Jared Lee Katzman
I am currently a PhD student and researcher at the University of Michigan School of Information.
Previously, I worked in industry artificial intelligence (AI) labs, building scalable AI applications, where I learned firsthand the challenges of addressing algorithmic fairness, accountability, and transparency in production systems. I am currently pursing a PhD with the goal of understanding the competing narratives around how to address the risks and harms of algorithmic systems and inform methods for responsible AI development, regulation, and resistance.
My present research investigates discussions around the governance of algorithms in different contexts: researching frameworks for developers to identify algorithmic harms; understanding AI framing in political discourses; and educating the public on the risks of algorithmic decision making. In addition, I am exploring different embodied practices to understand how movement and somatic experience can excavate and resist internalizations of dehumanizing cybernetic philosophies.
I am always interested in new conversation and research collaborations,
and I am open to consulting on responsible AI development. Please reach out,
my email is the word for the warmth of sun in winter (apricity) @ umich.edu.
Experience
UM Science, Technology & Public Policy Program
Collaborated with community organization Detroit Disability Power to conduct technical landscape analyses on how automated hiring algorithms are impacting people with disabilities. Evaluated current regulations around algorithmic hiring and their gaps to assist with advocacy and education campaigns.
Microsoft Research
Led research for multiple projects, including statistical methods for polling and election forecasting, interactive data journalism to improve comprehension of politically-relevant data, and survey design for measuring impact of news media. Investigate a taxonomy for fairness and inclusiveness in image tagging and captioning AI services. Member of Fairness, Accountability, Transparency, and Ethics (FATE) reading group. Teaching Assistant for Microsoft Research Data Science Summer School, Summer 2021, an intensive, four-week hands-on introduction to data science for college students in the New York City area.
Out in Tech
Started Out in Tech U’s Digital Mentorship Program, where LGBTQ+ students work on projects with professional mentors over the course of 8 week semesters. Managed 4 volunteer coordinators to design and facilitate all digital programming for participants across the United States. Grew the program from an initial 30 participants to over 400 members annually. Participated in the Spring 2018 cohort as a mentor; Worked with a mentee, with no previous experience in data analysis, on how to use Python, Jupyter Notebooks, Pandas, NumPy, and Plotly to visualize the impact of America’s Opioid Crises.
Amazon Web Services
Founding engineer on team developing Machine Learning Bias and Explainability products and research across ML platform. Designed and implemented data explainability feature for SageMaker Autopilot within 1 month of joining team pre-launch. Researched technological and scientific vision and strategy for 5 new product proposals reviewed by senior leadership of AWS’s AI labs. Developed Entity Linking models on top of BERT which improved precision by 10% compared to baseline methods.
Researched, developed, and released IP Insights, a deep learning algorithm that learns the history of users' IP addresses to identify anomalous behavior such as fraudulent logins or account takeovers. Discovered and removed bottlenecks in algorithm's performance by tuning MXNet's distributed training, leading to a 50% reduction in training time for 75% of the comparable cost of scaling. Presented a talk on scalable, distributed machine learning on MXNet at Amazon's internal developer conference to 110 attendees. Designed and facilitated workshops on machine learning and security.
Bridgewater Associates
Built a real-time analytics platform with AWS services to transform security controls and operations from a perimeter, defensive model to a data-centric, automated-reasoning framework. Redesigned a batch processing architecture as a real-time, serverless, streaming framework which decreased a security control’s effect time from 24 hours to sub minutes and reduced monthly costs by a factor of 10.