Interview with Professor Eric Mazumdar

Interviewer: Hannah X. Chen

Could you start by sharing a brief overview of your research and its applications?

I work at the intersection of machine learning and economics, and the motivation is really thinking about how we use machine learning algorithms in the real world, and what happens when we actually use them. In the real world, algorithms have to interact not only with people, but potentially also with other algorithms, and they have to deal with situations that we’ve historically seen machine learning algorithms struggle with. If you think back to applications of machine learning in the real world, there’s all these famous failures. There’s self-driving cars that crash because they can’t tell the sky from a truck, or that won’t cross an intersection because they can’t figure out the intent of other agents that might be present. There’s also all types of problems that happen in e-commerce markets, in social networks, and so on. And you can really think of these issues as arising because learning algorithms are coming into contact with people and other algorithms.

A big part of my research is just trying to understand what can go wrong when we use algorithms in environments containing real people who might have their own objectives. We’re also thinking about how we would design algorithms specifically for these environments so that they can have provable guarantees or at least economically meaningful measures of performance. My main application areas are in transport systems (ride-sharing platforms are a major example of this), e-commerce platforms, social networks, the delivery of healthcare — all of which have this kind of flavor of algorithms meeting people and interacting with other algorithms.

What do you see as the short-term or long-term directions of your work and the field as a whole? What is the end goal you would eventually like to reach?

The kind of pipe dream is this idea that we can develop algorithms that actually have a beneficial impact in society. And what I mean by that is having algorithms that don’t make things worse when they’re actually deployed in the environment — instead, they actively improve things. Maybe if it’s in transport systems, it improves congestion in cities. Maybe if it’s on e-commerce platforms, it lowers prices and makes the market more equitable. But doing this while keeping in mind issues of fairness and biases in data is really hard. But I think it’s a really important problem, because people aren’t going to stop using algorithms in the real world, and we really need to start figuring out how we do this properly, fairly, and ethically. That’s a big part of what my research is trying to do.

Is there anything you’re really excited about in your work right now, such as a current project or new idea?

Theoretically, we were able to show that one of the most important aspects when you’re using learning in the real world is how fast you’re learning. Often, we think that learning fast is always better, but when we started applying these types of ideas in games — where you’re learning against an opponent who’s also learning — we found it might not be better. And actually, the speed at which you learn changes the equilibria you might find, but different equilibria have different properties. Some might be more socially optimal, some might be better for individuals, and so on. So all of a sudden, we now have this way of achieving different solutions by changing how fast we learn.

Something I’m really excited about is that this is a phenomenon we’re seeing in real data from pricing on e-commerce platforms, for example. We see that people change how fast they update the prices relative to their competitors to get an advantage. Sometimes people update their prices very very slowly, and then they’re actually better off and are able to get more of the market. In other cases, they update much more quickly and can be better off as well. So there’s this interesting degree of freedom, which I don’t think has been explored much. I’m very interested in exploring the dynamics of learning algorithms in these game theoretic settings, which I think will be a really exciting area in the next couple of years. 

What are the biggest challenges that come with performing this research? What are the things you have to do to get your work moving forward?

A lot of my work is more on the theory side, mostly trying to develop algorithms that satisfy some theoretical guarantees and seeing that they work in practice. There’s two main challenges: on the data side, a lot of companies and entities that use algorithms in the real world don’t want to share their data because they don’t want to show that they’re doing something that isn’t working, or they’re scared of revealing some sort of inherent bias in their data. On the theoretical side, you have to bring in ideas from a lot of different fields to try and understand these problems. You need ideas from game theory and economics to deal with these issues of multiple agents all interacting, but then you need ideas from control and dynamical systems to think about the dynamics that are arising in these situations. And you need to bring in ideas from computer science to deal with the algorithmic parts. It’s really a research agenda that’s at this intersection of three different domains, so there are natural challenges regarding how they interact and whether you can make these things play nicely. It’s a really cross-disciplinary research agenda, and that comes with challenges, but also a lot of opportunities.

What has it been like working in both the CMS and economics departments at Caltech?

