Microsoft's AI and Speech Breakthroughs Eclipsed by New IBM Watson Platform
Researchers at Microsoft achieved what they say is a breakthrough in speech recognition claiming they've developed a system that's as effective or better than people with professional transcription skills. The software's word error rate (WER) is down to 5.9 percent -- an improvement from the WER of 6.9 the team reported in September. The milestone was enabled with the new Microsoft Cognitive Toolkit, the software that enables those speech recognition advances (as well as image recognition and search relevance). Microsoft announced both developments two weeks ago, though the timing wasn't the best as IBM was holding its huge World of Watson event in Las Vegas.
Watson, of course, is Big Blue's AI system made famous several years ago when it appeared on Jeopardy and, in advance of its latest rollout, made the talk-show circuit including CNN and CBS's 60 Minutes, where IBM Chairman, President and CEO Ginni Rometty talked up Watson's own achievements including the ability to discover potential cancer cures deemed not possible by humans, among other milestones.
Microsoft believes it has the most powerful AI and cognitive computing capabilities available. The Microsoft Cognitive Toolkit is the new name for what it previously called the Computation Network Toolkit, or CNTK. In addition to helping the researchers hit the 5.9 WER, the new Microsoft Cognitive Toolkit 2.0 helped the researchers enable what the company is calling "reinforcement learning."
The new open source release, like its predecessors available on GitHub, now supports Python including migration from C++. "We've removed the barrier for adoption substantially by introducing Python support," said Xuedong Huang, a Microsoft distinguished engineer, in a recent interview. "There are so many people in the machine learning community who love using Python."
Most noteworthy, Huang said is that the new software has a significant boost in performance, enabling it to scale across multiple Nvidia GPUs, including those added with the field-programmable gate arrays in the Azure cloud. Huang acknowledges the tool isn't as popular as other open source frameworks such as Google's TensorFlow, Café or Torch, but he argues it's more powerful, extensible and able to scale across multiple machines and environments.
"It' the fastest, most efficient distributed deep learning framework out there available," Huang said. "The performance is so much better. It's at least two to three times faster than the second alternative." The new Microsoft Cognitive Toolkit includes algorithms that can't degrade computational performance, he added.
For its part, IBM made some pretty big news of its own. The company released the new Watson Data Platform (WDP), a cloud-based analytics development platform that allows programming teams including data scientists and engineers to build, iterate and deploy machine-learning applications.
WDP runs on IBM's Bluemix cloud platform, integrates with Apache Spark, works with the IBM Watson Analytics service and will underpin the new IBM Data Science Experience (DSX), which is a "cloud-based, self-service social workspace that enables data scientists to consolidate their use of and collaborate across multiple open source tools such as Python, R and Spark," said IBM Big Data Evangelist James Kobielus in a blog post outlining last month's announcements at the company's World of Watson conference in Las Vegas. "It provides productivity tools to accelerate data scientists' creation of cognitive, predictive machine learning and other advanced analytics for cloud-based deployment. It also includes a rich catalog of learning resources for teams of data science professionals to deepen their understanding of tools, techniques, languages, methodologies and other key success enablers."
There are also free enterprise plans that include 10 DSX user licenses and a Spark Enterprise 30 Executor Plan, he noted. IBM claims more than 3,000 developers are working on the WDP and upwards of 500,000 users are now trained on its capabilities.
Has IBM already won the war against Microsoft, Google, Amazon and Facebook when it comes to intelligent cognitive computing, AI and machine learning? Karl Freund, a senior analyst at Moor Insights and Technology, said Microsoft has a long way to go to compete with IBM Watson for mindshare without question. "IBM is a brilliant marketing machine, and they are spending a lot of money to establish the Watson brand. Microsoft has nothing comparable, Freund said. "From a technology perspective, however, IBM has not convinced the technologists that they have anything special. In fact, most people I speak with would say that their marketing is ahead of their reality."
Freund said what IBM is offering is a collaborative development platform. "Microsoft is releasing their software as open source," he added. "IBM is all about the services, while Microsoft is seeking to gain broad support for their software stack, regardless of where you run it."
Microsoft's new toolkit and its multi-GPU support are significant, while Watson is likely to appeal to existing Big Blue shops including those with mainframes and organizations using the IBM SoftLayer cloud.
Posted by Jeffrey Schwartz on 11/07/2016 at 1:57 PM