How neural activity controls naturalistic behavior is a fundamental question in neuroscience. Using deep reinforcement learning to implement inverse dynamics models, Aldarondo et al. trained artificial neural networks that control a biomechanically realistic rat model in a physics simulation engine to imitate the movements of real, freely moving rats. They then related neural activity recorded from the dorsolateral striatum and motor cortex of real rats to the network activity of the ‘virtual rodent’ performing the same natural behaviors. Intriguingly, the structure of real neural activity was predicted more accurately by the virtual network’s activity than by any feature of the real rats’ movements. The authors showed that the network’s latent variability shapes action variability to attain robust behavioral control and predicts the structure of neural variability across behaviors. These findings may contribute to our understanding of how the brain computationally controls behavior.

Original reference: Nature https://doi.org/10.1038/s41586-024-07633-4 (2024).