I only recently started, but it’s been a very satisfying experience because from what I’ve seen, economics and the computing sciences are very tightly integrated at Caltech. The economists are familiar with a lot of the tools and ideas from computer science and engineering, and the people in engineering also have some connections with economics, so it’s been really interesting to work at this interaction and play both sides. In my view, Caltech is extremely well positioned in this new research area that’s emerging at the intersection of machine learning and economics. We have a very strong mathematical microeconomics core, along with very strong experimental economists that understand decision making, whose ideas we can then embed into our algorithms. On the other hand, it has a very strong computer science and mathematical sciences department. So it’s been a really fun, unique opportunity, and part of that is only possible because of the small size of Caltech, which makes it easy to work between divisions and departments.

What drew you to this area to begin with? How did you get started in either machine learning or economics?

The way I got into computer science, machine learning, and economics is a bit weird. I basically started by getting my hands on data. We were looking at data about how people park their cars in Seattle, and there’s this idea that a large part of the congestion in Seattle is caused by people just looking for parking. So the question was whether we could come up with a model for this and see if this is true. Something very natural that emerged from this problem is to model parking as a game between people: those trying to drive on the city streets and those looking for parking spots. We started doing that, and that got me thinking a little bit more about how we can even use learning algorithms in game theoretic settings. From there, I kind of moved more and more towards machine learning and game theory, and then economics more broadly. So it was not an immediate process — I didn’t go into grad school knowing that this was what I wanted to do, but it was a very natural thing because this research area allowed me to draw on ideas from game theory, which I thought was fascinating, but also apply tools from machine learning, dynamical systems, and control, all in the context of interesting, real-world problems.

What were your mentors like as you were working in multiple fields? What has your experience been as a mentor to students of your own?

What may be surprising is that in my undergrad, I was actually doing computational biology. I had a set of really good mentors who allowed me to explore my interests in that field and help me discover that the thing that excited me most in computational biology was actually mathematical modeling and coming up with dynamic models of how, for example, drugs propagate through biological systems. When I got to Berkeley for grad school, I again had mentors who could really get me thinking about big open questions and the emerging themes and questions that were arising as machine learning and algorithms are deployed in the real world. So I’ve been extremely lucky throughout.

Here at Caltech, there’s two main parts of mentorship for me. The first is giving people the freedom to be creative and come up with their own ideas while giving them guidance through high-level research goals and things like this. The second thing is mental health. I’ve had a lot of close friends struggle with mental health in grad school and undergrad, so for me, a big part of mentorship is making sure the person feels very supported on the day-to-day and has something to fall back on. A lot of these environments, grad school and even undergrad at such an academically rigorous place, can be very high-pressure, so you need to have strong support networks for students.

What are some of your favorite parts about being a researcher and professor?

The best part of doing research is being able to talk to people and all of the random conversations you have in hallways or in front of white boards. It’s also just interacting with students and people with such different backgrounds and interests. That’s kind of the standard answer, but it’s really the most interesting part for me, especially at Caltech where you can meet so many people who do so many different things. You can go from a conversation about fluid mechanics to game theory and auctions in the same hour. It’s all very exciting.

Do you have any advice for students who want to get into research, especially within your field? 

As I mentioned, I was in bioengineering when I started undergrad, and I ended up in EECS (electrical engineering and computer science); now, I’ve moved towards economics. I think being multidisciplinary and having varied interests is a good thing. Interesting things happen at the intersection of fields, so don’t be scared of trying to make connections between fields. 

Getting involved at Caltech, I think, is really a question of persistence. If professors aren’t answering (and I’m definitely guilty of this, too), it’s not because they don’t want to — it’s just that they get too many emails. We want to work with students at Caltech since you’re all so good, so it’s very important to not be afraid to just send the emails and knock on doors. I was definitely worried when I sent my first few emails looking for positions, but in the end it’s just biting the bullet and doing it.

Interview with Professor Frances Arnold

Interviewer: Maggie Sui

Biography

Professor Frances Arnold is the Linus Pauling Professor of Chemical Engineering, Bioengineering and Biochemistry at Caltech and a 2018 Nobel Laureate in Chemistry. Her research focuses on using directed evolution to create new enzyme function. Prof. Arnold received a B.S. in Mechanical and Aerospace Engineering from Princeton University and received a Ph.D. in Chemical Engineering from the University of Berkeley.

Trivia

Favorite book:  Too many to choose from.  I am now enjoying Ai Wei Wei’s memoir

Favorite Place:  Planet Earth

Favorite Food:  oysters

Favorite Protein molecule:  cytochrome P450 BM3

What type of research does your lab focus on?

