The world of science and technology is currently being reshaped by machine learning. At the heart of machine learning are algorithms whose aim is to mimic the complex processes of human learning, but on a near-unimaginable scale.
Although computers already are learning our preferences, enhancing our online experiences, and allowing for seamless automation to an unimaginable extent, for assistant professor Dimitris Papailiopoulos, who joined the electrical and computer engineering faculty in fall 2016, it is about how we can enable computers to learn concepts even faster.
In several modern applications, it can take up to several weeks for a computer to “learn” a certain concept. For instance, in driverless vehicles, the machine needs time to learn the difference between humans and cars, and intersections and cul de sacs—intuiting the same automatic connections that any human needs to make while driving. However, if researchers like Papailiopoulos can streamline this computational learning process, these novel machines will become much more efficient, accurate, and easier to interact with.
“In contemporary applications, learning algorithms are required to go through millions of examples to learn a concept—this is called the training phase,” he says. “For several modern machine-learning pipelines, this is a major computational bottleneck. The goal of my research is to significantly speed up this training phase.”
The current focus of his research is to reduce the typical duration of training from approximately a week, to only a day or two. Essentially, he’s trying to get computers to learn concepts that are in some cases as complex as those learned by humans, but at an extremely larger scale and rate.
Papailiopoulos, who grew up in Greece, developed an early interest in mathematics and computers, so electrical and computer engineering offered a perfect fit. He received his engineering diploma and master’s degree from the Technical University of Crete in Chania, Greece. He spent seven years there before moving to the United States, where he attended USC as a graduate student. He then followed his advisor to the University of Texas at Austin, where he received his PhD. After his PhD, he spent two years as a postdoctoral researcher at the University of California at Berkeley, where he was a member of the AMPLab.
During his PhD, Papailiopoulos mainly worked on large-scale data processing and data storage problems, which has led to several theoretical and practical innovations in the field of coding theory.
In many ways, his work with coding theory and data storage has been the foundation of his research. He started working on machine-learning problems a little more than three years ago, and is excited about the potential impact of developing and implementing new ideas in the field.
An academic at heart, Papailiopoulos moved to UW-Madison because of its strong research environment.
“I chose UW-Madison because I needed a place that would give me the academic and intellectual freedom to work on the things that I wanted to pursue,” he says. “I was also impressed by how excited people here were about my work and what I have to bring to the table.”
Like many new faculty members, he was also impressed by the highly collaborative and open environment on campus. “I think it’s one of the things that makes UW-Madison special, and isn’t necessarily true for many places; there’s a lot of opportunities for interdisciplinary and cross-disciplinary research here,” he says. “We have, for example, the Wisconsin Institute for Discovery, that brings together faculty from many disciplines like computer science, biology, electrical and computer engineering, statistics, with all sorts of backgrounds—this offers a fantastic space for researchers to collaborate with one another. You’ll find people from the social sciences and engineering working on a joint project, and I think that’s a very beautiful thing that you don’t find in many places.”
Papailiopoulos was hired through the college’s Grainger Institute for Engineering and is affiliated with the Wisconsin Institute for Discovery (WID) on campus. He’s part of the optimization group at the WID, which allows him to interact with people who use the same mathematical tools that he does, but often for entirely different applications.
“The optimization group offers a lot of intellectual diversity. You are exposed to concepts that may have been applied to, say, industrial and systems engineering, and you can get ideas about how you can apply those same concepts to some of the computational work in the field of machine learning,” Papailiopoulos says. “There are also all sorts of events organized by the group, which is another way to get your ideas out there, and get feedback from people of various scientific backgrounds.”
He is currently teaching a special topics class, ECE 901: Large-scale Machine Learning and Optimization. The class delves into algorithmic challenges associated with machine learning, specifically when applied to large data sets, which is very close to Papailiopoulos’s research interests.
“Machine-learning is a cutting-edge field right now, so there’s a lot of fascinating open problems,” he says. “I enjoy getting to share them with the students, and hearing interesting feedback and ideas on how we can resolve some of the existing algorithmic challenges.”
Madison has become increasingly high-tech, rated No. 5 in the country for high-tech jobs. For Papailiopoulos, Madison is a “hidden gem,” integrating both Midwestern charm with a growing tech scene.
“There’s a lot of tech industry here, and lots of young professionals inside and outside of the university, which creates a very lively and highly intellectual environment,” he says. “There are exciting things happening in Madison.”
Author: Lexy Brodt