Distinct neural architectures underlie motor skill acquisition and transfer in human sensorimotor cortex

Motor learning typically emerges from repetitive practice but can also arise through transfer or generalization, when improvements extend to untrained actions. However, it remains unclear whether training-based learning (training effect) and transfer/generalization-based learning (transfer effect) rely on shared neural mechanisms or whether they represent distinct processes supported by different cortical architectures. Here, we employed between-subject multivariate decoding of hyperaligned functional magnetic resonance imaging data to characterize the universal neural codes underlying the training and transfer effects and their resultant motor engrams in the sensorimotor cortex. We found that both learning processes are encoded bilaterally but shifted to the right (ipsilateral) hemisphere, with the training effect showing higher decoding accuracy than the transfer effect. The bilateral central sulci showed a clear mechanism-specificity towards the training effect, in contrast to the rest of the sensorimotor cortex. Interestingly, these learning processes consolidate two distinct motor engrams. Remarkably, the information structure of these processes is shared among individual brains, and the between-subject classification provided limited results when performed with standard, anatomical alignment. Our findings support the notion that motor skill learning is hierarchically organized in the sensorimotor cortex, and extend current models by revealing that training and transfer encode distinct and mechanism-dependent high-dimensional representations embedded in a neural code that is shared across brains.