We breed molecules like you can breed cats and dogs, selecting for the traits that are useful to us in a process that mimics evolution by artificial selection.  The ‘directed evolution’ methods we pioneered to engineer better enzymes are now used all over the world for everything from laundry detergents, cosmetics, treating disease to making jet fuel from renewable resources.  I envision a world where nature’s clean and efficient chemistry replaces dirty human manufacturing processes.  Wouldn’t it be wonderful to make our pharmaceuticals, cosmetics, fuels and chemicals like we make beer? Wouldn’t it be wonderful to have microbes eat our plastic wastes and convert them to valuable new materials?  That’s all possible with enzymes.  But someone has to make them do what humans want rather than what they do in nature. That’s what we do.

What aspects of research do you most enjoy?

I love discovering new things that enzymes can do.  Sharing the joy of discovery with others is almost as much fun. And since my students are always discovering new things, it’s a blast to go to work.

What influenced you to pursue science? Did you have any role models growing up?

My father was a nuclear physicist and also did a lot of engineering design. I wanted to be like him when I grew up, but I didn’t love nuclear physics. I tried a few things before I found protein engineering, at the beginning of the DNA revolution. 

What was the origin for your directed evolution idea? When did you first think of it?

When no existing design approaches worked to make better enzymes, I was getting a bit desperate. I decided to try lots of mutations, easily done randomly using error-prone polymerase chain reaction (PCR) and testing by screening in plates. I was thrilled when we found beneficial mutations in surprising places, where no one would have predicted. We accumulated those mutations in sequential rounds, and sometimes recombined them, just like breeding cats or dogs over multiple generations. That worked beautifully, and evolution became the only reliable method for generating better enzymes.  It was an instant ‘hit’ with people in industry, but took much longer to gain the respect of scientists, who always want to understand why. To understand why, however, it’s much nicer to start with the correct answer and work backwards. Directed evolution gave the correct answer.

How did you become the Nobel-prize winning scientist you are today? 

I don’t really know for sure, but it probably has something to do with Caltech. I came here naïve and with limited training, but a good idea.  I was able to implement my ideas here, and formulate even better ones with the help of critical colleagues and brilliant students and postdocs. I never stopped to worry whether it would be possible to make better enzymes or whether I would be able to do it. Nor did my students.

What is your work in Washington like?

Nobel Laureates are always warned not to get involved in things we know nothing about.  I threw that advice to the wind when I took on co-chairing the President’s Council of Advisors on Science and Technology (PCAST) in January, 2021.  I find it fascinating.  And I have already learned that science is easy compared to dealing with…people.  People are really complicated. We can have all the right science, but if people don’t trust it or won’t use it for other reasons, it goes nowhere. In this job, I am also seeing many fascinating new science problems and solutions.

I see this work (which is unpaid and takes a lot of time, by the way) as a way to pay back, to make sure that future scientists will have the same wonderful opportunities that I enjoyed.

What are some of the challenges you’ve recently faced? How do you overcome them?

We all face challenges: loss of loved ones, illness, failure.  If you have lived as long as I have, you will experience all of those, perhaps multiple times.  I find that the best way to face life is to be grateful for what I have, which is so much. I try not to dwell on what I have lost, nor do I feel sorry for myself. I am grateful to still be healthy and able to do something useful.

What are your interests outside of science?

I love music, gardening, hiking, meals and conversation with friends. I used to love travel, but I haven’t been many places lately…

What advice would you give to Caltech students and other young students interested in research?

Do it because you love it.  All else will follow.

Interview with Professor Aaron Ames

Interviewer: Sherry Wang

Can you give me an overview of your research?

Broadly speaking, what I do is study how robots walk. That’s kind of the core disciple we consider in my lab; that is, specifically achieving walking behaviors on robots that are as dynamic and dexterous and fluid as humans walking. So this actually involves quite a bit of math to try to understand and characterize what walking is, and once you can do that, you can apply it to a lot of application domains, especially domains where humans walk, or humans walking is part of it. What I mean is, robotic assisted devices, prosthetics, exoskeletons, etcetera. So I have a wide spread of robotic applications, but the real focus is dynamic locomotion.

How did you get involved in robotics, and in particular dynamic walking and dynamic motion?

