Skip to main content

Strategy for stronger statistics earns Po-Ling Loh an NSF CAREER Award

Written By: Sam Million-Weaver

The National Science Foundation granted Po-Ling Loh, an assistant professor of electrical and computer engineering at University of Wisconsin-Madison, a prestigious CAREER Award for her plans to improve statistical methods for modern computer science problems. The project could help self-driving cars avoid accidents or aid algorithms that make automatic diagnoses from medical images.

More and more technologies today rely on machine learning, or artificial brains known as neural networks, that parse out patterns from vast amounts of data in order to make decisions. When coping with messy information, classical statistical techniques sometimes come up short, which is why Loh plans to leverage an approach known as robust statistics.

“Deep learning is not robust,” says Loh. “Machine learning algorithms might make the wrong decision if they see some carefully chosen noise.”

The concept that an algorithm could be intentionally tricked by doctored data is called adversarial contamination—a growing concern as neural networks become commonplace. Security experts worry that in the future, for example, that a few strategically altered pixels in an image could make the steering system in an autonomous vehicle go haywire.

Traditional statistical methods aren’t equipped to deal with adversarial contamination. The classical theorems were developed to wrangle data sets that are, for the most part, well behaved.

“There’s a big difference between average conditions and worst-case scenarios,” says Loh.

Loh is developing new tools that could help neural networks outthink adversarial contamination. The methods will also be useful for scenarios where the data is inherently messy even without insidious outside influences, such as interpreting medical images. She’s begun working with the UW-Madison Department of Radiology to apply robust statistics to diagnostic algorithms.

“Radiologists are very interested in the practical applications of machine learning,” says Loh. “We’re very good at the theoretical aspects in the electrical and computer engineering department, so it’s a nice collaboration.”

Loh will also teach graduate students at UW-Madison how to harness robust statistical methods in a new course fall 2018. She also plans to educate the general public by giving talks on data science and statistics at high school math circles in Wisconsin and participating in public speaking engagements arranged through the UW-Madison Speakers Bureau.

The grant provides $400,000 of support over five years.