New version pushes Python language
All major companies investing in deep learning have some sort of packaged software that helps developers create neural network models. Google has TensorFlow, DeepMind released Sonnet, Facebook recently announced Caffe2, and Amazon touts MXNet.
Microsoft’s version was previously known as CNTK, and it was released as a beta in October last year. The new Cognitive Toolkit 2.0 has been expanded to support Python and Keras, a popular neural network API.
It’s a good move, considering Keras is a high-level framework, allowing researchers to experiment and build their models quickly. “[It] abstracts away a lot of details that most users don’t need to know about,” saving users from messy lines of code, François Chollet, the author of Keras and a Google engineer, previously explained to The Register.
Microsoft’s AI strengths lie in speech recognition. Last year, a team led by Xuedong Huang, Microsoft’s chief speech scientist, announced they had created a system capable of transcribing from speech at “human parity” level.
It’s no surprise that Cognitive Toolkit is particularly suited for recurrent neural networks and long short-term memory networks, both commonly used in natural language processing. Microsoft has made it easier for developers to branch out to different types of neural networks with Keras.
It will also include Java language bindings and new tools that compress trained models so it takes less resources to run applications like image recognition.
It’s already used internal applications like Cortana, its digital assistant, and is part of Microsoft’s wider goal to make AI more accessible.
The toolkit is open source and can be found here. ®