It’s an interesting story that is probably not wholly unique from a Caltech perspective. I was introduced to science fiction as an undergrad, so later in life than most people. I started reading Asimov and all these other authors, and I just got totally hooked. I was just fascinated with this idea of science fiction in general, but a big theme of a lot of Asimov’s books were robotics, and humanoid robots, and all the robots playing a part in our daily lives, and thinking about what that would really mean if we could really realize those behaviors. In undergrad, I did mechanical engineering and math, and I wanted to go to grad school to really understand this better. I think what appealed to me the most was just this deceptively simple thing that we do. Humans walk around all the time, they make dynamic movements all the time, and it seems like something we should be able to understand. Yet under the surface it’s such a complicated problem I find it intriguing. It’s one of those cases where you actually know a solution exists. Feasibility is guaranteed, because humans do it all the time. You just have to be smart enough to try to find that. I was purely theoretical in grad school. I didn’t touch a robot. I didn’t touch an experiment, I didn’t build anything, I just proved theorems and theorems and theorems. I think that made me a better person, but because my passion was robots–it was something I always really loved–I ended up coming back to it as a professor. It was a couple years into being a professor, and I had been doing some walking stuff and some simulation and everything, and I realized no one listens to me, and everyone probably thinks I’m crazy because the math I was doing was really out in left field. The only way I can actually convince people is if I get stuff to work. If I get physical hardware to do awesome stuff, then maybe they’ll listen to the math. I started doing that with simple robots, and making that one walk, and then I started getting addicted and making more and more robots; I worked with NASA on a humanoid. So I started getting really into the robotics side, while at the same time not losing the math that got me excited about this stuff and keeps me going to this day. Intrinsically, it was science fiction that led to studying a whole bunch of math, that led to me actually making robots to do stuff to demonstrate the math.

What are some difficulties and successes that you’ve seen in your research?

Difficulties always stem from the fact that the real world is hard. I always say to my students, and it’s a pretty silly quote but descriptive nonetheless, which is “hardware is hard.” There’s a “hard” in the “hardware” for a reason. The reality is that however smart you think you are–however clever you think you are–you go to put something on a robot and Mother Nature lets you know very quickly that you’re stupid, and you don’t know anything, and the world’s really hard. And you constantly are reminded of this. It’s not so much a difficulty; I view it almost as an opportunity, but it’s humbling. You’re constantly humbled, and what it takes to make anything work is so incredibly complicated. It’s exactly this difficulty that gives you opportunity, and what’s exciting and leads to your success story. Especially when you try to do what’s unique to my lab, (which is not to say that there’s no other labs that do it, but it’s a fairly unique thing) which is trying to do real theory–theorems and proofs and mathematically justifiable things–and then put that on hardware. So the goal is not so much to make the hardware do something, but the goal is to understand something at a deep mathematical level, and then make the hardware demonstrate that understanding. Recently, we worked with a French startup company called WonderCraft on their powered lower-body exoskeleton. And this is specifically aimed at paraplegics. There has never before this time been a powered lower-body exoskeleton that allowed paraplegics to walk dynamically. That is, walk without crutches, move their hands freely, and have a dynamic gait. And they [WonderCraft] took math that was developed by me–my lab, with collaborators and everything–but they took theorems that were in papers that we wrote, and they put it on the exoskeleton, and they actually got paraplegics to walk for the first time ever with a powered assisted device. So that was another one of these moments. You’re writing the papers, you’re proving the theorems–in this case, it was a paper that appeared in 2014–you would never think that in four years that that math would be helping a paraplegic to walk. Those are the moments that everything comes together, and it all looks perfect. And then you realize, “time to go back and try to do something new.” Now we want to have the next behavior, and nature starts smacking you down again. Then nature starts reminding you that you have a lot to learn. The success is great, but there’s always problems out there–which I find exciting. There’s never a lack of problems to solve.

What behaviors are you investigating now?

On the legged robots side, we really want to get robots out in the real world more and more. We want to have them do things that are impressive and dynamic, but in a really unstructured environments. Examples are, we want to take a robot on a hike. We want it to be able to hike one of the trails up in the mountains near Caltech. To actually be able to handle all these different terrain types autonomously, rather than just a pre-prescribed [route]. And not just walking; we’d love to do running and jumping. We’ve made a robot jump in our lab, which was amazing, but we’d like to get dynamic behaviors outside, in the real world. So on the walking robot side, that’s what it’s all about to us. It’s handling more and more complex terrain in a robust and dynamic way.

