We are delighted to introduce this year’s Harvard Data Science Initiative Postdoctoral Fellows! Our fellows support postdoctoral scholars with interests in advancing the field of data science. Fellows are independent investigators who pursue their own research in collaboration with Harvard faculty. Read more about the HDSI Postdoctoral Fellowship Program.
Melody Y. Huang
“I am currently a Postdoctoral Fellow at Harvard, where I work with Professor Kosuke Imai. My research sits at the intersection of social science and statistics. I am broadly interested in developing statistical methods to credibly estimate causal effects. During my time at Harvard, I’m looking forward to meeting other researchers and starting new collaborations. Before coming to Harvard, I received my Ph.D. in Statistics from the University of California, Berkeley, where I was fortunate to be advised by Prof. Erin Hartman. Outside of statistics, I enjoy running and board games.”
Melody Y. Huang, 2023 Wojcicki Troper HDSI Postdoctoral Fellow
Keyon Vafa
“I’m excited to join the Harvard Data Science Initiative as a postdoctoral fellow. My research focuses on developing machine learning methodology for applications in labor economics and political science, among other fields in the social sciences. I completed my PhD at Columbia University earlier this year, where I was an NSF GRFP Fellow and Cheung-Kong Innovation Doctoral Fellow. Outside of research, my hobbies include running, watching movies, taking part in trivia nights, and playing Monopoly Deal.
My data science journey actually began at Harvard, where I was an undergraduate in the class of 2016. I was first introduced to statistics and probabilistic thinking as a sophomore in Joe Blitzstein’s Stat 110 class (I still use the Stat 110 textbook today), and I ended up concentrating in statistics and computer science. Coming back to Cambridge, I’m looking forward to meeting people at Harvard at all stages of their data science journey.”
Keyon Vafa, 2023 HDSI Postdoctoral Fellow
Johannes Knittel
“From a young age, I was very curious about science and how the world works. As a child, I loved books that showed the insides of devices and appliances so I could speculate about their inner workings. I still remember being a bit disappointed with the depicted inside of a refrigerator. It looked too simple and I could not imagine how “energy” – that I associated with heat – could actually cool something. It would remain a mystery to me for a few years until I learned the physics behind evaporation-based cooling in school.
A pivotal moment in my life occurred when I had the opportunity to try out my dad’s old computer, running DOS and Windows 3.1. I was fascinated and quickly fell in love with computers, how I could run “programs” that oozed some form of inner machine-life. Again, I was very curious how computers work, how I could create such programs myself for automating all parts of life, particularly programs with a graphical user interface that allowed me to interact with the system in an intuitive way. This was right up my alley since I absolutely despise tedious, repetitive work (ChatGPT would call me a “monotony loather”). It eventually led to me studying computer science and even pursuing my PhD in that field, researching new methods for performing analyses on large datasets by combining machine learning with interactive visualizations.
Machine learning (or AI) is not a new topic, but with more computing power and better architectures, we have now reached a point where we can truly utilize AI in many disciplines for the better (mostly). Machine learning fascinates me for many reasons, not the least because of its connection with the mystery of the human brain and human intelligence. In more rational terms, it is essentially a way of producing programs automatically with the help of data. We can make use of so much information that is already available but hidden in datasets, without needing to analyze the data manually for obtaining explicit knowledge that we would then incorporate in our algorithms (monotony loathers love this).
The downside of this is that we lose our understanding of how programs work. Machine learning does not produce magic boxes. In contrast to the human brain, we can look at the wiring. We could write down all the operations that make up the trained model. We cannot comprehend the model, though, because there are too many intertwined operations. They form magic boxes to us, leading not only to a lack of trust but also to people overestimating or underestimating the capabilities and limitations of machine learning. In some cases it is ok to have these magic boxes, but in more delicate applications we might not trust the model sufficiently well that it would come up with reasonable and unbiased decisions most of the time. Even for non-critical applications, we might want to better understand whether the model has learnt actual concepts that could even lead to the discovery of new insights, or whether it focuses narrowly on what it was trained on using detailed yet shallow heuristics.
I am far from naive, we will not be able to fully understand big models with billions of parameters in any foreseeable future. If I can contribute to making machine learning models more explainable at least for some aspects in collaboration with my Harvard colleagues, I would be very happy. The good thing about mysteries is that they provide a wonderful source of flowery and deep philosophical debates among friends.”
Johannes Knittel, 2023 Wojcicki Troper HDSI Postdoctoral Fellow