We and our partners use cookies to understand how you use our site, improve your experience and serve you personalized content and advertising. Read about how we use cookies and your choices here. By continuing to use this site, you accept these cookies.

Facebook teaches machines what to forget to improve AI

Tossing out irrelevant information makes room for new memories.

Facebook's researchers are working on ways to improve its AI systems.
Graphic by Pixabay/Illustration by CNET

Facebook is teaching its artificial intelligence systems to forget irrelevant information so that computers can complete tasks more quickly and efficiently, the company said Friday.

The social network relies on AI to detect harmful content, such as hate speech and graphic violence, to rank content in its News Feed and carry out other tasks. With 2.85 billion people logging into the social network every month, Facebook has been relying more heavily on AI. 

Facebook's research scientists said in a blog post that it created a new method known as Expire-Span that teaches AI how to forget large volumes of irrelevant information. Each piece of information gets an expiration date, researchers said, freeing up a computer's memory space so it can focus on the necessary details for completing a task. Expire-Span predicts the most relevant information, Facebook said.

"As an example, if the model is training to perform a word prediction task, it's possible to teach AI to remember rare words such as names, but forget very common, filler words such as the and of," Facebook's research scientists Angela Fan and Sainbayar Sukhbaatar wrote in a blog post.

Facebook said that with Expire-Span, the company is one step closer to getting computers to retain memories like humans do. The human brain naturally preserves important information rather than every single detail, the researchers noted.

"The impressive scalability and efficiency of Expire-Span has exciting implications for, one day, achieving a wide-range of hard, humanlike AI capabilities that would otherwise not be possible," the researchers said.