such as Facebook, Google, and Amazon have progressed in various fields. Although Facebook has trained machine vision models on billions of images collected by Instagram, it has not previously announced any projects with similar video understanding ambitions.
"By learning from global public video streams in almost every country and hundreds of languages, our artificial intelligence system can not only improve accuracy but also adapt to our fast-paced world, recognizing the nuances and visuals of different cultures and regions. Clues," the company said on its blog. The project is called "Learning from Video" and is part of Facebook's "wider effort to build machines that learn like humans."
The resulting machine learning model will be used to create new content recommendation systems and review tools, Facebook said, but more can be done in the future. Artificial intelligence that can understand video content can give Facebook unprecedented insight into users' lives, allowing them to analyze their hobbies and interests, brand and clothing preferences, and countless other personal details. Of course, Facebook can already access this information through its current ad targeting process. Still, the ability to analyze videos through artificial intelligence will add a prosperous (and intrusive) data source to its store.