Every field of study is like its own thousand-layer cake. There’s always something more intricate and specific to research, and only the most passionate minds are willing to dig so deep. After some self-discovery, Dr. Alnur Ali became one of them, a scholar who can never be satisfied with surface-level knowledge, going as far as getting his postdoctoral degree at Stanford University.
A postdoctoral program is similar to a medical residency program in which students with PhDs have the freedom to work on their own research projects as a part of a larger academic community, according to Ali. Participating in a postdoctoral program is highly recommended for people interested in becoming STEM professors, Ali said.
Although Ali is on his way to successfully publish 3-4 papers by the end of the summer, he said there was no master plan. In fact, early on in his career, he said he couldn’t decide what to study for graduate school. He was interested in computer science, neuroscience, and cognitive science, but didn’t know which field actually made him want to learn more. Thus, Ali took a different route than his peers.
“Most people go to undergrad and then grad school. Some continue to postdoc to be a professor or work at Microsoft. Mine was the reverse. I started at Microsoft,” he said.
That’s right. Fresh out of undergraduate college, Ali went to do research at Microsoft for an early version of the Bing search engine. That was when Ali was first introduced to machine learning and its applications to recommendation systems. Although he read into neuroscience and cognitive science, Ali realized that he wasn’t as interested in diving further into those fields as much as machine learning.
“Neuroscience was too much at the micro-level like learning how each cell worked,” he said. “Cognitive science felt as if you were hanging out with a group of friends at night and someone asked ‘How do you think babies manipulate numbers?’ Any theory your friends came up with could be a subject in a cognitive science research paper. I wasn’t getting the answers I wanted.”
Because of this, Ali went back to machine learning which had everything he wanted: statistics, computers, making an impact, and a loose connection to how the brain worked. Soon, Ali lost himself in research and learning many artificial intelligence concepts on his own.
“I just needed to know more,” he said. “At Microsoft, I knew at a high level what I was doing, but not at a deep level. I didn’t know enough to help improve those models.”
This dissatisfaction with his own lack of knowledge helped Ali realize just how far he was willing to dig into the machine learning cake. Ali quit working at Microsoft and returned to school at Carnegie Mellon with a plan, confident in his passion.
Now, as a postdoctoral researcher, Ali does not consider his time at Microsoft as just a self-discovery detour, but as a valuable technical advantage to his research.
“I would recommend working before graduate school not just for life experience but to find the problems that really matter. Bing was a living, breathing search engine that real people used,” he said. “The problems it had actually mattered. So when I went back to school, it was the lens in which I viewed everything. I couldn’t focus on research if there was no possibility that they would be used in the real world.”
After all these years, Ali has totally immersed himself in recommendation systems and due to his experience at Microsoft, he already knew what problems the Bing search engine suffered through, and he spent his research programming ways to solve them.
One thing Ali noticed at Microsoft was that programmers were usually confident in getting data and training a machine learning model, but there was not enough attention to what happened after a model was deployed into production.
Oftentimes, programs begin to behave less efficiently or like how Ali liked to phrase it, “Your model does fine at first but then it starts sucking.”
The first problem a model hits is something called a distribution shift where the nature of users has shifted over time to new users the model wasn’t prepared for.
“Let’s say you use college student data to train your model but one day, your user base turns out to be retirees. Those are very different people,” Ali said. “An example from Microsoft was search queries. When Obama became president everyone started searching for Obama, but your model wasn’t really trained on Obama so it didn’t know too much about him before.”
Ali also observed a fairness issue in which the models would work better for a well-represented group of people in the data like white males, “but for guys like me, maybe not so much.”
Passionate about these issues, Ali programmed a solution with the ability to detect when inconsistencies within a model were occurring and could specify the cause or sub-group of people affected. A program that could automatically pinpoint the weak link of the model and retrain a targeted niche part of it in real-time would be very helpful. Of course, this is not the complete answer to every recommendation problem as there are tradeoffs that Ali further explores in his upcoming papers.
Regardless, the solutions Ali presented would not have been made possible without understanding the important problems he recognized during his industry experience. With it, Ali became aware of the problems affecting professional applications that impacted people worldwide.
In the end Ali found himself at a crossroads again on what he plans on doing after completing his postdoc. However, Ali said it’s clear to him that there are still ways to gain even more knowledge on the subject of machine learning and that the answers he receives will only lead to more opportunities in his future.
“If you’re really interested in something, questions will present themselves to you,” Ali said. “There’s probably an answer so keep going, go to a library, check out a book, and start reading. And you don’t know where that’ll lead you.”