Neural correlates of image memorability: Combining large-scale 7T fMRI with machine learning-based predictions

Abstract

We combine ultra-high-resolution fMRI with machine learning predictions to investigate the neural correlates of image memorability. Using ViTMem, a visual transformer-based tool, we estimated memorability scores for over 70,000 natural scenes from the NSD dataset. Our analyses reveal that higher memorability scores correlate with increased activation in ventral visual regions (fusiform gyrus, middle occipital cortex) and medial temporal areas (hippocampus, amygdala). Notably, we also observed significant engagement in fronto-parietal regions, indicating a broader network involved in modulating memory encoding. These novel findings demonstrate that ViTMem not only reliably predicts stimulus memorability but also provides a powerful framework for scalable investigations into the neural mechanisms underlying perceptual and mnemonic processes.