ARCS Scholar Lizzie Kumar has become a scientific sleuth at catching errors in artificial intelligence (AI) algorithms.

Studying under Dr. Suresh Venkatasubramanian, a top computer scientist and leader in the AI field, the University of Utah PhD candidate is conducting research on new methods that make it possible for machines to learn from experience, adjust to new data inputs, and perform humanlike tasks.

Most AI examples that we hear about today—from targeting web searches to detecting bank fraud or diagnosing breast cancer—rely heavily on machine learning technology.

According to Kumar, developers of machine learning algorithms build decision-making models based on sample data. Computers then identify patterns in the historical data and make predictions about how people are likely to respond in certain situations or on certain tasks, with minimal human intervention. The algorithms improve automatically with experience.

Yet the metrics used to build these complex algorithms frequently have faults that can result in severe consequences. Kumar explains: “People developing these systems often don’t have a deep understanding of what certain metrics can and can't tell them about the algorithms they’re building. If they are misinformed about the metrics or misinterpret them, developers may deploy a decision-making algorithm they think is performing well but in actual practice fails because of external factors, such as asking the wrong questions, not having enough data, or having too much data. In a medical setting, for example, the results could be fatal.”

Kumar is making it her mission to analyze and quantify how certain explanations of data-driven AI algorithms are formulated. “My colleagues and I are looking at whether these methods are confusing or outright misleading with respect to the statistical relationships among the features in the data.”

She hopes that her work ultimately will lead developers to consider how people interpret the output of the systems they build and create algorithms that make AI a valuable tool to augment human efforts in coming up with solutions to vexing problems.

However, for developers to create tools that make choices for the good, Kumar believes stricter guidelines and regulations are needed to prevent deployment of machine learning systems that intentionally or unintentionally cause or perpetuate harm.

Since machine learning is becoming commonplace in almost every aspect of our lives, she is using her work to lead the way to accurate, safe, and fair AI development practices, especially in areas with tremendous influence and impact on the world.

In 2019, an ARCS Scholar Award started Lizzie Kumar on her research path to understanding machine learning systems, with an emphasis on analyzing their social impact. Her ultimate goal is to inform the development of law and policy to prevent the intentional or unintentional deployment of harmful data-driven technology. In June 2021, Lizzie Kumar will graduate with a doctoral degree from University of Utah.