Salar Fattahi is graduating from Berkeley in Spring 2020 with a Ph.D. in IEOR. Salar recently accepted an Assistant Professor position at the University of Michigan and in a conversation with the department, he shared his views on experiences at Cal and plans for the future.
Tell us a bit about your childhood? Where did you grow up? What was it like there? How did you get interested in math?
I was born in Iran and my hometown is Urmia, a city in the northwestern part of Iran that is very close to Turkey, so my mother-tongue is actually Turkish.
I don’t know what made me interested in it, but I really liked math and physics since I was 6 years old. And then much later, after the nationwide university entrance exam, I got into Sharif University, which is a renowned University in Iran. I was doing electrical engineering, but the curriculum had a strong mathematical component to it. After my undergraduate [studies], I got into Columbia and started working with [Prof.] Javad [Lavaei]. I was in the EE program but was doing more math and optimization with applications in power and control systems. I spent one year at Columbia, got my masters and then moved with Javad to Berkeley. I got another masters here and hopefully soon, a Ph.D.
Just curious, what was it like living in New York coming from outside the US?
Actually New York was similar to Tehran. I did my undergrad in Tehran, and when I moved to New York I thought it’s very similar in the sense that it’s crowded and busy. Yeah, so it wasn’t a major shock, even though it was a big change.
Tell us about your research at Berkeley? Where did you start and what are you working on now?
The main theme of my research is to develop scalable and guaranteed computational methods for the efficient operation of large scale and safety critical systems, such as power systems, transportation systems, and brain networks. To do that, I develop tools in both optimization and data analytics, which are basically the two pillars of reliable computation in these kinds of safety critical systems. This is because within the realm of computational methods, there has always been a trade off between scalability of different computational methods and their performance guarantees.
For example, on one side, there are some very recently developed computational methods in deep reinforcement learning that can actually operate extremely complex systems with millions of parameters, like robots or self-driving cars. But the problem is that these kinds of methods often do not come with strong optimality and performance guarantees and therefore they’re prone to failure. On the other side, there are some extremely powerful, and very general computational frameworks in theoretical optimization that enjoy very strong optimality and performance guarantees. But these kinds of methods often quickly become intractable with the increasing scale of the problem.
Most of today’s systems such as power systems, transportation networks, or even brain networks are massive scale, so we need scalability in our computational methods and at the same time, they’re safety critical, so our computational methods should come with some sort of certifiable guarantees. The question that I’m interested in is to see if we can have the best of both worlds for these kinds of systems.
Can you elaborate on the brain-related applications of your work?
It is well-known that different brain regions interact with one another in response to different physical or mental activities. But we cannot directly see this connectivity network. And all we have is the limited number of fMRI scans on the brain. Understanding the underlying brain connectivity network is useful, especially for early discovery of different brain pathologies, such as Alzheimer’s. But the main issue is the scale of the problem. The full brain mass, for example, has more than 200,000, voxels, or nodes. So to fully understand this underlying brain connectivity network we essentially need to solve an optimization problem with more than 20 billion variables.
I try to answer the question of how to go from the data that neuroscientists collect from the brain to a meaningful description of the underlying brain activity. Our main method to solve these kinds of problems relies on the fact that the brain connectivity network exhibits some sort of sparsity structure, which helps us to solve the optimization problem that is related to brain inference. But these structures actually appear in other problems as well. For example, Power systems, transportation networks, and other interconnected safety critical systems have some level of sparsity in their structure. So this structure can be exploited in our computational methods to decrease complexity, improve runtime, and increase efficiency.
What sort of applications have you worked on in power systems?
Power systems are one of the largest and most expensive issues on earth, with an annual cost of $400 billion, and improving the operation by even 1% saves an enormous amount. In power systems a longstanding challenge is to match the demand and supply in real-time. This is a challenging task since today’s power systems have a lot of security constraints that should be satisfied at all times.The question is to then solve this problem, with all the difficulties, almost in real time. The problem is large scale and safety critical — a small error could lead to major blackouts. So we need to come up with better and more sophisticated computational methods that can actually exploit the underlying structure of the problem for efficient solutions.
Where do you see the problems you are working on going in the near future?
The problems I focus on, and pretty much all problems today, are becoming more data-driven. That is to say that data leads to the process of decision making. These kinds of problems have called for new methods that are at the intersection of optimization and machine learning. For example, there is an important problem related to power systems called state estimation — where you estimate the state of the system with a limited number of variables. This problem needs a combination of optimization and machine learning. Historically, the machine learning element was not understood well and also led to the northeast blackout of 2003.
So I would say that in the near future, most of the safety-critical systems will undergo a gradual shift from the legacy approaches where experts were the main decision-makers, to data-driven approaches where AI is in charge of decisions with human inputs. This transition requires a new scalable and guaranteed computational paradigm, and I believe that the field of OR will help us achieve this goal.
What was it like to work with your advisors, Profs. Javad Lavaei and Somayeh Sojoudi?
I started working with Javad six years ago when I joined Columbia. Javad has helped me center my research around optimization with applications in energy and control systems. Javad has not only been helpful with his inputs on my research but also encouraged me to employ new directions in my work which allowed me to publish more papers in top journals.
I started working with Somayeh three years ago and, under her supervision, I added machine learning and data analytics as new dimensions to my research. My work with her was also very fruitful and we published our joint works in top journals and conferences in machine learning. I am very grateful to them for their support in my Ph.D.
Congratulations on your new position as an assistant professor at the University of Michigan! Where do you see your research interests in the next 5-10 years?
I am very happy to join the University of Michigan. It will be a great opportunity to work with stellar researchers and students on problems I care about. The collaborative environment in Michigan is also something I look forward to. I am sure it will help me work on my long-term research plan, which is to develop more efficient computational methods for the next intelligent systems that are driven by data, such as smart infrastructure and smart cities, and to also induce elements of safety and robustness for such computational problems.
What excites you most about being a professor?
Added independence in research would be very exciting. I look forward to having the freedom to choose the problem and projects that I am passionate about to work on. Another aspect is to help the next generation of researchers by nurturing their passion towards research.
What is the most exciting thing about the field of IEOR right now?
I think the most exciting thing about IEOR is data-driven decision making. The data-driven nature of today’s interconnected systems is very important, and I think the field of IEOR will play an important role in the efficient operation of these kinds of systems. There is also a real push from different industries and federal agencies to integrate OR into real-world problems. In fact, very recently the department of energy announced millions of dollars in funding to integrate AI and ML into the energy sector and OR is at the core of this integration.
Do you have any advice for future Ph.D. students? What might have you done differently if you had a chance to start over?
My advice is to start doing research as soon as possible. It would be very helpful for Ph.D. students to start doing research from their first year. Find an advisor… and just start. If you don’t like the project you are working on, you can always change directions.
More importantly, don’t be afraid to talk to different people in the department, in your field of interest. Do not isolate yourself and work on your own. That wouldn’t give you any vision to work on impactful problems. Besides, it’s always helpful to collaborate.
What will you miss most about Berkeley?
First and foremost, I will miss my friends here; I will never forget our friendships and will certainly keep in touch. I will also definitely miss Berkeley’s weather and its unique culture. I will miss the collaborative and collegial environment we have in this department; it really helped me shape my research. Last but not the least, I will miss our departmental soccer team, The Martingales.
The IEOR department wishes Salar Fattahi the very best for his future!