Dynamical Neural Computation in Predictive Sensorimotor Control

How does the motor cortex function in predictive sensorimotor control? This is important for understanding neural control of movement, but dauting for the tangling of high-dimensional sensory and motor information. To address this question, we recorded neural population activity from the motor cortex while the monkeys were performing a manual interception task. Interestingly, on the neural states of single trials, we observed a low-dimensional ring-like neural geometry. This geometry, featured with ordered reach-direction clusters and tilted target-speed rings, also emerges in a three-layer RNN with appropriate inputs. Such a standard RNN, however, falls short of the capability in disentangling the sensorimotor interaction due to its homogenous hidden-units. Therefore, to scrutinize the interplaying process, we build another modular RNN to induce the neuronal tuning properties that may be relatively more sensory or motor. This RNN consists of two modules in the hidden-layer, one with adjustable connection weights and another with fixed ones. By making only the latter connected with the output nodes, these two modules work as ‘planning module’ and ‘execution module’, correspondingly. It turns out that the geometry of states from different modules differs as expected: the target-speed rings remain titled in the ‘planning module’ but go overlapped in the ‘execution module’, implying a divergence of involvement of sensory information. In addition, this model enables interception trials with the same endpoint and reversible inactivation, which so far can hardly be done in animals. The simulation results of short-term blocking target-location input suggest that sensory information to the ‘execution module’ is also necessary for accurate behavior. The present work reveals a specific geometric structure in neural space, and becomes an instance where models keep abreast with experiment: while the network’s reproduction supports the motor cortex as a dynamical system, the refined manipulation within network modules might provide further mechanical insights.