Engineers are problem solvers. They seek out problems, analyze them, and work to create logical, lasting solutions. Alina Zare, Assistant Professor in MU’s Electrical and Computer Engineering department, has spent the last few years solving the problem of how to get computers to learn from data – and, more ambitiously, how to get them to learn from uncertain or unclear data.
“A tagline that we say all the time is ‘What’s the story?’ or ‘You’ve got to tell a story.’” Zare notes. “When you’re presenting an algorithm or a solution to a particular problem, you’ve got to be able to lay out the whole story. What’s the problem, what have people tried to do, why it doesn’t really solve the problem yet, and then how your solution tries to fill that hole and what to do next.” The story of her recent work with machine learning begins with this problem: “People can learn from stuff that’s not very specific,” Zare observes, “but can we get a computer to be able to do that as well?”
Solutions to these problems and others are addressed at TigerSense, the Machine Learning and Sensing Lab at the University of Missouri. The lab “focuses on machine learning, pattern recognition, computational intelligence, signal processing, information fusion and related areas with application to sensing (primarily, understanding sensor data),” according to its website. These studies have applications as far reaching as landmine detection, ocean floor exploration, and agricultural advancement.
Supervised learning is a type of machine learning in which a computer takes in certain inputs and produces a desired output. One hindrance to supervised learning is inaccurate labeling and uncertainty within data. An inability to see or generate correct label information, high cost of collecting label information, and simple human error often leads to data not being labeled correctly — but Zare doesn’t believe that should mean a data set should necessarily be thrown out. She’s developing algorithms that will teach computers to learn from nonspecific and incorrect data.
The story of data uncertainty can be told in many different ways. For example, agricultural imaging that is collected over a large area can tell people where more fertilizer is needed, where weeds are growing, or where crops are dying. Teaching a computer to distinguish between diseased and healthy crops requires providing it with a breakdown, pixel by pixel, of what each plant looks like. This can be costly and time consuming. The information that does exist is often approximate, or it’s unclear how much of the material there relates to the problem. To date, no method for solving this supervised learning problem to an extent that can handle large amounts of uncertainty has been devised, so Zare is pushing the boundaries and working to create algorithms that can.
Zare’s research led to a proposal that won the National Science Foundation’s prestigious Early Career Development (CAREER) award, providing her more than $450,000 to develop solutions to these problems. The plotline of this story includes resolving unclear mapping data. She and her team placed targets on a college campus and then flew over the scene taking images using various technologies. On the ground, GPS units could identify where those targets were located but with an inaccuracy of up to five meters. To compensate, one would have to go out and manually find the targets again (still subject to human error), or just use the potentially inaccurate data. Can we train a computer to overcome that inaccuracy and still be able to learn from that data?
The idea for this work derived from her experiences with hyperspectral imaging —imaging that visualizes light across the electromagnetic spectrum and beyond what the human eye can see. Often, she would have labels for objects in the scene they were working within, but having to draw things by hand and manually mark data points grew into inaccuracy. “And this problem kept coming up over and over again — if we want to do supervised learning, we have to have accurate labeling,” Zare says. “And we just don’t have accurate labeling in the imagery. So the idea for this just grew out of trying to address that gap. Most methods need really accurate mapping, and so we’re trying to fill that hole.”
Zare’s work in hyperspectral imaging analysis gave introduced her to the agricultural applications of engineering. Much of agricultural research uses hyperspectral imaging to gather knowledge about fields and crops. With her knowledge of machine learning, Zare is able to offer insight into how to best analyze that imagery.
She and Felix Fritschi, MU Professor of Plant Science, have used imaging analysis to study root structures — instead of having to pull the roots out of the ground, lay them on a sheet of paper, and manually measure them, they can now pull the roots out of the ground, hold them in front of a structure with cameras that take pictures from multiple angles, and then have the computer measure and analyze the roots.
Zare believes engineering needs to be a diverse, creative think tank in order to find the best solutions: “It’s extremely creative. You want to be able to look at problems in a lot of different ways and angles. I think people will look at engineering as a very dry, math, very structured, but in fact it’s extremely creative and needs a diverse set of opinions to solve these things.”
Looking back, Alina Zare has always been interested in problem solving. In the late 1980s, her parents brought home a computer, a giant contraption with a gray screen and a green monitor. Alina thought it was fun to work on — she would do simple programming, she would type in a question and the computer would spit out different responses. She started playing games on Nintendo, and often found herself thinking about the changes she wished she could make. She never actually tried to change the games, but still, she wondered: what if she could?
Relatives who were in computer science had some influence on Zare; they showed her how to do simple programming, such as asking her computer different questions. At school, she liked math and problem-solving. She gets excited about working with different fields, like physics, plant sciences or medicine. “The ability to influence a variety of problems is pretty cool.”
An early experience that Zare credits with shaping her thinking came from a homework assignment for a Spanish class in high school. She had to write a recipe, and she remembers taking it to the extreme — she gave instructions about getting the pot out of the cupboard, for example. “I guess while I was doing that assignment, I had to think sort of computationally — being able to lay out all the steps. But also within each sentence I was very methodically following the grammar rules that she’s given us because by no means was I fluent in Spanish.”
When she started school at the University of Florida — where she earned her bachelor’s, master’s and doctorate degrees — Zare hadn’t spent too much time learning how to program or becoming fluent in coding languages. “I think I could think logically and in a kind of structured way. … You can learn while you’re in college. You don’t have to come in knowing everything. But that was challenging because you’re surrounded by people who you feel are a lot farther along than you in learning how to program, but you can pick it up I quickly.”
When she was job hunting, she looked to MU because she felt it would be a place that would allow her to interact with lots of different disciplines — and it has been. Now, teaching helps Zare organize her thoughts about her research. Teaching is a way for her to collaborate with students — she can work with them to solve their research problems and her own. “(I can) organize my thoughts on the topic and get questions from students. If I’m presenting a particular material, I have to organize how I think about it, how I want to present. And students will ask questions that will throw me a curveball — maybe I didn’t think about it that way, but that makes sense.” Zare believes problem-solving is best from with unique perspectives. Her students’ perspectives force her to reevaluate her own ideas and develop better solutions. “I like to bring in topics I’ve been looking at as questions to the class — what do you think about it? How would you solve this problem? Sort of like a deep thinking kind of activity. … It’s not just what’s been done in the textbook, but what we could do next.”
“Engineering is not really a field to come up with algorithms just for the sake of coming up with algorithms,” she explains. “They need to be better than what has been done before, or tackle a problem that hasn’t been tackled before.” In order to find and communicate a problem, engineers must think creatively and tell the story of their work. Engineers are charged not only with finding solutions to problems, but also with laying out the life story of that problem. “It’s certainly not a dry subject, and you can get very passionate and creative about it.”