Dimensionality and Topological Stability of Neural Representations in the Human Brain Predict Learning Outcomes

Recent advances in deep learning suggest that the dimensionality of model representations shapes generalization. We ask whether neural activity exhibits a similar principle during human learning. Using longitudinal fMRI collected from a real university course, we quantify representational geometry with intrinsic dimensionality and topological analysis. We find that learning outcomes depend on brain region: in association hubs supporting conceptual abstraction (e.g., Angular Gyrus), lower dimensionality predicts better performance, while in regions supporting complex perceptual processing (e.g., Temporal Fusiform Cortex), higher dimensionality predicts better performance. Topological analysis further shows that high-performing individuals form more stable structures in association hubs and that their representational topologies diverge more strongly from one another. Together, these findings suggest that effective learning in the brain relies on region-specific representational organization, with stable and individualized structures that support successful performance.