Microsoft is adding Python language support to its open source deep learning toolkit for developers. The kit, formerly known as Computational Network Toolkit (CNTK) and now called Microsoft Cognitive Toolkit, goes into a beta release today and features reinforcement learning for neural networking.
The Python backing extends the toolkit beyond C++ development for better accessibility. C++ "was good for really serious programmers," said XD Huang, Microsoft chief speech scientist, but "our mission is simple: to democratize AI," not only for web developers, but for different disciplines, even sociologists. Python support will offer extensiveness, expressiveness, and flexibility.
With reinforcement learning, an agent discovers the correct way to do something, such as navigate a room or form a sentence, via trial and error. A neural network is trained to interact with an environment with only occasional feedback in the form of a reward. In essence, training involves adjusting the network's weights to search for a strategy that consistently generates higher rewards. Reinforcement learning, the company said, could eventually lead to true artificial intelligence, in which systems make complex decisions on their own.
Thanks to a performance boost, Microsoft now claims leadership in working large data sets across multiple machines, as the kit leverages CPUs and Nvidia GPUs, and supports cloud deployment. Compared to that of the previous version, performance is nearly twice what it was before when scaling to eight Pascal GPUs in an Nvidia DGX-1, which is billed as an "AI supercomputer" by Nvidia.
Deep learning is an AI technique for developers and researchers to use large volumes of data, called training sets, to teach computers to recognize patterns from inputs including images and sounds. With the kit, large-scale deployments can be supported for developing both consumer and professional applications. It supports Microsoft's Azure's GPU cloud platform, currently in a preview stage.
Microsoft said it has used its deep learning kit for such purposes as a speech recognition system that recognizes words in conversation as well as a person does, matching or besting the error rate of professional transcription. The toolkit has since moved beyond speech recognition and into a variety of deep learning tasks, Microsoft said. Cortana, Microsoft's answer to Apple's Siri personal assistant, runs atop of CNTK, and cooling specialist Liebherr has used it in an application that recognizes individual food items in a refrigerator for use in taking inventory.