Computer Science Seminar: Samantha Chen (University of California San Diego)

Speaker: Samantha Chen (University of California San Diego)

Title: What kind of algorithmic problems can neural networks solve?

The seminar will be available for in-person and Zoom participation. To participate online, please email inquiry-cs@barnard.edu to receive the Zoom link.

Neural networks have transformed the way we approach problems from across domains from physics to medicine to ecology. However, a persistent and critical challenge remains in their ability to generalize to out-of-distribution inputs (i.e. input that differs significantly from their training data). One way this problem has been mitigated, in the domain of combinatorial optimization, is through the framework of algorithmic alignment, where one designs neural networks to resemble specific algorithmic paradigms (e.g. dynamic programming) in order to encourage the network to learn a generalizable subroutine. I will discuss how algorithmic alignment improves out-of-distribution performance for single-source-shortest paths and for a family of geometric shape-fitting problems known as extent measure problems. 

Samantha Chen is a 6th-year PhD student at University of California, San Diego advised by Yusu Wang. Their research is in the intersection of machine learning and computational geometry and has worked on problems related to graph neural networks, geometric deep learning, geometric algorithms, and optimal transport. Before Samantha was a student at UCSD, they did their undergraduate degree in mathematics and computer science at Carleton College in Northfield, MN.