Professor Karthik Sridharan earned his PhD from TTIC in 2012, and over the past nine years, he has crafted an impressive career in academia. Immediately after graduating from TTIC, he went on to be a Postdoctoral Researcher at the University of Pennsylvania Wharton School of Business, in the Department of Statistics. In the fall of 2014, Prof. Sridharan joined the faculty at Cornell University as an Assistant Professor in the Department of Computer Science, and has since been promoted to Associate Professor.
He chose to join the Department of Computer Science at Cornell because of the collaborative environment, and the emphasis on his area of research, theoretical machine learning. “I think research is the thing that I enjoy most, especially when I get to work with PhD students. One of the nice things about academia is that you can have long term projects, you work with a student for a while, about three or four years, so you get to build something with them,” said Prof. Sridharan.
He also enjoys working with undergraduate students at Cornell. “There are really bright undergrad students. It’s nice to see that you can have a positive influence on them, and see some of them wanting to do a graduate course, or to do research in the field. And it’s nice to know that the course that they took with you had an influence on them wanting to do this. That’s fun as well,” he said.
Prof. Sridharan became interested in machine learning during his third year of undergrad, at the Ramaiah Institute of Technology in Bangalore, India. After earning his B.E. degree, he came to the United States to earn a Master’s degree at the University of Buffalo. “I started reading about theoretical aspects of machine learning, and I really loved it,” he said. TTIC’s reputation for excellence in theoretical machine learning is what drew him to the PhD program.
“It was a new university when I joined, I think I was part of the third class in. I thought to myself, should I go to a place that has been around for a while, or to TTIC, but then TTIC really had the most fun people to work with. And so I ended up going there, and that was the best decision I made,” said Prof. Sridharan. He began the program working with Sham Kakade, and finished the program with Professor Nati Srebro as his advisor. He really enjoyed working with Prof. Srebro, and all of the Research Assistant Professors (RAPs) as well.
“The [RAP program] attracted a lot of researchers in the field, so apart from Prof. Srebro, I also got to work with Shai Shalev-Shwartz and Ambuj Tewari, who are all prominent people in the field, and they had a really good influence on me and helped me shape what I wanted to do. My thesis topic was heavily focused on online machine learning, and that was something that I picked up mainly because of Prof. Shalev-Shwartz and Prof. Srebro,” he said.
As a field, machine learning is still in fairly early stages. Though it can be found in many places, from unlocking your smartphone with facial recognition, to getting recommendations for what to watch next on Netflix, we are still relatively limited as to how we can use online machine learning. Prof. Sridharan believes that the key for machine learning right now is determining the best way to use all of the available data.
“[Machine learning] impacts us in a lot of ways. I think 90% of the time, hopefully, it’s in good ways, but it can also have a negative impact. As a researcher, I think that’s where a lot of our focus needs to be. We spent a lot of time figuring out how to make machine learning work, and there’s still a long way to go there, I’m not saying that it’s a problem that’s anywhere near where it should be. I think now it’s time to think about not only how well machine learning works, but also how do we make use of it in a way that’s helpful to society. Because it’s this raw power, we need to mold it into something that’s more useful, and that hopefully doesn’t do too much harm,” said Prof. Sridharan.
Much of his research began with an area called online learning, which is a process in which learning is continuous. For example, if one person is the “learner,” using a machine learning algorithm to learn something from a set of data, the data or person that the learner is interacting with is also evolving with time. Online learning is a machine learning framework that is capable of dealing with this kind of changing data.
So far, the basic framework of online learning does not work with a reactive learning framework, in which a learner takes an action, and the data source changes over time, but doesn’t necessarily react to the learner. “Once it starts reacting to [the learner], then we need to think about how we change or update machine learning algorithms to take care of this fact. Machine learning topics, like reinforcement learning and control, deal with a more reactive environment,” said Prof. Sridharan.
Recently, he has primarily been conducting research on machine learning in reactive environments. Part of this is figuring out how we learn — for example, if you make a decision today that goes into a machine learning algorithm, that decision will affect how the user reacts, so how does one learn in this type of scenario? Another question he addresses is a variant of this, where one must consider the societal impact this type of machine learning system could have.
“Let’s use the example of Google providing news to you. Google as an organization looks at your data and figures out that you like to hear news about music, and so it might look at this and start giving you more music articles. It figures out that you like jazz, so it starts giving you articles about jazz on day one. Maybe you just had a slight tilt towards liking jazz, but on day two, because it’s given you so many jazz articles, you start liking jazz even more and it starts to build up. You like jazz more and more and more. So when you were originally neutral, or only kind of interested in jazz, it converted you into a much bigger jazz fan,” said Prof. Sridharan.
“For music, maybe that’s not all that bad. But for news articles, for instance, which are more likely to be political, this might be a bad thing. This is a reactive machine learning problem where you want to learn how to give the users what they want, but you need to be careful about what you give them. If I convert you to be a fanatic on topic A, then as a learning problem, it’s very easy for me, because I just need to keep giving you articles about that. So my problem is solved as a machine learning person, because I figured out exactly what you want. In terms of what effect it has on people in general, it can be polarizing, and this is not good,” he explained.
