Microsoft: AI Isn’t Yet Adaptable Enough to Help Businesses
Microsoft’s top research executive says it’s too difficult to customize powerful machine-learning systems to an individual company’s needs.
The AI revolution may take longer than some expect to spread from Silicon Valley into other industries.
Recommended for You
Recent breakthroughs in machine learning have let tech giants such as Microsoft, Google, and Facebook build impressive new businesses and products powered by software that parses text and images. Some have launched cloud services they say can “democratize AI” by helping other companies do the same.
But Peter Lee, vice president at Microsoft’s research division, said this week that the most valuable, high-end machine-learning systems so useful to tech giants are still too inflexible and expensive for the company to offer its business customers.
“We are right now in terms of enterprise application of machine learning and AI concepts in an in-between spot,” said Lee at MIT Technology Review’s EmTech Digital conference in San Francisco this week. “AI is not a technology that has been reduced to practice; we have a small and highly paid cadre of craftsmen building the bespoke solutions.”
Lee cited AI-powered tools Microsoft has developed to help its salespeople close deals as an example. “I would like nothing more than to be able to sell those to every other company in the world,” he said, ruefully. That’s impossible right now, he added, because the technology would need to be laboriously customized to each new business.
Building software that uses AI to learn and adapt to the needs and conditions inside a particular business might seem like the answer. But Lee said Microsoft’s experiences trying to have software learn from the real world show that the technology isn’t currently mature enough.
When Microsoft launched a chatbot on Twitter that could learn from conversations last year, it quickly picked up racist language, for example. Microsoft’s translation system ruffled feathers last summer after analyzing online text led it to start translating “Daesh”—an Arabic name for Islamic State—into “Saudi Arabia” in English.
“These machine-learning systems have become more spectacular in their failures,” Lee said. Setting learning software loose in real-world situations still requires close supervision from expert humans.
There is clearly still money to be made selling more straightforward AI services to businesses. Microsoft’s offerings include services that do things like count faces in an image and convert speech to text. At the EmTech event, startup Clarifai showed off tools that help companies search and manage collections of images and video by understanding their contents.
Yet Lee implied that products like those can’t deliver the revolution—or new profits—that have been generated through tech giants’ internal projects in developing AI technology. “Most times the high-value applications of machine learning in the enterprise are the custom or bespoke solutions,” he said. “Every enterprise is a special case, so you want these to adapt to the special needs and conditions.”
Lee said work on software that learns by trying out different actions and seeing what works, rather than digesting static data such as text or images, might eventually get the industry out of this bind. He referred to how Google used reinforcement learning to master the game of Go and renewed interest in algorithms loosely inspired by evolution, recently highlighted by the independent research institute OpenAI.
“They’re not limited by scraping or mining the output of human intelligence,” he said. “You might conclude these systems have a better chance of understanding a business process and developing its own strategies.”
Keep up with the latest in artificial intelligence at EmTech MIT.
Discover where tech, business, and culture converge.
September 11-14, 2018
MIT Media Lab