Machine Learning At Scale: Making Billions Of Data Useful; Examining Biases in Facial Recognition Technology Through the Lens of Actor-Network Theory

My technical work and my STS research are connected through the concepts of Actor-Network Theory (ANT). My technical capstone is centered on developing and deploying a scalable data pipeline and machine learning models for predicting military aggression using multi-modal sensor data. In my technical project, I constructed a network of human and non-human actors to accomplish this goal. In parallel, my STS research paper analyzes the wrongful arrest of Robert Williams, the first known false arrest in the U.S. caused by a facial recognition mismatch. This failure resulted from the collapse of a sociotechnical network with various human and non-human actors responsible for the unlawful arrest. While the technical project focuses on optimizing algorithmic pipelines and the STS research on analyzing their failures, both entail a critical examination of how large-scale data systems interface with human decision-making. Studying the failure of a sociotechnical network like the one in the Robert Williams case provides critical insights into maintaining accuracy and accountability when constructing a similar network.

My technical project was developed during an internship with a startup specializing in data-driven strategy systems. I designed and deployed an end-to-end machine learning pipeline, which ingested, cleaned, and stored sensor data before feeding it into transformer-based time series models to predict threats. The project involved automating feature selection and model retraining while reducing model training time and maintaining high predictive accuracy. The system is a sociotechnical network designed to aid human stakeholders in making time-sensitive, mission-critical strategic decisions using reliable threat forecasts.

My STS research paper studies a similar algorithmic system used in law enforcement but focuses on why this particular sociotechnical network failed. In 2020, Robert Julian-Borchak Williams was falsely arrested outside his home in Michigan because of an erroneous match from a facial recognition system used by the Detroit Police Department. My research employs the STS framework of Actor-Network Theory to study Williams's false arrest as a consequence of the improper integration of human and non-human actors. It examines the systemic breakdown of law enforcement processes as a failure of a larger sociotechnical network, resulting from insufficient translation in the recruitment process of actors.

Working on both projects simultaneously helped improve my understanding of engineering practices. Building an actor-network during my internship motivated me to study the development and evolution of networks, specifically, why such networks fail. Constructing a network provided insights into the vulnerabilities that can creep into the design of such systems. Analyzing the Robert Williams case through the lens of Actor-Network Theory prompted me to reflect on my work in the technical project, including components that functioned well and aspects that could be improved. Working on the Robert Williams case helped me look at the network I built from an Actor-Network Theory perspective, and the knowledge gained will inform all my future engineering projects.

School of Engineering and Applied Science

Bachelor of Science in Computer Science

Technical Advisor: Briana Morrison

STS Advisor: Benjamin Laugelli

Technical Team Members: Saarthak Gupta