Harvard University to Use AI (Artificial Intelligence) Rats to Study Neuroscience

Figure | AI rats walk and feeding in the maze (Source: DeepMind & Harvard University)

In the study of artificial intelligence (AI), neural networks have become one of the core topics. Although the neural networks studied by AI are very different from the way neurons in the brain work, more and more scientists believe that the technology created by the brain structure is closely related to neuroscience. Studying the similarities and differences between the two can not only deepen our understanding of neuroscience but also make AI smarter. When studying neuroscience, rats are one of the common experimental objects, such as analyzing how the rat brain controls its movement. So can we study AI just like rats?

Researchers at DeepMind and Harvard University believe that this idea is feasible. They have created a virtual rat driven by AI that can perform multiple complex tasks in a simulated 3D environment. In studying how the so-called "AI brain" controls the movement of rats, neuroscience technology can come in handy. One of the authors of the paper, Jesse Marshall, a Harvard researcher, said that this research is like building a wind tunnel laboratory for neuroscience research, allowing researchers to test different neural networks with different degrees of biological authenticity. To understand how they respond to complex tasks and challenges.

Figure | Simulated rats and test environment (Source: DeepMind & Harvard University)

"Typical experiments in neuroscience research probe the animal ’s brain when performing simple tasks. Similarly, we want to understand how the brain generates and realizes flexibility, and then uses the results to design artificial agents (AI agents) with similar mechanisms " Jesse explained. The researchers said that they used a neural network to control the biological model of a rat in a simulated three-dimensional environment and then used neuroscience technology to analyze biological brain activity to better understand the mechanism of neural network control of rat movement.

The research results were published in the form of papers at the ongoing ICLR conference (International Representation Learning Conference). The ICLR conference project team believes that to use neuroscience techniques to understand the specific subjects of neural network control, this research adopts unusual ideas and provides a new research direction, which is exciting. The construction of virtual AI rats is based on real rats, and their muscles, joints and vision are all based on the measurement data of real rats. To make AI more realistic, virtual rats can also achieve proprioception, that is, muscle movement perception, a feedback system that informs the animal's body parts and their movement patterns.

To control the virtual rats, the researchers trained a neural network to guide them through four challenges: jumping through gaps, finding paths and feeding in the maze, escaping from the hilly environment, and racquets that meet the time interval requirements. After the rat completes the task, the research team will use neuroscience technology to analyze the recorded data of its "AI brain" neural activity to understand how the neural network achieves the motion control required to complete a specific task. In the real world, studying animal neural activity and linking it to specific behaviors is very complex, and most experiments are conducted in relatively strict experimental environments with relatively simple tasks. In the virtual environment, if you can simulate the rat's neural activity well and control it to complete multiple parts to form complex behaviors, such as foraging and racking, you can better match the neural activities and specific behaviors.

Since the AI ​​system that drives rats is self-developed, most of the analysis of the neural network operating mechanism is predictable and in line with expectations. But an interesting finding is that if nerve activity directly controls muscle strength and limb movement, then its duration seems to be longer than expected. This means that neural networks can express behavior on abstract scales such as running, jumping, rotating, and other intuitive behaviors. At the same time, neural networks also seem to have the ability to reuse certain representations of actions across tasks, and neural activities that encode behaviors often take the form of sequences. In other words, in the view of neural networks, although actions such as running and jumping need to coordinate multiple body parts and muscle tissues, once a specific movement pattern often appears, it can be abstracted into a specific movement method. For example, a rat (AI brain) finds that the hind limbs can jump up and down, then it will remember that this is a jump. If it needs to jump when performing other tasks, it will directly control its hind legs.

The researchers said that this abstract ability and cognitive model are generally believed to exist in animals, and have been observed in rodents and songbirds. The new findings further confirm the existence of this model, and this phenomenon naturally occurs in neural networks, and no reward mechanism is explicitly set. So while confirming the facts, the study also proved the similarity of neural networks and neuroscience. At present, researchers have open-sourced the virtual rat project, hoping that other teams can build on the existing model and test different neural network implementations in a virtual environment.

Blake Richards, a neuroscientist at McGill University in Canada who did not participate in the research, believes that although the full physiological authenticity is not simulated, the trained neural network still captures enough neural activity features that can be generated by neural activity. Influence making valuable predictions. "A major contribution of the paper is that it proposes a method of training neural networks by realistic means, which can more easily compare biological data and virtual neural network data," 

 See interesting videos of virtual IA rat performing different tasks

Virtual rodent -- neural dynamics (jPCA) during "two-tap" task for 3-layer policy

Virtual rodent -- neural dynamics (jPCA) during "forage" task for 3-layer policy

Virtual rodent -- neural activity stream during "gaps

Reference

https://spectrum.ieee.org/tech-talk/artificial-intelligence/machine-learning/ai-powered-rat-valuable-new-tool-neuroscience