Pretraining of Embodied Recurrent Networks Bridges the Gap Between Artificial and Cortical Neural Activities

Task-driven recurrent neural networks (RNNs) have been widely employed as tools for investigating neural dynamics in neural motor control research by modeling the motor cortex. RNNs are often implicitly assumed to learn the underlying computational mechanisms in accordance with biological neural circuits. However, the brain network has a highly structured and specific network connectivity and during individual development the motor cortex has acquired a rich repertoire of behavioral primitives via continuous learning of body control. Considering that the task-driven RNNs are often initialized randomly and trained directly on the specific task, how much these models can truly reveal about the motor cortex is still a crucial question awaiting further research. In this study, we propose a method for modeling the motor cortex pretrained on single reaching skills. Specifically, we use an RNN, receiving sensory feedback and task inputs, as the controller to produce motor commands that drive a musculoskeletal arm model. This model can perform reaching movements along a mini-jerk trajectory between arbitrary points in the workspace, prior to training on specific tasks. The model pretrained on single-reach task has more similarity with real neural data both on a neural geometry and neural dynamics level in center-out (CO) and random target touch (RTT) tasks than models directly trained on these tasks. Surprisingly, we observed the opposite pattern in a double-reach (DR) task, in which two targets appeared simultaneously, rather than presenting the next target after the completion of the prior movement as in the RTT task. This suggests that sequential movements are planned as an integrated unit, and this capability may be implemented at the level of motor cortical circuits. In summary, our results suggest that endowing the network with capabilities beyond the immediate task demands—through more systematic training or other methods—can help better understand the dynamics of biological neural circuits.