When Householder talks about computer vision—training devices to recognize objects through the “eyes” of their lenses—he’s electric with enthusiasm. Discussing the basics of the concept, Householder picks up a water bottle and places it behind a stack of t-shirts, leaving it partly out of view. He describes how if someone can help computers' electronic eyes fill in what’s missing, they can train computers to recognize the partly hidden water bottle again on their own.
The evolving technology has ramifications across health care, defense, apps, and even self-driving cars. It isn’t easy to implement, but where some see a challenge, Householder sees an invitation. It fascinates him to transform the complexity of natural phenomena into mathematical terms: in this case, turning the capacities of human vision into formulas. With enough practice and development, computers can recognize anything from tumors to buildings—from orbit.
Like a persistent teacher working with a child, Householder taught the computer to see, and understand what it was seeing. One major step: teach the computer to recognize an optical illusion. It’s as tricky as it sounds.
“You know those Pac-Man-shaped circles with the corners cut out, and there’s a square in there?” Householder asks, referring to an illusion like the Kanizsa Square. In this famous example, four “Pac-Man” shapes arranged in four “corners” tricks the human brain into seeing a square that isn’t really there.
The phenomenon occurs in our brains automatically, but a computer needs humans to help it connect those dots, so to speak. As Householder and his mentor, Assistant Professor of Mathematics Fred Park, worked diligently for weeks, the computer would get better and better at seeing shapes for what they were—even ones that only exist in the abstract, like the Kanizsa Square. The student and professor are excited by their promising results, and have submitted them for publication.
The complex concepts and calculations of computer vision were new territory for Householder. But he was excited from day one to rise to the task. As technical challenges kept popping up, Householder got into the habit of slowing down and revisiting the relevant reference material to gain a deeper understanding of the problem at hand.
“There was a lot of learning on the spot as I was going through it,” he said. But with an undaunted desire to learn, Householder picked up and honed his new skills with Park as his guide. The computer wasn’t the only one with a knack for learning and a good teacher.
Park saw his student’s tremendous growth during the challenging mathematics project. Householder learned how to nimbly work through any difficulties with outside-of-the-box engineering. He ultimately cultivated the kind of critical thinking skills that are of “utmost importance in any field,” Park said. “Here, he is gaining these skills firsthand at an accelerated pace.”
Park, too, is fascinated with the possibilities of training computers to mimic human cognition.
“Lots of principles from cognitive psychology are mathematically modeled in computer vision. It is a nice mix of math, computer science, and engineering,” Park said. “It is extremely rewarding and fun to learn new concepts and find real-world applications for them. I feel that in this age of the digital revolution, computer vision advancements will only grow and it is exciting to be part of this growth.”
Now, Householder’s ready to tackle a new problem: coaching a computer to be nimble in its thinking, too, through “machine learning.” When people want to train computers to automatically run complex calculations, what often happens is that they’ll train it to specialize and get really good at a certain set of data. But when new data of the same type comes the computer’s way, it doesn’t adapt well, he said.
Essentially, computers aren’t always great at adapting to new problems. Good news for them: Householder is getting better at it every day.
