Facebook’s AI Training Models can Process 40,000 Images Per Second

Most of the people familiar with Facebook’s different technology is familiar with the social network’s image recognition technology. However, for a long period of time, artificial intelligence researchers at the company have been working on to make the image recognition technology faster and recently, the researchers have shared results of their work.

The social networking company has announced the results during an event earlier. During the event, Pieter Noordhuis said that they were able to go through whole ImageNet dataset in under an hour. Researchers at the company achieved the results with help of Facebook’s custom GPU and new algorithms. What more, despite the speed, the results didn’t show any loss of accuracy. Talking about the results, Noordhuis said, “You don’t need a proper supercomputer to replicate these results.”

The new model works to associate images with words. In order to achieve the desired results, researchers entered a description of the images and the system showed all of the images with the associated classification without any kind of description tags. In short, the system is still learning how to associate images with descriptive words.

This is early to produce a verdict but it seems that the company is working to introduce the model to analyze even bigger dataset with billions of user uploaded images. Interestingly, the company had announced that the breakthrough is open-source thus anyone will be able to reap the benefits of the model. The company is currently planning on releasing more details about the training work in the Facebook Research page.

Featured Image: Pexels

Related Articles