What's new in Microsoft Cognitive Toolkit for deep learning?



The Cognitive Toolkit is an open source toolkit, focused on deep learning and artificial intelligence similar to Google's TensorFlow, which Microsoft released last year. While the initial version was remarkable in terms of speed, the second edition made available on Thursday puts more emphasis on usability with support for Python and Keras neural networking library.

It also includes the addition of Java language bindings for model evaluation and tools that compress trained models to run in real time even on resource-constrained devices including smartphones.

The updated Cognitive Toolkit not only provides a number of new features, but bug fixes too, albeit it isn't entirely backwards compatible due to deprecated and renamed aspects of the toolkit.

With Google's TensorFlow already supporting Keras, means this new release would make it possible for data scientists to easily port their code between two different backends.

Unlike other deep learning toolkit, Microsoft Cognitive Toolkit is built to work reliably with massive production data out of the box, and most AI workloads within Microsoft, from Cortana to Bing to the Emotion API in Cognitive Services, were all created using the toolkit.

Originally built for speech recognition systems, Cognitive Toolkit is very good at working with time series data for building recurrent neural nets.

Microsoft, however is battling with a bunch of other programming frameworks backed by other tech giants like Google's TensorFlow, which is perhaps the most popular open source machine learning framework, Amazon's Apache MXNet, and Facebook Caffe2 framework.

With this update, the Cognitive Toolkit’s performance for other kinds of neural nets was not just improved, but also the groundwork for making it easier to extend the system in the long.
Previous
Next Post »