In a challenge to correctly detect images from 100,000 photographs from Flickr and popular search engines, Microsoft bested the likes of competitors Google, Intel, Qualcomm, and Tencent. Microsoft was able to accomplish this using a method known as deep neural networks, which trains computers to recognize these images.
Apparently already a common technology amongst researchers in this particular field, Microsoft’s researchers system was more effective because it allowed them to use extremely deep neural networks, which can be up to 5 times deeper than systems any previously used.
Of the three categories entered by Microsoft in the ImageNet challenge: classification, localization and detection, Microsoft’s team came in first place in each. In the Microsoft Common Objects in Context challenge, a project initially funded by Microsoft but now run by other academics, the Microsoft team again won first place for image detection and segmentation. Microsoft’s winning entry had a localization error rate of just 9% and a classification error rate of 3.5%.
The contest, now in it’s sixth year, is organized by researchers from top universities and corporations around the world, and is becoming the go to standard for the field of research.
The program is also rather proficient at recognizing speech patterns, being the underlying software in Skype’s real-time translating capabilities.