Within just a few years, autonomous vehicles will likely be able to handle most of the driving, allowing humans to take their eyes off the road and step in only occasionally. Automakers refer to this level of autonomy as “Level 3.” But to get to that point, developers must solve how to enable humans and artificial intelligence-based systems to work as true partners effectively, conveniently and safely.
This past summer, Jian Hu, UM-Dearborn professor of industrial and manufacturing systems engineering, and Raameen Ahmed, a junior studying data science, explored the problem as part of the Summer Undergraduate Research Experience program. They partnered with Ford Motor Company to address a fundamental hurdle in AI technology: humans often don’t know what leads an AI system to make a given decision.
“Generally, today's AI systems are what we call a ‘black box,’” Hu says. ”Humans provide data and the AI produces a decision, but we don’t know what factors led to that decision. To produce more meaningful AI responses, we want what’s called ‘explainable AI,’ where we know not just the final decision but the factors that led to it.”
Explainable AI could make it easier for developers to debug autonomous systems, and it could also assist systems to become better at learning on the fly from human input.
The SURE project focused on achieving that goal in systems that process visual data, like the technology that interprets streams of incoming camera data in autonomous vehicles. The research team, which also included industrial and systems engineering graduate student Ming Gao, developed an algorithm that enables AI to analyze which attributes of a given image led it to classify that image in a certain way — for example as a road marking, an obstacle or a stop sign.
Ahmed says the SURE research not only enabled her to address a key industry challenge, it also offered a look at the ins and outs of research. The SURE program pairs professors with undergraduate students for up to 12 weeks to learn about the life of an active researcher and research-oriented careers, and to develop long-term mentor relationships.
“As someone who plans to get a PhD, I think this program is really something special,” Ahmed says. “It gives younger students or newer students an opportunity to really get into the industry, which I've appreciated. And there are so many helpful people to learn from.”
Students work 15-20 hours per week during the research period, receiving a stipend. Hu says it’s an important opportunity for students to take what they’ve learned beyond the classroom and tackle the kinds of challenges they might face in industry or academic careers.
“In real-world projects like this, there are a lot of difficult questions that require in-depth collaboration to work through, and this is an important way for students to gain that experience,” he says. “And the kinds of visual simulations we worked with in this project are widely used in industry, so Raameen was able to pick up a lot of career-relevant information.”
Hu explains that the SURE team’s project works by continuously ranking and re-ranking the importance of various attributes in its decision-making process, tweaking which are most important as conditions change. The technique could be especially useful in the human-machine interplay that defines Level 3 autonomy.
“Imagine you’re riding in a Level 3 vehicle on a snowy day; you might have to step in to prevent the vehicle from rolling through a snow-covered stop sign,” Hu says. “Our technology could enable the vehicle’s systems to learn from that human input, determining that the factors it usually uses to classify an image as ‘stop sign’ don’t work in the snow. It could then adjust its decision-making, finding another attribute that can identify ‘stop sign’ even in the snow.”
The team recently developed a paper detailing the technology. They believe their algorithm could help speed wider adoption of Level 3 connected vehicle capabilities.
“Level 3 connected vehicles are intriguing because they use what’s called distributed computing — using a combination of onboard systems and outside assistance from the cloud,” Hu says. “I think a wider rollout within the next few years is very feasible, particularly if we can improve the data infrastructure that connects vehicles to the cloud.”
He says the algorithm he and Ahmed developed could also be useful in other applications where humans and AI work together with visual data, such as systems that help doctors interpret MRI data. The team has applied for a patent through U-M’s Innovation Partnerships.
“I think it's the real-world impact that really interested me in this project,” Ahmed says. “It’s exciting to think that even a small advance could eventually have a major positive impact on peoples’ lifestyles.”
The SURE Showcase — featuring nearly 50 research projects across the university's four colleges — will take place from 4 to 6 p.m. today in the Renick University Center’s Kochoff Hall.
You can read about additional SURE projects from the College of Arts, Sciences and Letters, the College of Business, as well as the College of Education and Health and Human Services.
Article by Gabe Cherry
