Learning @ Georgetown

Change font size: A A A

Collaboration with Undergraduate Enhances Dr. Mark Maloof's Research

Sometimes the best research partnerships start in odd ways. Take, for example, Mark Maloof, Associate Professor of Computer Science, and Georgetown undergrad Jeremy “Zico” Kolter. During his sophomore year, Kolter—a philosophy major—took an introduction to computer science class from Maloof to brush up his self-taught programming skills. Shortly after the start of the class, Kolter approached Dr. Maloof about the potential for additional work to provide the challenge he was seeking.

Maloof had an experiment that he never had time to finish for his dissertation. Kolter ran the experiment, analyzed the results, and thought of ways to improve the learning methods.

“Zico came back with ideas about a new algorithm, Dynamic Weighted Majority,” says Dr. Maloof, whose research work focuses on machine learning, a branch of artificial intelligence. “He was really smart and motivated, and I wanted to keep him engaged.”

Under the direction of Dr. Maloof, Kolter continued to do research to develop the Dynamic Weighted Majority algorithm, which could have real-world application in programs like spam filters, which learn from past behavior of similar messages whether an email is spam or not. Previous scholars had developed the idea of Weighted Majority, but this algorithm maintains a fixed number of so-called "experts," which could affect Weighted Majority's performance in non-stationary environments. Kolter and Maloof's algorithm addresses this issue by dynamically creating and removing experts, hence the addition of “Dynamic” to “Weighted Majority.”

Dr. Maloof describes Weighted Majority algorithms using a real estate analogy. Say you have a team of ten property advisors who each start out making $1 per consultation. Assume you are able to tell immediately afterward whether the advisor made a good recommendation or a bad one. If an advisor gives a good recommendation, the salary remains at $1, but if it is a bad recommendation, the salary is halved. After the first consultation, advisors’ salaries would be either $1 or $0.50. After the second consultation, salaries could be $1, $0.50, or $0.25. The number of expert advisors does not change, but each one’s salary will, depending on their answers to the series of questions.

The important change in the real estate analogy is that Dynamic Weighted Majority accommodates a changing number of experts. Maloof and Kolter’s algorithm begins with one expert. If the recommendation was good, nothing changed, but if it was bad, that expert’s salary would be halved and a new expert would be brought in at full salary. If an expert got below a certain salary, he or she would be fired. Over time, then, the best advisors are retained, while the least successful ones are removed.

After this early work with Maloof, Kolter was hooked. He wound up double majoring in computer science and philosophy and received a GUROP fellowship for his part in the project. He is now continuing his love for machine learning at Stanford’s Computer Science Ph.D. program.

“Initially, I worked very closely with Mark,” says Kolter. “I did not have the background to deeply understand everything I was doing, so working with him was crucial for developing Dynamic Weighted Majority.”

Kolter expanded his research skills by tackling his own project next, which, with Maloof as his mentor, became his thesis. Kolter came up with another algorithm, Additive Expert, which addresses some of the theoretical and practical holes of the Dynamic Weighted Majority algorithm. Kolter presented their findings at conferences worldwide. Maloof plans to work their research into the next machine learning class he teaches.

“Georgetown places a premium on undergraduate education, and that translates into putting effort into the classroom,” Maloof says. “I think students getting involved in a professor’s research is critical for getting into top graduate programs.”

Dr. Maloof also emphasizes the importance of funding availability for students like Kolter through GUROP. He says it’s gratifying as a professor to be able to engage students in his research projects, while providing them the necessary financial support so they can devote their time to conducting such front-line research.

“My favorite part about teaching is to look in students’ eyes and tell they’re excited about something they’ve discovered,” he says. “It’s very gratifying to see them go on.”

In addition to machine learning, Dr. Maloof teaches introductory computer science courses and an artificial intelligence class.

For more about the Computer Science Department’s cutting edge work, see our article on Dr. Brian Blake’s work on augmented reality devices in a previous issue.

Print Article

Related Stories