Predicting conceptual understanding through key information encoding during a STEM lecture

Poster Session C - Sunday, April 14, 2024, 5:00 – 7:00 pm EDT, Sheraton Hall ABC

Yeongji Lee¹ (yeongji.lee.gr@dartmouth.edu), David Kraemer¹; ¹Dartmouth College

When students are presented with new concepts during a science lecture, successful understanding relies on effective encoding of individual units of information as well as subsequent integration of these encoded pieces into a unified network that captures their underlying relationships. Using the embedding space of a large language model, a prior study revealed that the semantic network structure of a lecture can predict which units of information are most likely to be recalled, and also which units of information are most indicative of subsequent understanding of the central concepts. These results identify the key time points of a lecture where the information presented is critically predictive of subsequent comprehension of the main concepts of the lecture. In this follow-up fMRI study, participants watched the same video lecture – focused on several Newtonian physics concepts – and subsequently were asked to verbally recall what they remembered and learned from the lesson while still inside the scanner. Focusing on the specific time points where information is presented that most strongly predicts subsequent understanding of the central concepts, we use multivariate neural classification and related analysis methods to characterize the neural patterns associated with encoding of information for successful subsequent conceptual understanding. These findings highlight new approaches for understanding how conceptual understanding is built up over time – such as over the course of a lecture – and reveal new insights about the neural processes underlying this type of abstract knowledge acquisition that is critical for STEM education.

Topic Area: LONG-TERM MEMORY: Semantic