Arithmetic learning is associated with developmental increases in similarity between brain activity and artificial neural networks

The ability to learn simple arithmetic is often thought to rely on the associative nature of human memory, where repeated exposure to arithmetic problems strengthens the relations between operands and outcomes. This view is reminiscent of learning in large language artificial neural networks (ANNs), where training involves statistical associations between symbolic inputs and outputs. Based on this parallel, we hypothesized that brain activity during arithmetic problem-solving should be increasingly similar to ANN-derived representations with learning and development. We used fMRI to investigate the relation between brain activity during single-digit addition problem-solving and ANN features in 104 participants across four age groups (8-, 11-, 14-year-olds, and adults). Vertex-wise encoding models were used to predict brain activity using latent features extracted from pre-trained ANNs. Results revealed that the ANN-based model better predicted arithmetic-related activity in older than younger participants in a specific region of the left precentral sulcus. Prediction accuracy was higher for smaller than larger problems, consistent with the idea that associations between problems may decrease in strength with problem size. Furthermore, ANN prediction patterns aligned with a model in which each addition problem was increasingly represented separately with age. On the one hand, these findings may support the idea that, with learning and development, some neural representations of arithmetic problems become discretely organized in memory and increasingly resemble ANN-like processing. On the other hand, the lack of relation between the ANN-based model and activity in other brain regions suggests some degree of dissimilarity between arithmetic processing in humans and ANNs, suggesting the existence of additional mechanisms for simple arithmetic learning in humans.