Researchers at Harvard University have teamed up with scientists from Google DeepMind to develop an artificial brain for a virtual rat. Nature, This revolutionary advance opens new doors to studying how the brain controls complex movements using advanced AI simulation techniques.
Construction of a virtual rat brain
To build the virtual rat brain, the research team used high-resolution data recorded from a real rat. Harvard researchers worked closely with the DeepMind team to build a biomechanically realistic digital model of the rat. Graduate student Diego Aldarondo worked with DeepMind researchers to use deep reinforcement learning, a powerful machine learning technique, to train an artificial neural network (ANN) to act as the virtual brain.
The neural networks were trained to use inverse dynamics models that are thought to be used by the human brain to guide movement. These models allow the brain to calculate the required trajectory and translate it into motor commands to achieve a desired movement, such as reaching for a coffee cup. Using reference trajectories derived from data from real rats, the virtual rat's neural network learned to generate the forces needed to produce a variety of movements, including movements that were not explicitly trained on.
“DeepMind was developing a pipeline for training biomechanical agents to move around complex environments. We simply didn't have the resources to run those simulations and train the networks,” Orbetsky noted. The collaboration was “incredible,” he added, highlighting the critical role DeepMind scientists played in making this groundbreaking achievement possible.
The result is a virtual brain that can control a biomechanically realistic 3D rat model within an advanced physics simulator, closely mimicking the movements of real rodents.
Potential uses
Virtual rats with artificial brains offer a new approach to investigating the neural circuits responsible for complex behaviors: by studying how an AI-generated brain controls the movements of a virtual rat, neuroscientists can gain valuable insights into the complex workings of a real brain.
This breakthrough could also pave the way for the design of more advanced robotic control systems. “Our lab is interested in fundamental questions about how the brain works, and this platform, as an example, can be used to design better robotic control systems,” Orbetsky said. Understanding how the virtual brain produces complex behaviors could enable researchers to develop more sophisticated, adaptive robots.
Perhaps most intriguingly, this research could give rise to the new field of “virtual neuroscience,” in which AI-simulated animals serve as convenient, fully transparent models for studying the brain even in diseased states. These simulations could provide an unprecedented window into the neural mechanisms behind a range of neurological disorders and lead to new therapeutic strategies.
The next step: More autonomy for virtual rats
Building on this groundbreaking work, the researchers plan to give virtual rats even more autonomy to solve similar challenges faced by real rats. Orbetsky explains: “Our experiments have given us a number of ideas about how such challenges might be solved and how the learning algorithms underlying the acquisition of skilled behaviors might be implemented.”
Giving virtual rats additional independence will allow scientists to test theories about learning algorithms that enable them to acquire new skills, which could provide valuable insight into how the real brain learns and adapts to new challenges.
The ultimate goal is to better understand how the real brain produces complex behaviors. “We hope to test these ideas in virtual rats to better understand how the real brain produces complex behaviors,” Orbetsky says. By further refining and expanding this innovative approach, neuroscientists and AI researchers can work together to unlock the mysteries of the brain and create more intelligent and adaptable systems.
