By Diana Lutz, news.wustl.edu
Kilian Q. Weinberger, PhD, assistant professor of computer science & engineering in the School of Engineering & Applied Science at Washington University in St. Louis, has won a prestigious Faculty Early Career Development Award (CAREER award) from the National Science Foundation (NSF).
The awards are given “in support of the early career-development activities of those teacher-scholars who most effectively integrate research and education within the context of the mission of their organization” with the goal of “building a firm foundation for a lifetime of integrated contributions to research and education.”
Eighteen CAREER awards are currently “active” at Washington University in St. Louis.
Weinberger will use the projected five-year, $440,000 award to perfect a type of machine learning that could be useful for a broad array of applications.
Weinberger’s CAREER project, “New Directions for Metric Learning,” seeks to solve one of the fundamental problems of machine learning: how to compare individual texts, images or sounds. If an algorithm could perfectly determine whether two instances of a data type are similar or dissimilar, most subsequent machine learning and data analysis tasks would become trivial, he says.
“A common similarity measure between two data instances is the total squared difference of their attributes,” Weinberger says. “With this metric, similar instances end up close together and dissimilar instances are far apart. Although this distance is a convenient and intuitive measure of similarity, it ignores the fact that the meaning of similarity is inherently task-and data-dependent.
“For example, one person might be interested in organizing articles by author, whereas a second might organize them by topic. Given the nature of their respective tasks, both should use very different metrics to measure document similarity.”
To deal with this difficulty, domain experts adjust their data representations by hand — but this is not a robust approach. It would be better if a software program could “learn” the metric (or data representation) that works best for each specific application, and this is the approach Weinberger plans to take.
“Such a metric can be learned,” Weinberger says, “by mapping the digital representation of the data into a high-dimensional representation, which is then deformed to move similar points closer together while keeping dissimilar data instances apart.
Back to News Directory