Dylan Zapzalka

Email: dylanz@umich.com

Download CV

I'm a first-year Computer Science Ph.D. student at the University of Michigan advised by Maggie Makar. Before landing at UMich, I majored in Computer Science and Mathematics at NDSU where I was advised by Saeed Salem. I also worked as a software engineer at Veritas Technologies. My research interests are in the areas intersecting machine learning and causality. In particular, my research involves finding ways to make machine learning models robust to spurious correlations and fairer to marginalized communities. My goal is to become a researcher in the industry after I obtain my PhD in 2028. Outside of academia and work, I enjoy reading and powerlifting.

Education

Aug. 2023 - Present

University of Michigan

Ph.D. Student

I'm currently a Ph.D. student advised by Maggie Makar studying machine learning.

Aug. 2019 - May 2022

North Dakota State University

Undergraduate Student

I studied computer science and mathematics and was advised by Saeed Salem.

Aug. 2017 - May 2019

Central Lakes College

High School Student

I attended Central Lakes College for my last two years of high school.

Experience

May 2022 - Aug. 2023

Veritas Technologies

Associate Software Engineer

Worked as a software engineer where I developed features to integrate NetBackup with cloud providers.

Jan. 2022 - May 2022

North Dakota State University

Research Assistant

Worked as a RA where I wrote a paper on how to make adversarial examples for GNN malware classifiers.

May 2021 - Aug. 2021

Pearson VUE

Software Engineer Intern

Interned at Pearson VUE where I worked on their credential management system.

Papers & Projects

Current Project

My current project involves developing a novel causally-motivated regularization scheme to make machine learning models more robust to confounding variables. Unlike previous approaches, this regularization scheme does not need access to auxiliary labels with information on the confounding variable. Instead, this approach utilizes known mediators between the input and the target variable. This regularization scheme is useful in instances where a wealth of knowledge is known about the causes of a target variable, but not much is known about possible confounders.