Brain Decoding from a Deep Neural Network of Visual Memory Reveals the Maturation Trajectory of Memory Systems across Ages 3-7

Memories from the first four years of life are largely inaccessible by adulthood, yet these early experiences profoundly shape cognition throughout development. Understanding this phenomenon of infantile amnesia is critical for revealing the mechanisms of early memory development. Here, we developed a new content-based behavioral measure of memory that quantifies how closely children's memory patterns resemble those of adults, which we term the memory maturation score. Using a movie-watching fMRI dataset spanning age 3 to adulthood, we trained a cross-validated model on adults' movie-evoked brain activity to predict adult's propensity to remember each movie snippet, estimated by ResMem, a deep neural network trained to predict visual memorability. We then applied the model to each child's brain activity while viewing the same movie to assess their memory maturation (i.e., cross-decodability from adult patterns). Children exhibited adult-like neural patterns of memory around age 4, extending prior behavioral findings. Structural analyses showed that the volume of the left posterior hippocampus, but not age or total brain volume, significantly predicted children's memory maturation. Memory maturation around age 4 thus likely reflects age-related neural development of hippocampal memory systems, rather than an age increase itself. Functional searchlight analyses revealed a trajectory of memory maturation: the inferior temporal gyrus reflected mature memory-related processing from age 3 onward, the left posterior hippocampus by age 5, and higher-order regions involved in social-affective processing and multimodal integration by age 7. Together, these findings suggest that hippocampal maturation between ages 3 and 5 drives the maturation of memory patterns.