A minimal recurrent neural network models the robustness of interleaved practice on motor sequence learning

Effective motor skill learning depends critically on the structure of practice, not solely on practice volume. This study explores the differential learning outcomes when multiple procedural skills are acquired simultaneously through interleaved practice (IP) versus repetitive practice (RP) structures. In humans, IP yields superior long-term retention compared to RP, which often leads to significant forgetting. To explore the computational basis of how distinct practice structures impact motor sequence learning, we implemented a minimal recurrent network-the Elman network-to perform a sequential motor task. The network was trained on sequences presented in either RP or IP order. While RP led to faster error reduction during training, IP produced both superior trained set performance and better generalization to novel sequences. These findings demonstrate that the benefits of IP emerge from the interaction between input variability and basic temporal recurrence alone, without requiring complex biological plasticity mechanisms or specialized contextual processes. Our results suggest a shared computational principle linking human motor learning with catastrophic forgetting in artificial neural networks: variability across practice contexts forces the formation of more robust and generalizable internal representations. This work offers a parsimonious account of IP benefits and informs both theoretical models of learning and strategies for optimizing motor rehabilitation.