Microsoft today announced that it is making it easier for developers to use its Computational Network Toolkit (CNTK) to build their own deep learning applications. The company first open sourced this toolkit in April 2015, but at the time, it was hosted on Microsoft’s own CodePlex site and was only available under a restrictive academic license. Now, the team is moving the project to GitHub and to the MIT open source license.
This tool is behind of Cortana and Skype Translations
CNTK is an open-source deep-learning toolkit that became available back in April 2015. However, when it was still on CodePlex, it was restricted by an academic license, which means that it was virtually unused beyond scholarly use. By uploading the project on GitHub, Microsoft will be able to get more developers on the action.
While Microsoft’s old license made the project accessible to academics, it wasn’t really geared toward production usage and tinkering outside of the academic environment. With this new license — and by having the project on GitHub — Microsoft hopes to attract other users as well.
As Microsoft’s chief speech scientist Xuedong Huang notes in today’s announcement, CNTK is highly optimized for speed. “The CNTK toolkit is just insanely more efficient than anything we have ever seen,” Huang said. Those other projects Huang is referring to include the likes of Google’s recently open-sourced TensorFlow, as well as projects like Torch, Theano and Caffe.
Microsoft argues that one of the advantages of CNTK is its ability to run on a single core, as well as on a large cluster of GPU-based machines. The company also says that it can scale across more machines than other projects (but that’s obviously a claim we can’t exactly verify).
The company says that one of the reasons why it has the upper hand against others is because it can run on a single core or on a “large cluster of GPU-based computers,” according to Microsoft Principal Development Manager Chris Basoglu, who’s involved with the toolkit. More to the point, the researchers argue that “it can scale across more GPU-based machines than other publicly available toolkits,” which offers a huge advantage when it comes to large-scale experiments or computations.
The team discovered that GPU is efficient in processing algorithms used in technology that can recognize images and movements as well as speak, hear and comprehend speech. As such, Microsoft uses powerful computers with impressive GPUs to run CNTK internally. For example, the company’s popular virtual assistant Cortana takes advantage of CNTK for speech recognition.
“We further introduce the computational network toolkit (CNTK), an implementation of CN that supports both GPU and CPU. We describe the architecture and the key components of the CNTK, the command line options to use CNTK, and the network definition and model editing language, and provide sample setups for acoustic model, language model, and spoken language understanding,” the team says [pdf] in a publication.
While it’s important for Huang and his team to deliver the internal needs of Microsoft via a tool such as CNTK, they also want other developers who are geared toward deep learning to have access to the same resources.
“The reason we did this is we want to give our users flexibility to make changes. That strengthens the ecosystem and the toolset we have,” Huang.
With the latest development, it appears that the advancements in artificial intelligence will certainly have a better overall progress, as it encourages more researchers to join the mix.
AI algorithm learns can to write political speeches
At the University of Massachusetts Amherst, Valentin Kassarnig has created an artificial intelligence machine that has learned how to write political speeches that are similar to real speeches.
Kassarnig used a database of almost 4,000 political speech segments from 53 U.S. Congressional floor debates to train a machine-learning algorithm to produce speeches of its own.
According to MIT Technology Review, the speeches consisted of more than 50,000 sentences each containing 23 words on average. Kassarnig categorized the speeches by political party and by whether they were in favor of or against a given topic.
Here is an example of a speech it automatically generated:
“Mr. Speaker, for years, honest but unfortunate consumers have had the ability to plead their case to come under bankruptcy protection and have their reasonable and valid debts discharged. The way the system is supposed to work, the bankruptcy court evaluates various factors including income, assets and debt to determine what debts can be paid and how consumers can get back on their feet. Stand up for growth and opportunity. Pass this legislation.”