Microsoft has posted on Github CNTK toolkit for deep learning

Microsoft has released the open source source code of tools that the company uses to accelerate development in the field of artificial intelligence: the Computational Network Toolkit kit is published on Github .

According to the official blog, researchers had to create a set of CNTK tools out of necessity. Lead Microsoft Speech Recognition Specialist Xuedong Huang says that the developers in his group were concerned about how to speed up the process of speech recognition by computers, and the tools available were too slow.

Therefore, a group of volunteers volunteered to solve the problem on their own, using their own solution, which put the performance in the first place.
')
Efforts paid off in full.

In internal tests, CNTK showed better performance than four other popular computing tools that developers use to create models of in-depth training in tasks like speech recognition and pattern recognition, due to better communication capabilities.

“The CNTK toolkit is incredibly more efficient than anything we have ever seen,” Xuedong Juan admits.

A performance increase of this kind is extremely important for a rapidly developing field of depth learning, because some of the most difficult tasks of depth learning can be calculated for weeks.



Over the past few years, the field of deep learning has taken a big step forward, as more and more researchers are launching machine learning algorithms using deep neural networks. Many believe that in-depth training has the potential for improvements in artificial intelligence, Microsoft writes.

The progress achieved has allowed us to create systems that accurately recognize and even translate speech , as well as image recognition systems that can even answer questions about them.

At Microsoft, the CNTK toolkit is used on a cluster of powerful computers with a GPU. It turned out that GPUs are ideally suited for processing such algorithms, thanks to them there was such a breakthrough in the technologies of synthesis, speech recognition and understanding, image recognition and movement on video.

Everyone can benefit from CNTK tools: from a developer with a limited budget and a single computer to a company capable of raising a whole cluster of servers from a GPU. The Microsoft developers claim that CNTK scales up to work in parallel on many GPU machines much better than other publicly available tools of this kind.

In April last year, Microsoft opened the CNTK for academic researchers through Codeplex under a more limited free license. Since Monday, CNTK is available absolutely for everyone.

Independent developers express the pleasure of such positive actions by Microsoft, but at the same time they do not miss the opportunity to pin up the company: “Only Microsoft could open the githaba docx documentation,” writes one of them.