Interview with Hadi Salman, AI Resident


Hadi Salman, AI Resident
Tell us about yourself and why you decided to join the AI Residency Program.
I did my undergraduate work at the American University of Beirut and I graduated in 2016 with a double major in mathematics and mechanical engineering. I was always fascinated with intelligent systems and this drove me to join the Robotics Institute at Carnegie Mellon University, where I obtained a Masters in Robotics in 2018. During this time, I worked on topics at the intersection of robotics and machine learning, and it became clear to me that I wanted to delve more into and specialize in AI.
I applied to Microsoft’s AI Residency program because I thought it would be a perfect opportunity to solidify my knowledge and experience in AI, and get the chance to dive deeper in specialized topics. My impression about the residency was that it is a place where you can freely work on any AI topic you want alongside experts in the field, with the advantage of having a lot of compute resources (and my impression was correct!). Thus, it was clear to me that I would join the program if I got accepted.
In summer 2018, I interned at Uber ATG, where I worked at the intersection of machine learning and uncertainty modeling for autonomous driving. After that, I joined the Microsoft AI Residency program and the experience so far has exceeded my expectations!
Have you been able to grow your AI and ML skills in the first seven months of the program?
Absolutely! My AI skills have grown substantially since I joined the program, both in terms of implementation and analytical skills.
I learned how to build robust, large-scale experimentation pipelines for machine learning over a compute cluster. For instance, during the first few months, I worked on a project that required parallelizing millions of small jobs over a cluster of approximately 3000 CPUs. Many bottlenecks showed up when running at such a scale and I had never dealt with such scenarios before joining the residency.
Furthermore, I was lucky to be mentored by super-smart mathematicians and theoretical computer scientists focusing on solving AI problems. This boosted my math skills and made me a more rigorous researcher. It also made me look at AI from a totally different perspective than before joining the residency.
What are your goals for the next four months?
I am currently working on a couple of super-exciting ideas for building robust deep learning models. So far, I have published two papers and my goal is to finish some of the follow ups on these papers and hopefully get one or two more papers done before the end of the residency.
What is your favorite accomplishment thus far into the program?
My favorite accomplishment so far was building a low-latency job scheduling pipeline on Azure. This feature didn’t exist before on Azure and it is super useful when one encounters millions of jobs, each of which requires only few seconds to finish. In such a scenario, any overhead in job scheduling would cause extreme delays. I developed this pipeline as part of my first research project on LP-relaxed robustness verification of neural networks, which required parallelizing millions of small linear programming jobs over a cluster of approximately 3000 CPUs.
What is your typical day as a resident like?
The life during the residency is nice and flexible. I would describe it as the life of a grad student, but with no pressure and with way more compute resources! I typically head to the office around 10 or 11 in the morning. I spend most of the day at my desk working. I also attend some meetings, and I sit in on many talks and reading groups, and I discuss research ideas with the group of researchers I am working with. The residents usually have lunch together and we regularly plan activities outside work. The atmosphere is healthy and friendly, and the residents share ideas and expertise.