Log into Nature, how does the brain control movement? DeepMind designs virtual animals with “AI brains”

Editor | Radish Skin

Animals have very subtle control over their bodies, so they can perform a variety of behaviors. However, how the brain achieves this control remains unclear. To deepen our understanding, we need models that can link control principles to the structure of neural activity in animals.

To achieve this, researchers from Harvard University and Google DeepMind built a "virtual drunk animal" that uses artificial neural networks to drive a biomechanical simulation model of rats in a physical simulator.

The team used deep reinforcement learning to train virtual agents to imitate the behavior of freely moving mice, allowing the researchers to compare "recorded neural activity from real mice" with "virtual agent behavior" models that simulated their behavior Compare. This "network activity of virtual agent animals" can be used to explore the brain's learning and reasoning processes, thereby increasing understanding of these processes. Additionally, the team's deep learning models could help develop smarter robots and other autonomous systems.

The model was able to accurately imitate the movements of real mice, a significant achievement that is expected to improve scientists' understanding of how the brain controls complex coordinated movements.

It is difficult for the most advanced robots to replicate this result, but the research team believes that their discovery can greatly improve the flexibility of future robots.

The research was titled "A virtual rodent predicts the structure of neural activity across behaviors" and was published in "Nature" on June 11, 2024.

#Humans and animals control their bodies with ease and efficiency that is difficult for engineered systems to imitate. This is caused by computational simulation, the technical aspect of sports neuroscience. The reason is that, relative to models of causal production of complex, natural movements, neural activity in motor systems rarely has clear explanations.

These models of biogenesis differ in that neuroscientists attempt to infer motor system function by linking neural activity in relevant brain regions to measurable movement characteristics, such as the kinematics and dynamics of different body parts.

This approach is problematic, however, because the laws of physics inherently relate motion characteristics and therefore can only describe behavior rather than generate it. To solve this problem, the research team proposed a new approach: using virtual animal models associated with control models to infer computational principles.

Illustration: Using MIMIC to compare biological and artificial controls within the behavioral spectrum. (Source: paper)

The research team developed a "virtual rodent" in which an artificial neural network (ANN) drives a biomechanically realistic rat operating in a physical simulator Model.

When building this system, a balance needs to be struck between tractability, expressiveness, and biological realism. The researchers chose the simplest model that could reproduce the mice's behavior and predict neural activity.

The model uses deep reinforcement learning to train ANN to implement the inverse dynamics model. The input is the future movement reference trajectory and current body state of the real animal, and the output is the actions required to achieve the desired state. Researchers can compare the neural activity of real rats with the activity of virtual rodent networks based on related data.

This approach has two main advantages: First, the model is causal and can physically reproduce the behavior of interest, not just describe it. The second is to focus on identifying the functions implemented by brain areas rather than just descriptions of information flow.

“We learned a lot from the challenge of building ‘embodied agents’: AI systems must not only think intelligently, but they must also translate that thinking into practical actions in complex environments.” Matthew from Google Deepmind "Taking the same approach in a neuroscience context appears to provide insights into behavior and brain function," Botvinick said. The results showed that neural activity in the sensorimotor striatum and motor cortex was altered by virtual rodents. of network activity is more accurately predicted, consistent with these two regions achieving inverse dynamical control.

Illustration: virtual mouse. (Source: Deepmind website)

Furthermore, underlying changes in the network predict the structure of neural changes across behaviors and confer system robustness in a manner consistent with the minimal intervention principle of optimal feedback control.

These findings reveal that biomechanically realistic virtual animals through physical simulations can help explain the structure of neural activity across behaviors and link this to theoretical principles of motor control.

Furthermore, this approach demonstrates the potential for artificial controllers to manipulate biomechanical models to reveal computational principles of neural circuits. Virtual animals can serve as a platform for virtual neuroscience to simulate the effects of variables that are difficult to infer in experiments on neural activity and behavior.

This area of ​​research is critical to the development of advanced prosthetics and brain-computer interfaces. By reconstructing neural circuits, insights gained from this study could lead to new ways to treat movement disorders. Additionally, the study noted that virtual rats provide a transparent model for studying neural circuits and the impact of disease on these circuits.

Next, the researchers plan to let the virtual mice autonomously solve tasks encountered by real mice to further deepen their understanding of the brain's skill acquisition algorithms.

In the future, scientists may build brain-inspired network architectures to improve performance and interpretability, and explore the role of specific circuit structures and neural mechanisms in behavioral computation.

Paper link: https://www.nature.com/articles/s41586-024-07633-4

Related reports: https:// decrypt.co/235086/virtual-rat-ai-brain-harvard-google-deepmind-robotics

The above is the detailed content of Log into Nature, how does the brain control movement? DeepMind designs virtual animals with “AI brains”. For more information, please follow other related articles on the PHP Chinese website!