AI Gets Smarter with Microsoft's Cognitive Toolkit 2.0
Microsoft's second release of its open source deep learning framework earlier this month adds support for Java bindings, Spark, and Keras.
- By Michael Domingo
Artificial intelligence just got smarter with the 2.0 release earlier this month of the Microsoft Cognitive Toolkit. This version adds support for Java Bindings, Spark, and Keras.
AI sounds like we might be talking about the future of computing -- but there's important work being done with it now. As an example, Microsoft researchers worked with a team of digital crime investigators with the Federal Trade Commission to thwart malware scammers, and key was the use the Cognitive Toolkit.
A typical scam being run currently is a pop-up ad that warns computer users of a virus on their computers, which in turn tells them to call or contact a company to help get rid of the virus. It's not just one company working these scams, but hundreds, and locating them is tricky, since those companies would often continue relocating their digital footprints. To track them down, the researchers and investigators used a mix of tools: the computer vision API from the Toolkit to scan and gather information at faster than human speed to eventually pinpoint the origin of the scams; Power BI to gather and slice the data into something meaningful; Azure to host and run the apps.
(The full story of the discovery, capture, and prosecution covers quite a few other Microsoft technologies; you can read the whole story on The Official Microsoft Blog here.)
With the Cognitive Toolkit at 2.0, there are a few more new features:
- Support for Java Bindings, Spark: Besides evaluating models in Python, BrainScript, or C#, Cognitive Toolkit models can now be evaluated with a new Java API, which means it can now be integrated into Java-based deep learning models and can be used on the Spark platform.
- Model Compression: The Toolkit now includes "extensions that allow quantized implementations of several FP operations, which are several times faster compared to full precision counterparts," which makes evaluating models on lower-powered computing devices such as smartphones much more accurate.
- Keras Support in Preview: The added Keras support is also new, and being tested in a public preview. The open source Keras neural network library is popular with developers for its user-friendliness when used for deep learning applications, and now that the Cognitive Toolkit supports it, code that's written with Keras can easily be used by the Toolkit without having to change it.
Details on these features is in this blog post from the Cognitive Team.
About the Author
Michael Domingo is a long-time software publishing veteran, having started up and managed several developer publications for the Clipper compiler, Microsoft Access, and Visual Basic. He's also managed IT pubs for 1105 Media, including Microsoft Certified Professional Magazine and Virtualization Review before landing his current gig as Visual Studio Magazine Editor in Chief. Besides his publishing life, he's a professional photographer, whose work can be found by Googling domingophoto.