Animals have the ability to control their bodies, which allows them to perform a variety of behaviors. How such control is carried out by the brain, however, remains unclear. Improving our understanding requires models that can relate regulatory principles to the structure of neural activity in animal behavior. To facilitate this, we created a ‘virtual rodent’, where an artificial neural network used a biomechanically realistic model of a mouse in a physics simulator. We used deep reinforcement learning to train a virtual agent to mimic the behavior of freely moving mice, thus allowing us to compare the neural activity recorded in real mice with the network activity of a virtual mouse imitating its behavior. We found that neural activity in the sensorimotor striatum and motor cortex was better predicted by the network activity of the virtual mouse than any aspects of the movement of the real mouse, corresponding to both regions that use dynamic forces. In addition, latent network variability predicted the structure of emotional variability across behaviors and provided robustness in a manner consistent with the minimal intervention goal of appropriate response control. These results show how realistic simulations of biomechanically realistic animals can help explain the structure of neural activity throughout behavior and relate it to theoretical principles of movement control.
Here is a new Nature article by Diego Aldarndo, et.al. Via @sebkrier.
Source link