Do your IT experts know what the company's machine learning algorithms are doing? They almost certainly do not, according to Sam Madden, professor of electrical engineering and computer science at MIT -- and that's a problem.
"If you can't understand why the algorithm is going to work, when it's going to work or, worse, when it's going to fail, then you're going to have to be very cautious about putting these things into practice," Madden said at the recent MassIntelligence conference in Boston. The conference was a joint effort between the Massachusetts Technology Leadership Council and MIT to bring industry and academic experts together to discuss advances in artificial intelligence (AI).
Thanks to the enormous stores of data now available for crunching, massive compute power and better algorithm performance, machine learning has made impressive advances in image, speech and pattern recognition. There are machine learning applications that can read lips, beat the reigning world champion in the game of Go and see well enough to drive a car -- sort of.
But here's the catch, Madden explained: Most machine learning systems require Ph.D. expertise to make them operational; and even with expert guidance, machine learning algorithms are opaque, making it difficult to know why an algorithm arrived at the conclusion it did or what it's learned, according to Madden.
Making matters more obscure, the data scientists who develop the algorithms often work independently from the engineers who deploy them, Madden said. The separation between the two groups makes it difficult to adapt the models over time and can even lead to failures.
The infamous case in point: Microsoft's Tay a conversation bot targeted at 18 to 24 year olds, began making racist and sexist comments 24 hours after it launched. The bot, which made MIT Technology Review's worst-tech list of 2016, was programed to mimic those it engaged with and, so, was taught the behavior. In another example, researchers tricked an image recognition algorithm to incorrectly identify images. Rather than a poodle, the algorithm saw an ostrich.
"If I can't understand why [the algorithm] said, 'This is an ostrich,' then it's hard for me to have a lot of confidence if I want to put this algorithm into production on some website," Madden said. And if an AI-forward company like Microsoft has to shut down algorithms due to unintended behavior, other industries might want to think twice about embracing the technology, he said.
The computer science and artificial intelligence laboratory, aka CSAIL, at MIT wants to shed light on the black box of today's machine learning systems with a new initiative, SystemsThatLearn@CSAIL. The initiative, which Madden co-directs, is focused on accelerating the development of the next generation of machine learning algorithms, which he described as large-scale machine learning systems that perform complex, humanlike tasks. Think conversational systems or self-driving cars.
If these next-gen algorithms are going to survive outside of lab environments, transparency is vital. "People are going to have to be cautious about letting a car drive them 100% of the time, unless the algorithm is just really good and it can understand when it's going to work well and when it's not going to work well," Madden said. "If it just randomly fails 1% of the time, that's not going to acceptable to people."
In its quest to shed light on machine learning's black box, SystemsThatLearn@CSAIL had to break down some academic barriers. The program joins the research teams that develop algorithms at MIT with the research teams that develop the large-scale systems the algorithms run on. "These two groups are sort of like oil and water," Madden said. "We're trying to bring them together ... to solve these really hard problems."
Madden told MassIntelligence conference goers that the initiative will also work on building tools and algorithms for large-scale systems, including how machine learning can optimize the operations of these systems. He provided an example of researchers applying machine learning to the TCP protocol, which controls how packets are transmitted on the internet, to boost its performance.
AI and tide tables
Platform providers, like Google and Amazon, were also a point of discussion at the AI event. They are making it easy to inject AI into the business via APIs. That's good news for small companies and startups that may not have the manpower to develop machine learning or artificial intelligence algorithms on their own, according to Mike Frane, vice president of product management at Windstream, an IT services and communications provider in Little Rock.
"Startups don't have to focus on more of the mundane plumbing of AI," he said. "They can get hyper-focused on the differentiator they want to introduce to the marketplace."
Robert Mawrey, who works on the internet-of-things platform at MathWorks Inc. in Natick, Mass., agreed, pointing to a pet project of his as an example of how ready-made tools can be a time-saver.
Mawrey is somewhat of a tide-table expert. For years, he's been studying how the wind affects tide predictions. "When the wind blows, you can almost toss the tide forecast out the window," he said. He wanted a more accurate tool. He likened his tide project to that of a startup trying to get its business off the ground. "Limited resources, one guy, a complicated problem: How am I going to solve it?" he said.
Mawrey, who isn't a trained AI expert, took the years of data he'd gathered and applied a ready-made machine learning algorithm -- from MathWorks, of course -- to the problem. "I came up with something pretty good," he said.
AI in the ER
Artificial intelligence's predicted impact on the enterprise
Deep learning's new lease on life