Computer Sciences welcomes Assistant Professor Kirthi Kandasamy, focusing on machine learning

Kirthevasan (Kirthi) Kandasamy comes to UW-Madison from a postodoc at UC-Berkeley. He will research and teach about machine learning, which he says he happened on accidentally, and he’s never looked back! Kansdasamy works more specifically on using game theory to help competitors share data when they might not be inclined to do so but it is in their best interest. Read about this below – and watch a beautiful video from his home country of Sri Lanka!

Hometown: I am from Sri Lanka. I was born in Jaffna in the north of the island, and grew up in the capital Colombo. Nothing compares to life back home. Here is a nice video.

Educational/professional background: I was a postdoc at the University of California, Berkeley, working with Ion Stoica, Mike Jordan, and Joey Gonzalez. I completed my PhD in Machine Learning at Carnegie Mellon University, where I was co-advised by Jeff Schneider and Barnabas Poczos. Prior to CMU, I completed my B.Sc in Electronics & Telecommunications Engineering at the University of Moratuwa, Sri Lanka.

How did you get into your field of research? It was by sheer accident. When I was in college, I was looking up some material on regression and numerical methods. Somehow, I ended up watching a machine learning lecture series on Youtube. Never looked back since then.

What are your areas of focus? Broadly, I work on machine learning (ML). I am specifically interested in decision-making under uncertainty, game theory, economics, nonparametric methods, and computer systems.

What main issue do you address or problem do you seek to solve in your work?   ML systems are ubiquitous these days, and increasingly, they interact directly with people (or organizations). When building out such ML systems, it is easy to forget that people are rational, intelligent, and often selfish entities. Suppose an ML system is relying on input/feedback or input from a user and then using that information to make decisions that affect the user. Instead of providing that information truthfully, a user may be dishonest so as to manipulate the decisions taken by the system in their favor.

Such concerns are exacerbated when there are multiple users. We may wish to design ML systems which make fair decisions that benefit all users; however, some users might choose to manipulate outcomes to their advantage. Building effective ML algorithms in such situations requires leveraging ideas from game theory and economics to account for strategic and fairness considerations.

Please tell us about something you’re working on in layperson’s terms, so that non-computer scientists at UW-Madison and the general public can understand what you’re passionate about. 

One of my current problems on the above themes actually applies ideas from game theory to a use case in machine learning: specifically, on sharing data among competitors.

Since AI/ML has surged in popularity in the last decade, people nowadays realize the value of data. Moreover, organizations/people recognize that there is value in sharing data, even in competitive environments. For instance, banks are usually competitive, but they could share data on fraud detection—this is good for the banks and for society at large. This seems like a win-win situation. But organizations are understandably apprehensive about sharing data with competitors as they might be worried that competitors may not act in good faith, e.g. intentionally withhold data, or poison the data they are sharing.

We are designing algorithms which disincentivize such bad behavior. This requires quantifying the value one organization’s data provides to the others and then using that to determine how much of the others’ data it should receive in return. Each organization should feel that their best strategy is to share their data honestly.

What’s one thing you hope students who take a class with you will come away with? Learning to look at the big picture before delving into the details. This is an important skill when solving challenging problems. Students should learn to break down the big problem into smaller sub-problems and keep repeating until they have manageable problems that are easy to solve.

What attracted you to UW-Madison?  UW-Madison has very strong faculty in CS, EE and other engineering and scientific disciplines. The ML faculty here are also particularly strong. There are many natural collaborators and exciting opportunities.

I also really liked Madison as a city. It isn’t too busy. At the same time, it is very vibrant and has a lot to offer.

What was your first visit to campus like? I came here around mid-March for my interviews. The weather was deceptively nice for that time of the year (I didn’t need a jacket when I went out to get dinner on my first night!). I was able to walk around campus and the city, which was really nice. I also enjoyed the conversations with the faculty and the students, both technical and otherwise.

What are you looking forward to doing or experiencing in Madison? So many things to do! Visiting the Dane County Farmers Market, spending time at the Memorial Union Terrace, and walks along the lakes and the many parks in Madison, to name a few. I have also not biked in a while — Madison seems like the ideal city to resume biking. 

Do you feel your work relates in any way to the Wisconsin Idea? If so, please describe how. My work combines ideas from many fields, such as machine learning, statistics, and game theory. As such, the ideas can be fairly technical. At the same time, many of the problems are motivated by concrete use cases. Once we develop a mathematical solution, we often deploy it in some application to evaluate its real-world impact.

Hobbies/other interests: Outside of work, I enjoy cooking & baking, lifting, and reading.