Ahead of its Build conference, Microsoft today released a slew of new machine learning products and tweaks to some of its existing services. These range from no-code tools to hosted notebooks, with a number of new APIs and other services in-between. The core theme, here, though, is that Microsoft is continuing its strategy of democratizing access to AI.
Ahead of the release, I sat down with Microsoft’s Eric Boyd, the company’s corporate vice president of its AI platform, to discuss Microsoft’s take on this space, where it competes heavily with the likes of Google and AWS, as well as numerous, often more specialized startups. And to some degree, the actual machine learning technologies have become table stakes. Everybody now offers pre-trained models, open-source tools and the platforms to train, build and deploy models. If one company doesn’t have pre-trained models for some use cases that its competitors support, it’s only a matter of time before it will. It’s the auxiliary services and the overall developer experience, though, where companies like Microsoft, with its long history of developing these tools, can differentiate themselves.
“AI is really impacting the way the world does business,” Boyd said. “We see 75% of commercial enterprises are doing more with AI in the next several years. It’s tripled in the last couple years, according to Gartner. And so, we’re really seeing an explosion in the amount of work that’s coming from there. As people are driving this forward, as companies are driving this forward, developers are on the front lines, trying to figure out how to move their companies forward, how to build these models and how to build these applications, and help scale with all the changes that are moving through this.”
What these companies — and their developers — need is more powerful tools that allow them to become more productive and build their models faster. At Microsoft, where these companies are often large enterprises, that also includes being able to scale up to the needs of an enterprise and offer the security guarantees they need.
As companies start adopting machine learning, though, they are now also getting to a point where they have moved from a few tests to maybe running a hundred models in production. That comes with its own challenges. “They are trying to figure out how to manage the life cycle of these models,” he said. “How do I think of the operational cycle? How do I think about a new model that I’m ready to deploy? When is it ready to go?”
Only a few years ago, the industry started moving to a DevOps model for managing code. What Microsoft essentially wants to move to is MLOps for managing models. “It’s very similar to DevOps, but there’s some distinct differences in terms of how the tools operate,” Boyd noted. “At Microsoft, we’re really focusing on how do we solve these problems to make developers way more productive, using these enterprise tools to drive these changes that they need across their organization.” This means thinking about how to bring concepts like source control and continuous development to machine learning models, for example, and that will take new tools.
It’s no surprise then that adding more MLOps capabilities is a major part of today’s releases. The company is integrating some of these functions into Azure DevOps, for example, that allows them to trigger release pipelines. The company is also giving developers and data scientists tools for model version control, for example, to track and manage their assets and to share machine learning pipelines.
These are very much tools for advanced machine learning practitioners, though. On the other side of the spectrum, Microsoft also announced a number of automated machine learning tools, including one that essentially automates all of the processes, as well as a visual model builder, which grew out of the Azure ML Studio. As Boyd told me, even companies like British Petroleum and Oregon’s Deschutes Brewery (try their Black Butte Porter if you get a chance) now use these tools.
“We’ve added a bunch of features into automated machine learning to simplify how people are trying to use this kind of work,” Boyd noted.
Microsoft today also launched a number of new services in its Cognitive Services lineup, including a new personalization service, an API for recognizing handwriting and another one for transcribing conversations with multiple speakers. The personalization service stands out here because it uses reinforcement learning, a different machine learning technique from most other Cognitive Services tools, and because it is far easier to implement than similar services. For business users, there’s also the Form Recognizer, which makes extracting data from forms easy.
What’s more interesting that the specific features, though, is that Microsoft is shifting its emphasis here a little bit. “We’re moving away from some of the first-level problems of ‘here’s the table stakes, you have to have an AI platform,’ to much more sophisticated use cases around the operations of these algorithms, the simplification of them, new user experiences to really simplify how developers work and much richer cognitive services,” Boyd explained.