Modeling Target Reaching in the Motor Cortex Using Recurrent Neural Networks

Abstract

The introduction of multielectrode arrays capable of recording from hundreds of neurons simultaneously in the motor cortex (and other areas) has energized the field of motor control neurophysiology. The information obtained with such recordings, however, presents challenges with interpreting the functional significance of the simultaneous activities of large numbers of neurons during voluntary movement. One effective means to address this problem is to apply a dynamical systems approach, wherein population activity is considered to traverse a state space with each axis of the space represented by the firing of a single neuron. Despite the high dimensionality, neural trajectories in the motor cortex typically occupy only a low-dimensional subspace. One consistent observation based on recordings in monkey motor cortex during reaching movements is that the subspaces during movement preparation and movement execution are different from one another. An unanswered (and largely unasked) question has to do with the necessity and purpose of these different subspaces. To address this question, I modeled neural activity in the motor cortex using a recurrent neural network (RNN). The inherent structure of RNNs is like that of recurrently connected motor cortical neurons. I trained two identical RNNs on a delayed reaching task to eight targets distributed in a circle (‘center-out-task’). The inputs were target location, timing of target presentation and go cues. Outputs were ‘hand’ kinematics representing movements to the targets. The two RNNs differed only in how target location was presented to the network; in one case the target signal was extinguished upon movement onset (‘Interrupted input RNN’), while in the other, the target signal remained on (‘Sustained input RNN’). In general, the activity patterns of individual neurons in the RNNs showed similarities to those recorded in the monkey motor cortex during this task. Interestingly, neural dynamics in the interrupted input RNN exhibited a similar orthogonalization of preparatory and movement related subspaces. However, for the sustained input RNN, preparatory and movement subspaces were practically the same. This finding suggests that in the biological system, gating of sensory feedback about target location occurs near the time of movement onset, leading to a different representation in neural state space.