Finding the Needle in a Digital Haystack

Surveillance cameras. Smart thermostats. Baby monitors. Home automation systems. Self-driving cars. In recent years, tens of millions of such “edge devices” have been connected to the Internet. It’s projected that by 2020, there will be as many as 50 billion such devices making up the Internet of Things (IoT)—dwarfing the number of traditional computers and smart phones—and turning out an ever-expanding amount of data.

Most of that data? It’s normal and shouldn’t trigger a reaction. However, when anomalies occur—such as an abnormal heart rhythm sensed by a wearable EKG, a pedestrian in the way of a self-driving car, or an alarming financial transaction—they must be identified and addressed immediately. But with so much data out there, humans can’t possibly review it all. With that dilemma in mind, math major and computer science minor Elise Courtemanche ’18 and Professor of Mathematics and Computer Science Brian Turnquist spent summer 2017 on an Edgren Scholar project focused on getting computers to learn to define “normal” and then to detect anomalies. They used mathematical models and real-world data to determine ways to detect needle-in-the-haystack anomalies and report them in real time, using far less computation than industry standards.

“Think of parents,” explains Turnquist. “They can be sitting listening to a baby cry through a monitor, and they just know when something is different. They may not even be able to tell you what it is that’s different, but they know. People have a God-given ability to detect changes like that, to know when something’s up. It’s extremely difficult, but if we can replicate this human-like type of anomaly detection, then each internet-connected device can become ‘self-aware’ of its own normal and only report the anomalies to the cloud.”

Turnquist explains that in IoT, this machine learning—a subfield of artificial intelligence in which computers “learn” what’s normal or abnormal for a certain context—is increasingly important for the safe and efficient incorporation of autonomous technologies into our existing infrastructures. For instance, a hotel might have hundreds of surveillance cameras, Turnquist explains by way of example. “Nobody can possibly look at all that data, but every camera can ‘learn’ what’s normal for its corner. They all remain silent, looking at what they’re always looking at. But if something unusual begins to happen, they can sense that and then show the security team the most relevant feeds,” he says. Courtemanche explains the algorithm is based on recognizing a sudden increase of distance between vectors in a high-dimensional vector space.

In order to test the algorithm in a benign, easily accessible way, Turnquist and Courtemanche installed a webcam in Turnquist’s office window in the Clauson Center (CC). Facing Kresge Courtyard and the path to Nelson Hall, the camera picks up a feed of outdoor foot traffic. “Every single quadrant in the field has its own anomaly detector attached to it,” explains Turnquist. “One might only see blue sky, and they don’t know what the others are seeing. They each just sit there quietly as long as things are normal. But if something enters into that box that it hasn’t seen before, it lights up and says ‘Hey, look at this.’”

Turnquist explains the far-reaching implications of this work, something IT companies have had on their radar for years but that’s only now being developed in earnest. In healthcare, what if a specialist didn’t have to look at every radiology scan? We could do more scans without increasing the workload of radiologists. While there are many exciting implications of this work, there are also potentially worrisome ones. Streamlining data allows for an ever-expanding IoT, which could increase life expectancy and crime-fighting capabilities but also break down personal privacy. “These advancements aren’t inherently positive or negative. They just are. We have to decide whether the implications are exciting or frightening—or both,” says Turnquist. “But this is the beginning of something that—10 years from now—could be a huge part of our lives.”

Courtemanche says the potential of working on research like this—right alongside established faculty members—was something that drew her to Bethel over the larger state school she was considering. “It was a chat with Dr. Turnquist, on a tour of Bethel, that really changed my plan. To be able to have exposure to this kind of research at the undergraduate level has been a game-changer.” This spring, she’s working with Turnquist on getting this project published, and then she’s turning her sights on her post-Bethel plans: a Best Semester program at Oxford, studying logic. “It’s closely related to math. I love the formal logic and philosophy of mathematics. I see this as a great way of broadening my skills on a subject I already love. Then it’s grad school, maybe out east. Wherever the research is interesting!”

Courtemanche’s future in mathematics is bright, but for her, the journey has been as rewarding as the progress made toward her goals. On this particular project, her skills and Turnquist’s were complementary, but they shared a deep excitement over their findings. Courtemanche remembers one particularly exciting breakthrough moment when she and Turnquist stopped and said to one another, “I love math. Math is so cool!”

Read about the rest of the Edgren Scholars who did research in 2017, or find out about math and computer science programs at Bethel.