Microsoft executives say the next frontier in artificial intelligence will involve using human professionals’ expertise to train machine learning models. “Machine teaching” is essentially an interface that sits atop a machine learning layer to give people without the ability to code a way to train and deploy systems.
Machine teaching can be applied to a number of areas, including text classification, conversational dialog, computer vision, and robotics, Microsoft VP for Business AI Gurdeep Pall told VentureBeat in a phone interview.
Recent AI startup acquisitions and years of Microsoft Research work will fuel initiatives to make machine teaching more widely available in the future, he said. Microsoft Research created the machine teaching group promising such capabilities in 2015.
Initiatives in these areas will rely on nearly a decade of work by Microsoft Research and AI Group, existing tools like the company’s AirSim simulator, and Bonsai and Lobe, San Francisco Bay Area AI startups Microsoft acquired last year. Lobe focuses on deep learning without code, while Bonsai is designed to help enterprises train systems used in manufacturing, building management, and robotics.
Pall said Microsoft machine training will also seek to incorporate the Robotic Operating System (ROS), the popular robotics OS that came to Windows 10 last fall.
Potential use cases include nurses teaching a robotic arm how to move patient samples or experts helping automate the movement of oil drilling platforms, Pall said.
“I think AI has to break away from being something that only AI experts can work on, otherwise the application of the AI and its impact on the world is going to be very limited. I think we have to get to the place where it’s not that 10,000 AI experts in the world are the only ones who can actually apply AI in different scenarios across many vertical domains,” he said.
Tools that increase access could empower millions of people to create AI models, Business AI general manager Mark Hammond said in a blog post today.
If AI is going to become more esoteric and harder to understand, it is heading in the wrong direction, Pall said, adding that machine teaching is meant to transfer knowledge from a human expert’s brain to machine learning algorithms in fields like finance, manufacturing, and customer support.
“I think there have to be some very concrete steps to make it a lot more accessible to the domain experts who really don’t understand what is happening inside the engine but … can actually specify the tasks, train the machine, and see whether it is working well or not and accordingly tweak it,” he said. “We believe this fits this really important need in this chasm that exists today between the AI experts and domain experts in the world.”
While AI advances in recent years have emphasised machine learning frameworks that require an understanding of Python — like TensorFlow and Microsoft Cognitive Toolkit — companies like Microsoft have been making platforms and tools to help people who don’t know how to code, train, and deploy AI models.
Additional details about upgrades to Microsoft’s tools for training AI models are expected at the Build annual developer conference being held May 6-8 in Seattle.
The company’s commitment to machine teaching follows the release of several Google Cloud solutions, including the debut of the Anthos hybrid cloud platform, Cloud AutoML Tables, and BigQuery ML for AI with tabular data.
Microsoft plans to distinguish itself from competitors by remaining focused on machine teaching aimed at solving real-world problems, Pall said.
“I think of it as a horizontal thing,” he said. “Whether it’s from building customer service bots or autonomous systems or identifying objects in computer vision tasks, all those things can be done better or faster with a machine teaching approach with a human expert in the loop.”