Purdue computer science PhD student in Indianapolis thrives using AI to model human cognition and learning

INDIANAPOLIS — As a Purdue University master’s degree student in electrical and computer engineering in West Lafayette, Karen D’Souza was impressed by the depth and breadth of the research offered through the university’s Department of Computer Science. So much so that D’Souza was convinced to return to Purdue to complete her PhD in artificial intelligence and machine learning in the educational space.

Since fall 2021, D’Souza has been a vital member of the research group led by Snehasis Mukhopadhyay, a Purdue computer and information science professor in Indianapolis who became her PhD advisor.

“It seemed like a great fit for me to dive into cutting-edge research,” says D’Souza, who earned a bachelor’s degree in computer engineering in Bengaluru, India. D’Souza was especially drawn to the Purdue project that, through the use of AI, aimed to boost participation of chemistry students from all backgrounds. “Overall, Purdue University provided a great research environment for me to thrive in the PhD program in computer science,” she adds.

Thriving indeed. While interviewing for the research position with the team led by Mukhopadhyay, Chemistry and STEM Education Professor Pratibha Varma-Nelson, and Indiana University computer science professor Shiaofen Fang, D’Souza learned more deeply about the group’s AI in education project. That also piqued D’Souza’s interest since her goal was to build a unique pedagogical solution while working closely with the chemistry department and an existing online class.

The team’s research also fit D’Souza’s interests in developing solutions in the education field, specifically for multimodal machine learning and human-in-the-loop AI. Here, humans actively  participate in the training, evaluation and operation of ML models, providing valuable guidance and feedback. She also was struck by the interdisciplinary nature of the program’s research, combining computer science (AI), chemistry education, psychology and even sociology.

“At Purdue University in Indianapolis, I applied to the AI in education project out of a deep curiosity about how and why machines can be designed to model human cognition and learning processes,” D’Souza says. “I explored each of these areas in depth, ultimately developing an integrated framework for pedagogical AI.”

Responding to need for AI tools for educators

As data from multiple modalities of text, images, audio and video increased — accelerated by the shift to hybrid and remote learning during the COVID-19 global pandemic — D’Souza says a clear demand emerged for AI tools that could provide context-aware feedback to educators. 

“Multimodal AI offers a holistic approach to a complex data representation challenge, while a human-in-the-loop approach ensures that I was able to make context-aware corrections by including humans at every iteration,” she says.

D’Souza’s research centers on an online undergraduate chemistry class that employs a pedagogical tool called cyber peer-led team learning (cPLTL). In this format, small groups engage in peer group learning, tackling complex topics without a traditional instructor. Her work models AI to assess various dimensions of peer learning including critical reasoning, teamwork and problem-solving. But there was a hitch. Confronted with a lack of public education datasets owing to privacy concerns, she explored generative AI solutions using large language models.

“One of the highlights of my research was comparing AI analysis from large language models with human evaluation. This was valuable because it was never explored before using generative AI,” she says.

While concentrating on her studies and research aspirations as a Boilermaker, D’Souza jumped at the chance to develop the necessary skills outside the classroom and laboratory — in case she chose a career collaborating as an academic with industry:

• She participated in the semester-long AI Safety Purdue course, a student-run initiative, where D’Souza gained insights into mechanistic interpretability and its critical role in developing safe AI systems before deployment.
• She actively participated in weekly peer group learning with the Purdue Machine Learning Club, staying up to date with the latest advancements in the field.
• And she joined Purdue’s Society of Women Engineers, networking with industry and gaining tips to grow as a leader.

When Mukhopadhyay encouraged D’Souza to consider teaching, she spent a year as a lecturer for a Purdue undergraduate computing class — an experience she found more rewarding than she expected. “I quickly learned that classroom discussions sparked new ideas helping me articulate complex topics more effectively even in my research. It was one of the most valuable experiences of my time at Purdue,” D’Souza says.

Tapping learning activities beyond the classroom

Beyond the classroom, she secured internships with Dell, IBM Research and the Pacific Northwest National Laboratory, gaining valuable industry perspectives that enriched her studies and strengthened her research forte. “These experiences were instrumental in my dissertation as they allowed me to directly apply my knowledge to the research,” she says.

Mukhopadhyay says D’Souza’s research, which has focused on developing essential algorithms to model a multimodal and human-in-the-loop machine learning solution from multiple data channels, will have applications in collaborative STEM education.

“Her work is the first of its kind in STEM education, where integration of traditional machine learning and LLMs (large language models) using a human-in-the-loop approach has not been explored previously. Research in these areas requires a strong foothold of theoretical concepts,” Mukhopadhyay says. “And Karen’s academic competence, persistence, self-motivation and open mindedness to discover new ideas helped drive the project forward.”

D’Souza also is working on Mukhopadhyay’s high-profile health care project, which uses AI to predict the possibility of sinusitis surgery. She says the skill sets drawn from her dissertation focus transferred seamlessly to health care since both education and health care face similar privacy and security risks. Both domains also lack publicly available data.

This collaborative research effort was awarded a five-year, $3.3 million National Institutes of Health R01 grant. Mukhopadhyay’s AI portion accounts for $1.4 million of that total. The predictive AI models he and his team helped develop could eliminate cases where there is no medical benefit to sinus surgery at all, relieving hefty financial burdens, as well as significant physical and mental costs for millions of Americans.

“While these case studies are often perceived as big data problems, they can in reality be approached as small data problems,” D’Souza says. “The project itself is a one-of-its-kind opportunity to develop an advanced AI solution that tackles both multimodal data and integrates subject matter experts to enhance accuracy and outcomes.”

While continuing her PhD, D’Souza also is working as a computational scientist in generative AI at the Idaho National Laboratory, offering her a unique opportunity to serve as a postdoctoral researcher in probabilistic risk modeling. At the laboratory she leverages her expertise to develop advanced AI applications for nuclear science and technology.

“Dr. Mukhopadhyay has been monumental in my success at Purdue,” D’Souza says. “I attribute my success to his guidance and the dynamic research environment he fostered. He encouraged me to collaborate with experts across various fields including computer visualization, multimedia, education, and psychology, broadening my perspective and strengthening my research. Beyond that, his unwavering enthusiasm – even when experiments did not go as planned — was truly inspiring and had a profound impact on my journey.”