After a peripatetic academic route finally led him into physics and mathematics majors at the University of Southern California, Daniel Strouse found himself contemplating the brain.
Strouse, a Department of Energy Computational Science Graduate Fellowship recipient at Princeton University, was researching quantum information theory. He wanted to develop algorithms for fast, efficient computing systems that incorporate the strange laws of quantum mechanics.
“We were working very hard, designing these algorithms for computers that didn’t exist yet, but there was this incredible, efficient computer right in front of us – right inside all of us – that we didn’t really understand,” Strouse says. He studied the work of other researchers, including his present doctoral advisor, William Bialek, and found it’s possible to mathematically frame problems the brain must solve. “And by focusing on the problems that the brain has to solve, you could understand some of its organization.” That led him to think even more broadly about how the challenges biological systems face shape their designs.
Those same kinds of mathematical approaches, Strouse says, are applicable to the growing field of machine learning, which focuses on extracting knowledge and meaning from data. Our brains face similar issues: choosing from a flood of sensory information to draw meaning and make predictions. Strouse says it’s about “how do you learn, basically? And how do you teach a computer to learn?”
“Machine learning is exciting, because when you’re focusing very abstractly on how you analyze data, you can work on almost anything. It’s like you’re understanding a bunch of things at once.”
A major part of Strouse’s research considers how our brains decide which information from the environment is important. We’re almost constantly awash in so much sensory information we can’t possibly process or memorize it all. How do our brains decide what to focus on?
“Our basic hypothesis is that things that are predictive of the future are the important things,” Strouse says. By applying ideas from machine learning, he and his colleagues developed a mathematical formula that describes the approach. Using this theory, “we can make predictions about how brains respond to moving images.”
Strouse’s research could have other applications, such as data compression and clustering. More generally, its combination of math, theory and computational neuroscience is at the scientific forefront. “A lot of people are very excited about that intersection of neuroscience and machine learning because there’s been this long history of back and forth between the two fields.”
The difficulty the researchers face, however, is similar to the difficulty the brain faces: There’s a lot of information to sort through. Strouse and his colleagues must create a statistical description of the environment before they can identify and choose the specific features that have predictive power. “When you start doing that you realize, wow, the environment has so many variables, it’s so huge you can’t possibly measure all of that and model it in a computer,” he says. Machine learning and neuroscience research must face the obstacle of “figuring out how to break off a small chunk that you can focus on and actually solve.”
At USC, Strouse drifted from filmmaking to business and Chinese language studies to chemical engineering before he took a physics class. He realized mathematics and science were more than just applying formulae. “It was physics as a framework for understanding everyday things around us, and I was completely fascinated by this.” By the time his undergraduate career was over, Strouse had taken five more classes with the instructor, Associate Professor Paolo Zanardi, and worked with him on quantum information science research.
Strouse also spent a year in Great Britain, studying at the University of Cambridge’s Computational and Biological Learning Laboratory. It’s one of many countries he’s visited.
Strouse will seek a postdoctoral research post after graduation, but hasn’t decided whether to continue studying machine learning or theoretical neuroscience. Both are appealing, but he’s not sure which he wants to concentrate on for the short term.