Aligning Neural Population Patterns Facilitates Motor Learning Transfer

Motor learning transfer, the ability to apply skills acquired in one task to enhance performance in a related task, is driven by changes in neural ensemble activities. However, the long-term evolution of neural population dynamics during motor learning transfer remains unclear. Specifically, how do the neural patterns reorganize and stabilize over an extended learning period in the new task? To investigate the neural mechanisms in the motor cortex that enable such transfer, we employed a Brain-Machine Interface (BMI) paradigm in rats. In the experiment, rats first mastered a lever-pressing task before proceeding to a more complex but related lever-discrimination task. We analyzed neural ensemble activities by projecting them into a low-dimensional linear feature space that captures the most prominent dynamic structure. Within this space, we represented neural patterns as clusters and developed an iterative method to align similar clusters between the two tasks. Leveraging this alignment, we introduced a novel decoding approach, Clustering Alignment-based Transfer on Kernel Reinforcement Learning (CATKRL), which utilizes parameters learned from the initial task to enhance efficiency in the new task. Our results revealed that neural patterns for learned actions form distinct clusters with consistent shapes and centroid distances across tasks, and these patterns exhibit rotational evolution during learning. By integrating the cluster alignment mechanism into the RL decoder, we achieved a faster training speed with less data in the lever-discrimination task. Our results suggest that aligning neural pattern clusters can enhance BMI decoding efficiency by leveraging consistent neural representations. This approach not only provides valuable insights into the brain mechanisms underlying motor learning transfer but also holds promise for advancing multi-task learning in neuroprosthetics.