Today at Brown

Erik Sudderth says his research interests “sit at the boundary of the machine learning and computer vision communities,” which is partly what led him to computer science at Brown.

“I’m in machine learning, essentially a branch of statistics. My whole business is interacting with people who have a lot of data to look at and want to see what’s really going on,” he said. “Brown has a fair amount of collaboration between computer science and other departments, and it has a really great applied math department that’s interested in the intersection of applications and mathematics. That’s more where I am.”

Computer vision caught his interest because it poses a set of very difficult problems in machine learning. In “supervised learning,” for example, the computer is given a set of defined categories for objects it is supposed to find — much as certain cameras have “learned” to identify and highlight human faces in a snapshot. That’s interesting and useful, Sudderth said, but he wants the machine to do more.

“A lot of my work is reducing the amount of supervision that’s required,” he said. “Supposing I gave you a whole bunch of pictures that were taken at the beach. You don’t need me to give you categories or labels to figure them out. You look at them all together and make your own categories — sand, sky, water, people, birds. Can that be done in an automatic way by the machine?”

That would be an important step forward. The millions of unlabeled images on the net will always dwarf the ones that have proper meta-data, Sudderth said. “It’s easy to cluster data in various ways, but the trick is to have the machine come up with clusters that mean something, that are close to the things you are interested in identifying.”

Doing that with video may require more computing power, but the problem is easier. “One of the tricky parts in two-dimensional images is that objects are flattened and pasted in front of each other,” Sudderth said. “If you are moving around in a 3-D video scene, that problem is solved — and it was solved partly because people in the special effects industry cared about it. They needed to place synthetic objects into scenes.”

Sudderth began his studies in “the more mathematical part of electrical engineering” — signal processing, information theory, control and communications — at the University of California–San Diego and continued at MIT (Ph.D., 2006), where “electrical engineering and computer science are in the same monstrous department.” His interest in developing models that could provide an interface between large, complex data sets and the real world led him toward the computer science side of the department.

Computer science itself is changing, exploring areas of common interest with many disciplines. Machine learning, for example, has much to offer whenever people and the systems they use need to interact with complex, “messy” data. Places like Google and Facebook depend on that kind of expertise, Sudderth said. “The most successful companies in this Internet world are the ones that have strong machine-learning groups.”

When he begins teaching this fall, Sudderth will offer an introduction to machine learning aimed at advanced undergraduates and beginning graduate students — “a mathematical foundation and roadmap of the tools that are already out there for people who want to be users of machine learning. Computer science now has a fundamental role in other disciplines. One of the challenges is to bring more of that to the undergraduate curriculum.”