And in the same light, translating those behaviors that we’ve achieving on walking robots to robotic assisted devices. We actually have one of these exoskeletons by WonderCraft in our lab at Caltech now. We’re putting our new, more advanced stuff on it and testing it out, and we’d love to get that device walking outside, and walking on sand and dirt. Right now it’s walked dynamically, but in clinical settings. Everything we’re doing is pushing outside the lab. Start to handle the complexity of the real world. And that’s where a lot of interesting theoretical problems come from too. How do you quantify that uncertainty? How do you potentially bring in machine learning to make sense of that uncertainty, but use it in the context of the models and the theories that governs the motions of the systems. That’s where we’re pushing: trying to get robots into the real world, and have them try to do fun things.

Do you see the field progressing toward the “singularity” that science fiction authors write about?

I’d love to think that we’re that close, but no, we’re way off from any sort of “singularity”. We can’t even make robots do simple stuff still. If you think about the million years of evolution that it took for humans to do the things that we do, we haven’t recreated a fraction of it. We’ve come an incredibly long way. When I started doing experimental robotics, I had really simple robots and just making them walk dynamically was a big deal. This was before Boston Dynamics days. At the time, Boston Dynamics had quadrupeds that were cool, but no walking robots–no bipeds–and that was only like eight years ago, so it wasn’t that long ago. And now, we have humanoid walking robots. I take our biped from our lab outside and have it walk on grass and dirt, completely untethered, 3D walking dynamically. We’ve made a huge progression towards getting robots out of the lab and into the “real world”, but it’s still very scripted. When we have our walking robots walk on grass or dirt, we get it in various specific scenarios, and we tune the robot right, and we can’t have it go walk some mountain trail, or go walk downtown Pasadena and pick me up a cup of coffee and bring it back to me. These higher level behaviors and reasoning are an enigma to this day. And in some ways, the world of robotics is also bifurcated at this point. There’s a whole wealth of people that are working on getting robots to do dynamic things, and then there’s a whole wealth of people working on machine learning and AI, and those are different people. The things that [machine] learning is good at are things that can’t, right now, be translated to actual robots. And conversely, the things that we can make robots do doesn’t currently doesn’t inform machine learning. I think that’s starting to change, and people are looking for that intersection. [However,] to talk about a singularity is to talk about a robot that can do everything from walk effortlessly, to actually reason about its environment, much less become conscious. We’re just not even close. Right now, if I want to lose my robot, I’ll walk into a volleyball court and it will fall on its face, I guarantee–any robot will. Soft sand, or go in some snow, and it’s over. Until we can do those basic things that animals do so easily, much less humans, we’re a long ways off from achieving a “singularity”.

What are some of your interests outside of science?

I still love science fiction to this day. Books, movies, all of it, I find it really exciting. There’s nothing like a good, corny science fiction movie when you’re burned out mentally, or a good sci-fi book to rekindle that excitement in science. Little things just around LA. We try to take hikes–I have a dog, so I love going on walks with my dog and my wife, and going on hikes around Pasadena.

Do you have any career advice for students interested in robotics?

It’s an exciting area with a lot of potential. It’s interesting because the robotics push has been going on for a while now. It was a “dead” area for a while–very small–and it just bloomed research-wise in the last decade. And it doesn’t look like it’s slowing. We’re seeing more and more robots going into the real world. One of my students is the CTO of a startup that’s doing cooking robots–Miso Robotics–and he’s getting these robots into Dodger Stadium and cooking burgers. So you’re starting to see robots in the real world, so it’s an exciting time to be in the domain. My advice for students that are interested is to, at the undergraduate level, get involved with research labs. There’s no replacement for getting your hands on hardware and starting to play with these things early on in your undergraduate studies. Go to grad school because that’ll give you a depth in robotics that’s hard to emulate. Get involved in research because that’s where the exciting stuff is still happening. That’s not to say that industry doesn’t have exciting things, but the next generation [is] being developed [in research labs]. Get involved early, stay involved, and be part of some of these fun projects. The great thing about robotics is that it doesn’t take years of training to do something. You get in with an undergraduate degree–the standard courses you have–and pretty quickly make something happen that can motivate you to go deeper and deeper into the area.