No assumptions when tackling problems

Friday, June 12, 2020

While a high school student in Toronto, Maysum Panju knew that the University of Waterloo was the destination for math. While working on his undergraduate degree and Master’s in computational math, he started learning about machine learning. That led to his interests in developing algorithms and theoretical proofs and he decided to start his PhD in statistics.

Now in his fifth year of his PhD, Panju is focused on tackling a problem called symbolic regression. Symbolic regression asks why we should assume that we know what the pattern is based on the input variables, and instead looks at the data to find the best possible model that will fit without making any assumptions that we know the output. It produces a rule in the form of arbitrary mathematical expression with no pre-specified structure.

“A common application of this could be learning the natural laws of physics. If you’re given data about a physical system, you don’t necessarily need to know the different physical laws, you can uncover the laws by uncovering the data sets only,” explained Maysum. “That’s the problem I’m trying to solve.”

The availability of neural networks has given Panju a new way to tackle this problem. The more powerful, versatile tools provide improved models. Working with his supervisor Ali Ghodsi, he’s making progress towards developing models that work with consistent success.

Panju has always been interested in practical applications, but enjoyed the efficiency, accuracy and data requirements of theoretical proofs. His graduate work in statistics has allowed him to pull these two interests together. His presentation of that research earned him an award from the Department of Statistics and Actuarial Science last year, and he recently competed in the Three-Minute Thesis (3MT) competition.

Remote video URL