AI Hippocampus press

Our lab’s work recently appeared in Neuropsychologia, reporting a new test of whether individual differences in brain structure can predict individual differences in spatial navigation performance using modern deep learning approaches. The article was covered by UTA, MSN, Neuroscience News, and Medical XPress.

In the paper, led by Ashish Sahoo we trained and compared graph convolutional neural networks (GCNNs) and 3D convolutional neural networks (3D-CNNs) on T1 MRI data (N = 90) to predict navigation ability measured with an objective virtual-reality spatial memory task (participants drew maps of a highly realistic virtual environment). While the models fit the training data well, prediction in held-out test data was weak—suggesting that (a) much larger datasets and/or richer behavioral measures may be needed for reliable brain–behavior prediction, and (b) in healthy younger adults, hippocampal/macroscopic structural features may not be a primary driver of navigation ability.

This project was a cross-disciplinary collaboration spanning psychology, engineering, and industry partners (including colleagues at NVIDIA), bringing together expertise in spatial cognition, MRI, and machine learning to clarify what brain structure can—and cannot yet—tell us about navigation behavior.