Adam Coscia, a third-year Ph.D. student in Human-Centered Computing at Georgia Tech, is developing visual analytics tools to interpret large language models, a type of black box AI model. During the GVU Center Research Showcase in April 2023, he presented two tools that he is building which helps to interpret and explain large language models as part of his work with the AI-ALOE Institute.

Coscia’s first tool, KnowledgeVIS, visualizes the associations that language models learn to make during training, such as harmful gender biases. This tool enables AI-ALOE researchers to identify problems during training and prevent them from reaching students. After training, AI-ALOE’s language models learn to perform specific tasks. To understand how they work, Coscia has been visualizing backpropagation, attention scores, and different layers of the language models deployed in iTell, an intelligent textbook that can automatically score student writing.

He hopes that his novel methods will be adapted by other teams at AI-ALOE, enabling everyone to evaluate their AI before deployment. Coscia sees AI-ALOE as an amazing opportunity to improve human-AI interaction and innovate across multiple domains, ultimately having a positive impact on how researchers and learners interact with AI technology in the classroom.

About GVU Center Research Showcase

The GVU Center Research Showcase at Georgia Tech is a top event for tech enthusiasts and industry professionals looking for new ideas. Research themes include AI, AR, gaming, robotics, and more from top researchers from 20+ labs. Learn more at GVU Center Research Showcase 2023.