The limbic navigation system as a hierarchical RNN
This paper introduces a hierarchical recurrent neural network that, through path-integration training, successfully reproduces key neural representations, circuit motifs, and dynamic phenomena of the brain's limbic navigation system, offering a biologically interpretable framework for linking neural mechanisms to spatial behavior.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
To understand how an animal finds its way home, scientists look to a specialized network inside the brain known as the limbic navigation system. This system acts as an internal compass and map, allowing creatures to track their position and orientation without needing to see landmarks. It relies on distinct types of nerve cells that fire in specific patterns: some signal which way the animal is facing, others measure how fast it is moving, and a third group creates a grid-like pattern that marks location in space. For decades, researchers have tried to understand how these different signals are processed and combined. While some models have successfully recreated the firing patterns of individual cells, they often treat the brain as a single, flat machine, missing the fact that the real brain processes direction and speed in separate, connected stages.
A team of researchers at University College London has built a new kind of computer model to bridge this gap. They created a hierarchical recurrent neural network, a system designed to mimic the brain's layered structure. Instead of feeding the computer a single, pre-packaged speed and direction signal, they gave it two separate inputs: one for how fast the animal is turning and another for how fast it is moving forward. The model is organized into four distinct layers that pass information along in a specific order, much like a factory assembly line where raw materials are transformed step-by-step into a finished product. The first layers handle the turning signals, while the later layers combine those directions with speed to calculate the animal's exact location in a two-dimensional space.
When the researchers trained this system to perform the task of path integration—keeping track of where it is based solely on movement—it learned to do more than just solve the math problem. The network spontaneously developed internal representations that looked strikingly similar to the real biological cells found in the brains of rats and mice. The first layers of the model produced cells that fired when the animal turned left or right, mirroring angular velocity cells. As the signal moved to the second layer, the cells became more focused on direction alone, resembling head direction cells. In the deeper layers, the network generated cells that fired in a grid pattern, just like the famous grid cells found in the entorhinal cortex of the brain. Crucially, the model also produced cells that combined direction and location, a complex type of cell observed in nature but rarely seen in simpler computer models.
The researchers did not force the model to look like the brain; they simply let it learn the task. Yet, the internal connections it formed matched known biological patterns. The network developed a ring-like structure where activity could flow smoothly in a circle, a feature thought to be essential for maintaining a stable sense of direction. It also created specific, asymmetric connections between cells that allowed the system to shift its internal map correctly when the animal turned. When the scientists tested the model's robustness, they found it could navigate for much longer distances than it had ever seen during training, suggesting it had learned a general principle of movement rather than just memorizing specific paths.
To see how these parts work together, the team performed virtual experiments by disabling specific groups of cells within the model. When they silenced the cells responsible for detecting leftward turns, the model's internal sense of direction began to drift continuously to the right. When they cut the connections between the direction-sensing layers and the location-sensing layers, the grid-like patterns that mark location began to break down and lose their shape. These results mirror real-world experiments where damaging specific brain areas in animals leads to similar disorientation, confirming that the model has captured the causal relationships between different cell types.
In a final, sophisticated test, the researchers introduced a rhythmic pulse to the model, mimicking the brain's natural theta waves, which are slow electrical oscillations that occur during movement. By having two groups of turning cells take turns inhibiting each other in rhythm with these waves, the model began to sweep its internal map back and forth. This produced a "theta sweep," a phenomenon where the brain briefly simulates locations to the left and right of the animal's current position in rapid succession. This behavior, recently observed in real animals, emerged naturally in the simulation without the researchers explicitly programming it to happen.
The study suggests that the brain's complex navigation system is not a random collection of parts but a highly organized hierarchy where direction and speed are processed separately before being combined. By building a model that respects this structure, the researchers have created a digital twin that can reproduce the behavior of real neurons and the circuits that connect them. This approach offers a powerful new way to test ideas about how the brain works, allowing scientists to run virtual experiments that would be difficult or impossible to perform on living animals. The work indicates that the specific architecture of the brain's navigation system is not just a biological accident, but a necessary solution to the problem of tracking position in a moving world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.