Google DeepMind and Harvard build virtual rat with AI ‘brain’

Researchers from Google DeepMind and Harvard University have built a virtual rodent powered by artificial intelligence to better understand how the brain controls movement.

The virtual rat is powered by an artificial neural network that mimics the neural activity of its real-life counterpart, giving researchers a chance to compare the two.

While animals have “exquisite” control of their bodies, allowing them to perform a diverse range of bahaviors, how the brain implements such control remains unclear, researchers said.

To get a better understanding of how the brain works, researchers trained the virtual rodent to mimic the whole-body movements of freely moving rats in a physics simulator, where an artificial neural network actuated a biomechanically realistic model of the rat.

“We then compared neural activity from the real rat’s brain to the activations of the virtual rodent’s artificial neural network when performing the same behaviors,” lead author Diego Aldorando said in a thread of posts on X.

“We found that the virtual rodent’s neural networks, which implement inverse dynamics models, were better predictors of neural activity than measurable features of movement, like the positions or velocities of the joints, or alternative control models,” Aldorando said.

With @Harvard, we built a ‘virtual rodent’ powered by AI to help us better understand how the brain controls movement. 🧠

With deep RL, it learned to operate a biomechanically accurate rat model - allowing us to compare real & virtual neural activity. → https://t.co/GaToq3AWTQ pic.twitter.com/HnWZLk2mcE

undefined Google DeepMind (@GoogleDeepMind) June 13, 2024

Researchers used deep reinforcement learning to train the virtual agent to imitate the behavior of freely moving rats, according to the paper published in Nature.

The results of the study demonstrated “how physical simulation of biomechanically realistic virtual animals can help interpret the structure of neural activity across behavior and relate it to theoretical principles of motor control,” the paper read.

According to Aldorando, their research approach can be applied in neuroscience and facilitate the study of aspects of neuromotor control that are hard to experimentally deduce. It could also be “instrumental in modeling the neural control of increasingly complex animal behavior.”

Rowan Cheung, founder of the Rundown AI newsletter, said the study could “massively open up new research with testing on AI animals and expand robotics.”