Microsoft open-sources CNTK deep-learning toolkit

January 26, 2016 // 2:11 p.m.

Tags: #artificial-intelligence #cntk #deep-learning #google #machine-intelligence #microsoft #open-source #speech-recognition #tensorflow #xuedong-huang

Microsoft is following in rival Google's deep-learning footsteps, announcing the release of its Computational Network Toolkit (CNTK) under a permissive licence.

Deep-learning networks are big business, thanks to increasing computational power and the growing tide of what marketeers like to call Big Data. Following Google's decision to release its TensorFlow machine intelligence software library under an open-source licence, rival Microsoft has opted to do the same with its CNTK software.

Developed by a team led by Xuedong Huang as a means of improving the company's ability to recognise speech, CNTK is claimed to offer a significant performance boost over rival packages - including TensorFlow - thanks to an ability to scale beyond a single machine. Using a single GPU, CNTK is claimed to equal or beat rivals TensorFlow, Theano, Torch 7, and Caffee; across two GPUs, it is claimed to beat them all by a healthy margin; installed on the company's Azure platform and scaled to eight GPUs running on two virtual machines, it offered several times the performance of its rivals in internal testing. 'The CNTK toolkit is just insanely more efficient than anything we have ever seen,' Huang boasted of the results.

Internally, Microsoft uses CNTK on a GPU-driven high-performance computing cluster to drive its speech recognition systems. In April 2015, the company made the software available to academic researchers. Now, the software has been published on Microsoft's GitHub repository under a permissive licence of its own design allowing for free use, modification, publication, distribution, and even resale.

The release of the CNTK toolkit comes in the same week that the death of artificial intelligence pioneer Marvin Minsky, aged 88, was reported.
Discuss this in the forums

QUICK COMMENT

SUBSCRIBE TO OUR NEWSLETTER

WEEK IN REVIEW

TOP STORIES

SUGGESTED FOR YOU