advertisement
advertisement

Facebook’s Image-Recognition Tech Is Teaching 40,000 Images A Second To Understand Context

The company said it was able to train industry standard data sets that used to take days in just an hour.

Facebook’s Image-Recognition Tech Is Teaching 40,000 Images A Second To Understand Context
[Photo: Flickr user Carlos ZGZ]

The ability to process things like digital camera effects, image filters, augmented reality, medical imaging, and much more could be a whole lot faster, thanks to research being done by Facebook.   

advertisement

If you’re a regular Facebook user, you’ve no doubt noticed that over the years, there’s been more and more emphasis on images and on automatically identifying the people and the objects in them. For the service’s 1.94 billion monthly users, artificial intelligence and machine learning are behind the ability to quickly surface meaningful baby pictures, vacation selfies, and pet action photos.

Now, the company’s AI and machine-learning teams have developed methods for training data sets—the procedure that’s used to teach visual recognition models to distinguish between large numbers of images—that are faster than anything else available today. The company said today it has come up with a system that’s capable of training 40,000 images per second per machine, making it possible to train on a 1k data set—the industry standard training set—in less than sixty minutes, and with no loss of quality. Until now, that was something that could take days, or even months to do.

The work is key, given that every tech company is handling a rapidly increasing amount of data these days, and counting on AI and machine learning to parse it. But with that increasing amount of data has come longer training times, something Facebook says has led to slowdowns in research and development time.

By moving to a system with 256 graphical processing units from one with 8 GPUs, Facebook was able to achieve a scaling efficiency of 90%.

Keeping with its philosophy of sharing its work freely and widely, Facebook says it’s open-sourcing the hardware stack it developed to achieve the improved training times. And that could benefit many companies that are taking in massive amounts of visual data and need to process it quickly. From digital camera tools aimed at consumers and advanced medical screening to helping develop autonomous cars and much more, this development could speed up R&D and make things easier for many people and companies.

About the author

Daniel Terdiman is a San Francisco-based technology journalist with nearly 20 years of experience. A veteran of CNET and VentureBeat, Daniel has also written for Wired, The New York Times, Time, and many other publications.

More