In the world of machine learning, the buzzword these days is “deep learning.” It’s a technique that has been a popularized by Geoff Hinton, who is now at Google and previously worked at Microsoft Research, as well as other computer science researchers like Yann LeCun who are looking for better ways to teach computers how to recognize objects and speech.
Facebook, too, has done quite a bit of work in the area and today, the company is open-sourcing some of its projects around the Torch7 computing framework for machine learning. Torch has long been at the center of many machine learning and artificial intelligence projects in academic labs and at companies like Google, Twitter and Intel.
Facebook today is launching optimized tools to increase the speed at which deep-learning projects that use Torch run. One allows developers to parallelize the training of their networks using multiple GPUs simultaneously. Another improvement ensures that training the convolutional neural nets at the center of many deep learning systems can be trained 23 times faster when compared to the fastest publicly available code today.
In addition, Facebook is launching a number of additional tools that bring more speed to other parts of Torch, as well. Some of these are modest, but many of Facebook’s projects results in 3 to 10x improvements over the default tools.
All of this is pretty technical, of course, and you can read more about the details here.
What matters, though, is that deep learning techniques (or at least their results) are slowly starting to show up in a lot of the software we use every day.
Google+ Photos, for example, uses it to allow you to find images in your photo library. And at CES last week, Nvidia spent most of its keynote discussing how it uses deep learning to classify objects that a camera on a car may see in order to further its research in autonomous driving.