Reducing bias in an AI-enabled STEM learning environment

May 8, 2023
Class of children sit in computer lab working on monitors

The University of Oregon is a contributor to a national artificial intelligence research institute focused on developing learning technologies to accelerate young people’s achievement in science, technology, engineering, and math (STEM). 

The U.S. National Science Foundation announced today it is awarding $20 million over five years to a team of institutions to establish the Inclusive and Intelligent Technologies for Education (INVITE) Institute, based in the College of Education at the University of Illinois.  

Three children look at laptop screen.

At the UO, the research effort will be led by Daniel Lowd, an associate professor in the Department of Computer Science, who studies learning and inference with probabilistic graphical models, adversarial machine learning, and statistical relational machine learning. Lowd will lead the effort to collect, analyze, and share large-scale datasets from diverse populations to create digital environments that are welcoming and inclusive for underrepresented groups such as girls and students of color. 

These data will be used to develop intelligent learning systems that are socially aware — flexible and responsive to students’ interests, sociocultural influences, cognitive needs, and behavior — surpassing current technologies by inferring individuals’ needs and dynamically adjusting instruction. 

A key purpose of the INVITE research is to broaden engagement with and learning of STEM among historically marginalized groups at the prekindergarten-12th-grade levels by investigating emerging AI techniques and building intelligent technologies. 

Lowd and several graduate students will minimize noise and bias in the datasets, which could lead to bad models and analysis, as well as a learning environment that will ultimately not be representative of a diverse group of students. 

“Much of my research is in adversarial machine learning, which studies how people can manipulate machine learning models — anything from spam filters to image classifiers to large language models like ChatGPT,” Lowd said. “In some cases, this is by manipulating the input, such as changing a few words in a spam email to bypass a spam filter. In other cases, it might be by creating false patterns in the training data, so that any model trained on that data will make certain mistakes.  

“My hope is to apply this kind of adversarial analysis to ensure that the datasets and the models we train on them are as fair and robust as possible.” 

The new generation of learning systems that the INVITE group hopes to create will be radically more responsive to learner needs, behaviors, and development and be designed to support the whole learner, beyond discipline-focused achievement. Use-inspired research will focus on how children communicate STEM content, how they learn to persist through challenging work, and how teachers support and promote noncognitive skill development. The resultant AI-based tools will be integrated into classrooms to empower teachers to support learners in more developmentally appropriate ways. 

Lowd’s portion of the $19.5 million grant (over five years) is $1,109,778. Other collaborating institutions include University of Illinois, University of Florida, Temple University, and the Educational Testing Service, which offers the GRE (graduate school admissions) and TOEFL (English proficiency) tests, among others. 

— By Kelley Christensen, Office of the Vice President for Research and Innovation