What is your specialty?
I work in the space between statistics and machine learning.
Statistical models are nice because they’re interpretable, probabilistic, and reliable to train. But they can struggle with large, complex datasets. On the other hand, machine learning models are great for large, complex datasets, but many of them are black boxes which are hard to interpret, and the training process can be unreliable and resource intensive.
My research is about developing new methods which blend the best of these worlds.
How did you end up in that area?
My PhD was actually in cognitive science, running experiments on how people think. But I very quickly got much more interested in modeling data and neural computation. It involved a lot of math, so I started building up my math chops from scratch, going through linear algebra and differential equations. After that I started doing real analysis, dynamical systems, and probability. When I finished, I realized that I wasn't going to use my PhD in cognitive science directly. I just really liked math and modeling data.
So I went back to grad school and got degrees in math and statistics. Then I had this amazing job offer at a cybersecurity company building malware classifiers on huge datasets and building intruder detection models from active real-time sensors.
They were applying statistics and machine and machine learning back in 2012. It was a new world back then for cybersecurity. We had an incredibly diverse team approaching the problem from all these different perspectives. It was a great experience. After six years I just really missed the research. I went back to do a postdoc in machine learning at Tufts University in Boston and then I got my current job as a research associate at Harvard Biostatistics with a professor whose research I really admire.
What do you admire about that professor’s research?
He is just incredible. He scratches all of the intellectual itches I have. He is deeply good at the theory and deeply insightful about the big picture of things and how things work. But he still understands how to run a good experiment. He's genuinely very dedicated to his application, which is cancer detection.
How will your expertise apply to SMART FIRES?
SMART FIRES is right up my alley. The selection problem of finding where you would want to place a controlled burn is similar to other selection problems that I’ve worked on throughout my career. For example, in cancer research, you’re looking along the chromosome and you’re trying to see if there are places in the chromosome where there’s something weird happening called copy number alterations. If you see that, it’s an early clue that someone might have cancer. I’ve also worked on selection problems in other areas of research - stroke recovery, malware analysis, and soldier performance.
Another thing is that I’m very interested in is spatiotemporal problems where you have irregular time stamps or irregular spatial locations. Sensors don’t necessarily give information in regular time and aren’t always synced with each other. This problem came up at the cybersecurity company when we were developing models for intruder detection based on how people use their computers. I’ve recently been developing some new methods for such data that blends techniques from statistics and machine learning.
What are you most excited about with your new position?
I am excited to work on SMART FIRES, to teach classes related to probabilistic modeling of data, and to work with talented colleagues and students.
What do you do outside of work?
First and foremost, hanging out with my friends and family. I have a 9 and
a 7-year-old so I’m lucky enough that they usually still want to hang out with me. My hobbies have been changing recently because we’ve been moving around, but I enjoy hiking and racquetball. My goal is to learn how to skate ski.