From multiple neural cortical networks to motor mechanical behavior: the importance of inherent learning over separable space-time length scales

An important question in neuroscience is how different cortical areas bind during the planning and execution of voluntary, goal-directed behavior. Learning visually-guided reaches can provide important theoretical and experimental insights into this problem, particularly when combined with Brain-Machine Interface (BMI) and multi-electrode measurements over extended periods of time across multiple cortical regions. We exploit the force-field paradigm [1] that alters the arm dynamics of the subject to monitor the ensuing adaptive processes in order to understand across multiple regions the differences between a habitual reach and a reach that requires learning. We quantify the translation of movement plans into their physical implementation by studying the representation of time [2] in relation of its well documented separability from the spatial components of motion trajectories [3]. Previously the internal representation of environmentally-dependent forces on position and velocity was found to be time-invariant [2]. We aim at explaining this feature in relation to the motor system's plasticity [4] during closed loop BMI. To this end we followed the evolution of tuning, mean firing rate levels and spike-time statistics across separable cell classes simultaneously recorded in the pre-motor and motor cortical regions of rhesus macaques as they adapted to new movement dynamics imposed by an external mechanical device.

We find that (1) several stable spatio-temporal representations co-exist in a given cell which permits identification and selection of different motor programs to operate the external device, and (2) these multiple representations can be extracted from the multi-electrode neuron spike patterns reflecting various spatial re-parameterizations compatible with the ones imposed by the external mechanical device. A neural theoretical formulation in terms of a Hodgkin-Huxley excitatory and inhibitory neural ring network is used [5] to model multi-electrode spiking statistics, explicitly considering the separation of different motor dynamical times.