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

Area of Science:

• Cognitive Science
• Neuroscience
• Computational Neuroscience

Background:

• Motor skill acquisition relies on practice structure, not just volume.
• Interleaved practice (IP) shows superior long-term retention in humans compared to repetitive practice (RP).
• The computational underpinnings of these practice structure effects on motor learning are not fully understood.

Purpose of the Study:

• To investigate the computational basis of how interleaved practice (IP) versus repetitive practice (RP) structures impact motor sequence learning.
• To determine if simple computational models can explain the benefits of IP in motor learning.

Main Methods:

• Implemented a minimal recurrent neural network (Elman network) to simulate motor sequence learning.
• Trained the network on sequential tasks using either RP or IP training structures.
• Compared network performance and generalization capabilities between the two training structures.

Main Results:

• Repetitive practice (RP) resulted in faster initial error reduction during training.
• Interleaved practice (IP) yielded superior performance on trained sequences and better generalization to novel sequences.
• The benefits of IP were observed solely from the interaction of input variability and temporal recurrence, without complex mechanisms.

Conclusions:

• The benefits of interleaved practice (IP) in motor learning can be explained by basic computational principles involving input variability and temporal recurrence.
• Variability in practice contexts promotes the development of more robust and generalizable internal representations.
• This study offers a parsimonious computational account for IP benefits, informing motor learning theories and rehabilitation strategies.