Microsoft Neural Network defeats Google and Intel in image recognition competition

Microsoft Research excelled in several categories at the sixth annual ImageNet Image Recognition Competition. She managed to surpass competitive systems from Google, Intel, Qualcomm and Tencent, as well as from a number of startups and research laboratories ( results ).

The champion system is called "Deep Residual Learning for Image Recognition", and an article describing the technical principles of its work has been published in the public domain for free access .

“We trained a neural network with a depth of more than 150 layers,” the researchers describe the method. - At the same time, a deep residual learning framework was used, which facilitates the optimization and convergence of extremely deep neural networks. The method of deep residual learning allows you to get additional accuracy when the neural networks are much deeper than those used previously. This advantage in accuracy is not observed in many conventional neural networks when they are deepened. ”

In the illustration, a neural network with residual learning is shown in the right column.



Deep learning technologies are now being actively studied by many large corporations. With the help of neural networks, they increase the efficiency of internal systems and improve the quality of custom products. In a humorous style, Microsoft demonstrated the capabilities of its developments in recent applications for determining age and mustache rating. Commercialization of image recognition technology takes place through the API within the project "Oxford» ( by Project Oxford ), beta testing that began a month ago.



According to the terms of the ImageNet contest, the program must correctly detect and classify objects in 100,000 photos from Flickr and from various search engines, choosing from thousands of thematic categories (ant, banana, apple, etc.).

Microsoft's development showed a classification error rate of only 3.5%, and localization errors - 9%.

In previous years, the winners of the competition in terms of the classification of objects were Google, startup Clarifai and NEC.

“We did not even suggest that this idea alone [deep residual learning] could be so important,” said Jian Sun, one of the program’s authors, on the official blog .