Going back to the music example, if the program were to figure out that you like a particular artist, and continue giving you content about that artist for a week or two, you may be happy with that. But after a while, the end user may get tired of seeing content about that artist, and could leave the platform altogether. Even from an organization’s perspective, in the long term, they could lose customers.
This is another example of a control problem, where the learning algorithm is making predictions, or providing a control. For example, with facial recognition, the system can collect images, both of faces and of other things, and learn what a face looks like. This could work well in the future, but currently, for tasks where a machine learning algorithm is trying to learn and interact with users, the algorithm must be aware that it is in a reactive environment. This is more like a control, or a reinforcement learning problem.
Prof. Sridharan has been looking at both types of problems. He has been exploring how to use what we have learned from reinforcement learning and control literature to deal with problems like polarization, fairness, and other related topics.
“Fairness is another thing that’s very similar to the polarization issue that I talked about, in that fairness is basically where you’re trying to make sure that you’re treating all users equally, but you want to be aware of the fact that users also change over time. It’s a dynamic environment. You need to think about how to enforce fairness, when you deal with such reactive environments,” he said.
He has also been conducting research on non-convex problems, partial information problems, and their key challenges. In traditional machine learning, researchers are used to what are called full information problems, in which the training data contains all necessary information regarding the benefits and drawbacks of one model vs. other possible models.
This can be computed for every one of the training points, but when the models are deployed online in partial information problems, you choose to deploy one model over another, and only after deployment do you get feedback about that particular model.
“Dealing with this partial information is a key question. It turns out that when you take any of these reactive environment problems, and put in this partial information angle, the problem doesn’t just linearly become harder, it becomes exponentially harder. Dealing with such problems is a key aspect of new research,” Prof. Sridharan explained.
He credits his time at TTIC with teaching him a lot of essential values and skills that he has used throughout his professional career. TTIC helped him learn how to do good research, and really dig into problems. Prof. Sridharan also emphasized the importance of not trying to go after publication too soon, or doing research only with the goal of publishing a paper.
“I think you want to do research that you really enjoy, and you want to publish good work. If you have good work, you can wait for however long to publish. It’s going to come out and you don’t need to be worried. So I think not going after publications and just going after good research is something that I learned by example from the folks at TTIC,” he said.
His time in the PhD program also helped him learn how to give an effective talk. “I’m still learning how to give good talks, but I think of how much progress was made from when I started at TTIC to when I left. I still remember during my thesis proposal, they told me, oh, you’re done? I passed, but I was asked to come back and give the talk again, because they said I needed to learn how to give a talk so that people who are not exactly in my subfield understand what’s going on,” said Prof. Sridharan.
He also appreciated that during his first year and a half at TTIC, his advisor was completely fine with him focusing more heavily on classes rather than research. They understood the value of having a strong foundation, and he enjoyed having the ability to learn from more seasoned researchers in the field before conducting more of his own research and publishing papers.
Aside from coursework and research, he also valued the social aspects of life at TTIC. As a small academic community, the program has a close-knit, friendly atmosphere. Oftentimes faculty, students, and RAPs would all go out for lunch together. “It wasn’t like students were all in one place, and the faculty were all in another. It was very friendly, and I think what I enjoyed most was the open door policy. Folks were all very approachable, so if I had a problem, and knew this particular person was working on a similar topic, I could feel free to go and chat with them,” he said.
The open door policy extended beyond academics as well. During his time at TTIC, Prof. Sridharan also tried rock climbing for the first time with friends from the program, and swimming in Lake Michigan. Faculty members hosted dinners and barbecues that were open to the whole TTIC community.
He fondly remembers the chaos in the atrium leading up to deadlines for submitting papers to conferences in Europe. “Some of the big conferences used to have deadlines in the middle of the night, and literally an hour or so to before the deadline, you could see a bunch of people climbing up and down the stairs, trying to get their paper submitted at the last minute. It was fun, and everyone would order takeout in bulk and have dinner together while they were working on their papers,” he said.
Prof. Sridharan has accomplished a significant amount of research in machine learning since leaving TTIC ten years ago, and one of the accomplishments that he is most proud of is establishing a rigorous foundation for online learning, and figuring out how to derive online algorithms. This research was published across two to three papers that established a connection between computer science and math, in a field that is typically considered more math-based.
“In a sense, we were able to kind of see that there is what’s called an empirical process theory behind online learning. So that is something of a foundation, that’s more from pure probability theory, and functional analysis. We were able to not just do that, but show that you can actually use some of the some of the results from this kind of subfield in math to derive algorithms that you can actually use in practice,” he said.
In the future, he would like to continue working on reactive machine learning problems, specifically reinforcement learning style problems. Over the next four to five years, he is planning on exploring rich observability style problems, in which an action is made, and the observations come from a very large data set, making them more complicated. Typically, without extra assumptions, it is very difficult to learn these problems effectively.
“The classic tabular setting doesn’t quite work in these problems, so trying to understand their true complexity, and how to derive algorithms in a principled fashion is a goal for the next few years. The other goal that I’m really trying to push at is how to build machine learning algorithms that would be able to prevent polarization. The idea is to use some of what we learned from these RL control style problems and try to formalize polarization and bias amongst users through mathematical tools,” said Prof. Sridharan. He hopes to use these problems to derive algorithms that will minimize polarization and bias among users.