Many academics choose to study one thing and one thing only. Eric Mazumdar is not one of those academics. As an assistant professor of computing and mathematical sciences and economics at Caltech, Mazumdar uses tools and ideas from economics to understand emerging problems in machine learning. Some of those problems have grown immensely more complicated with the development of algorithms that steer our choices about the clothes we wear, the food we eat, the movies we watch, the information we digest online, or even whether to grant someone bail or give them a home loan. But algorithms are a double-edged sword; for all the complexity they add to our lives, their behaviors can be easier to model and understand than human behaviors.
We recently spoke with Mazumdar, who came to Caltech in 2021 after receiving his doctorate in electrical engineering and computer science from UC Berkeley that same year.
You study economics. You study computer science. You got an undergraduate degree in electrical engineering. How did you transition from electrical engineering to economics? Is there a through line there?
My journey was actually a little bit more complicated. When I started on my bachelor's, I was actually interested in bioengineering for a while, and then I realized more and more that the part I liked most about the research I was doing was the math and the modeling of the dynamics of biological systems. That took me from bioengineering to electrical engineering and computer science.
Then I started my PhD working in a field known as control and dynamical systems. It's a relatively old field, but, at the time, I was thinking that the field's tools hadn't been applied yet to a really pressing area—this emerging area of machine learning in society, where the dynamics of the algorithms have real impacts on people's lives. The decisions that these machine learning algorithms make can have consequential impacts, but the dynamics are really complex because you have algorithms interacting with people.
To me, the through line is thinking about the underlying dynamics in all of these systems.
Can you define dynamics for me? Is there a field-specific definition of dynamics?
Dynamics are really just any process that evolves over time. A lot of our day-to-day interaction with algorithms happens in dynamic ways and at different scales. For example, on a small scale, you have individuals interacting with Netflix or TikTok, and their preferences change, and content changes over time. But then you can have dynamics on larger scales, like traffic dynamics. Anyone who commutes knows that traffic patterns change and evolve throughout the day. In the morning, traffic tends to move toward urban areas and in the afternoon, traffic moves toward residential areas. That's a time-varying process. Dynamics are just any time-varying process.
I would imagine that these kinds of dynamics can get rather complicated when you talk about something like Netflix, because it's not just people's preferences. The algorithm is also adapting to their preferences, so maybe someone starts out with a preference but then the algorithm learns from them and it feeds back to them, and it might change their preference.
You use exactly the right word there. They get complicated because of these feedback loops between algorithms and people. That interaction, as it plays out over time, can get extremely complicated and give rise to behaviors that we wouldn't want to happen or things that are surprising.
Examples of this are pretty common. One of the ones that we were looking at recently was what happens if multiple firms are trying to sell the same product on Amazon, and they're all trying to set their prices. If they use the wrong pricing algorithms, the fact that they're interacting with each other can cause prices to go crazy and spike unnaturally. It shows that the feedback loops in these systems are super important and are not very well understood yet.
What is it about economics that interests you?
A large class of algorithms is being deployed into everyday life for decision making that has an impact on people's lives. If these algorithms are making consequential decisions for people, we have to take into account people's objectives and people's preferences and how people reason. Those are ideas and tools that I think economics has historically dealt with. Instead of reinventing the wheel, I think that there's a huge amount of benefit we can get from integrating ideas in economics into algorithm design and machine learning. That's what I work on.
Machine learning and artificial intelligence (AI) have been in the news a lot lately. Is this a challenging field to be in because of how quickly things are changing?
For sure. On the one hand, things have changed very fast even since I started grad school, which was 2015; the recent progress in deep learning and large language models like GPT-3 and 4, in particular, has been really impressive. However, on the other hand, I think that things have moved so fast that a lot of important questions have gone unaddressed. I think there's a huge amount of work to be done at the interface of economics and machine learning in terms of understanding the fairness, bias, and manipulability of algorithmic decision making. As an example, in recent work, we trained a simple algorithm on top of a large language model (like GPT) to predict skills and jobs from resumes, and we found that by adding a simple signal in only 0.1 percent of the training data, we could consistently get a resume scored highly for skills that it actually wasn't associated with. The fact that these models, which can be so impressive, are also so easy to manipulate is really telling.
An overly simplistic way of thinking about why we've missed out on understanding these problems is that machine learning has progressed extremely rapidly in tasks that are similar to pattern recognition, but in decision making, I think it's moved a little bit slower. To me, that's because decision making, fundamentally, has to deal with uncertainty, planning through dynamics, and the behaviors of other agents. Those are problems that people are still thinking about in economics and that we don't have huge amounts of data for. Adding the complexity of algorithms doesn't really make the problem any easier.
Economics is an old field, but machine learning is relatively new. What does it add to the mix when you bring these machine learning tools into economics?
I think we understand machine learning algorithms a lot better than we do people, because, from a certain standpoint, we can control what algorithms do when we use them in game-theoretic settings. We have a lot more control over the end behaviors that can happen. That tends to actually simplify some of the problems in the end.
How do these algorithms simplify economics problems for you? Are they able to recognize patterns that are difficult for humans to recognize? Is it that they are able to work through a big pile of data more easily?
They make it easier because we have more math to describe them. There are a lot of mathematical models for how people make decisions in economics, behavioral economics, and mathematical psychology, but if we're coding an algorithm, we know exactly what's happening. That kind of certainty allows us to simplify and understand the dynamics we study.
What brought you to Caltech?
One of the things about Caltech that's exciting to me is how interdisciplinary it is. I'm jointly appointed in the Division of the Humanities and Social Sciences and CMS [the Computing + Mathematical Sciences department], but I also have an affiliation with the Control and Dynamical Systems program. The opportunity to work across disciplines and talk to people in different areas is a very rare thing in academia, and I think that's something that Caltech does extremely well because of the small size. I can interact with the economics faculty, with the experimental economists, with the control folks, with the robotics folks, or with people who work on networks.
As an example, we're doing a research project right now with one of my students that is exploring how the use of algorithms in social media impacts polarization in society. We're trying to model the interaction of an algorithm over a large population of people. Modeling large populations of agents is something that's commonly done using partial differential equations (PDEs), and Caltech has a really strong core of theorists in applied math who work on understanding PDEs. Just by talking to them, we have gotten several ideas on how to model these populations of people when they interact with learning algorithms. We used that to come up with a model that tells us when and how polarization can emerge.
OK, my last question: Where is your favorite place to travel to, and why?
My mom is from France and my father is from India, so I've had to go to India and France quite a lot, but the one trip that has stayed with me was a trip to Egypt when I was pretty young. We were able to start from the bottom of the Nile and go all the way to the top, and see all of the sites. It was just incredible to see all of these historical sites that had been there for millennia. That's always been one of my favorite trips. It's such a different culture and history than the ones that I'm used to in France, India, and the U.S.