Data Centers

Facebook drives 2.5x performance boost in machine translation using neural networks

Using its scalable deep learning framework, Caffe2, and recurrent neural networks, Facebook was able to further scale translations happening on the social network.

Using recurrent neural networks (RNNs) and its Caffe2 deep learning framework, Facebook was able to achieve a 2.5x efficiency boost in its machine translation efforts, the company announced in a blog post on Thursday. As a result of this success, Facebook is moving all of its machine translation models to neural networks, the post said.

To give some context for the scale of that transition, it's important to note that Facebook handles some 4.5 billion translations everyday on the backend of its network. All of those are now effectively being handled, in part, by a form of artificial intelligence (AI).

According to the post, using Caffe2 "significantly improved the efficiency and quality of machine translation systems. As such, other Facebook product teams, such as speech recognition and ads ranking, have been using the framework to train RNN models as well, the post said.

SEE: Special report: How to implement AI and machine learning (free PDF) (TechRepublic)

When Facebook initially open sourced Caffe2 back in April 2017, it didn't support RNNs. However, the post noted, the Facebook AI team has been working on the "building blocks" for RNN use cases like machine translation and speech recognition for the past few months.

For developers and data scientists, Caffe2 offers a generic RNN library with an RNN engine that executes RNN cells, the post said. In a way, RNN cells are a type of smaller Caffe2 network that can be utilized through APIs.

The work being done by Facebook highlights the potential business value of emerging AI tools like RNNs. An infrastructure move of that magnitude is extremely difficult for most companies, so it has to be weighed against the potential value it can bring to the company implementing it.

SMBs, and even some larger companies, may not be able to manage a shift that big, or may simply not have enough data to benefit from using these tools. Fortune 500 firms and very large enterprises, however, should begin consider if their most data-intensive processes could benefit from AI tools.

The 3 big takeaways for TechRepublic readers

  1. Facebook used Caffe2 and recurrent neural networks to get a 2.5x efficiency boost in its machine translation efforts.
  2. Due to the success of its work, all of its machine translation models, along with some other processes, are being moved to neural networks.
  3. Large enterprises with a lot of data should seriously consider the use of AI tools like deep learning and neural networks to improve performance.

Also see

neuralnetworks.jpg
Image: iStockphoto/cosmin4000

About Conner Forrest

Conner Forrest is a Senior Editor for TechRepublic. He covers enterprise technology and is interested in the convergence of tech and culture.

Editor's Picks

Free Newsletters, In your Inbox