Microsoft makes deep learning toolkit open source

To accelerate the development of artificial intelligence with Microsoft’s Computational Network Toolkit, the natural language research division has made the CNTK toolkit open source. According to its creators, CNTK is much more efficient than four other machine learning toolkits.

Microsoft made the toolkit public before, but under a much stricter license. CNTK is licensed through GitHub under MIT, giving others more freedom. Making the toolkit further open source is the result of the successes that have been achieved since the kit was released for academic research in April last year.

According to the head of the speech department Xuedong Huang, the toolkit is many times faster than comparable machine learning toolkits. The comparison is made with Theano, TensorFlow, Torch 7 and Caffe. Google recently made TensorFlow open source.

Source: Microsoft Blog

Also, CNTK would be the only toolkit that is scalable across more machines. In the test in which CNTK is almost twice as fast as its own program with four GPUs in one machine, it scales over two machines with four GPUs each. According to the researchers, the program can be scaled over large clusters with many GPUs.

The program runs on Windows and Linux and is written in C++ and works with Nvidia’s cuDNN 4 library, but can also run on CPUs, as can be read in a presentation on CNTK late last year that also indicated that the toolkit move to GitHub in the foreseeable future.