Modeling the neural control of movement with a virtual rodent

Abstract

Brains control bodies to generate behavior. Nevertheless, there is a bias in motor neuroscience to overlook the body and focus on the relationship between neural activity and measurable features of behavior. While this approach has been invaluable in describing the neural representation of movement, it has largely failed to provide a comprehensive understanding of its neural control. Such an understanding requires models that causally generate animal behavior and explain the patterns of neural activity in behaving animals from principles of control theory. To probe this line of inquiry, we developed a 'virtual rodent', a biomechanical model of the rat actuated by artificial neural networks in a physics simulator, and evaluated its utility as a model of the neural control of movement. First, we studied the contextual organization of motor activity in artificial neural networks trained to perform several complex motor tasks with deep reinforcement learning. By analyzing the network population dynamics with respect to the virtual rodent's behavior, we found that the network's representation of movement qualitatively resembled observations from neural recordings in real animals and was organized into two classes encoding task-specific behavioral strategies and task-invariant motor dynamics. Second, we trained the virtual rodent to imitate the behavior of freely-moving rats, thus enabling the one-to-one comparison of neural activity recorded in real rats to the network activity of a virtual rodent performing the same behaviors. We found that the virtual rodent's network activity predicted the single-unit and population activity of the sensorimotor striatum and motor cortex better than any features of the real rat's movements, consistent with both regions implementing inverse dynamics. Moreover, we found the virtual rodent's latent variability was structured in a manner consistent with the minimal intervention principle of optimal feedback control and predicted the structure of neural variability across behaviors, suggesting that the brain may organize neural variability in accordance with this principle. Together, these results demonstrate the utility of integrating biomechanical modeling, physical simulation, and artificial control in studying the neural control of complex behavior.