Structural generalization and continual learning enabled by factorized entorhinal-hippocampal memory and entorhinal-parietal action circuits

To navigate the world, the human brain performs a feat of engineering that remains unmatched by standard recurrent neural networks on specific navigation tasks. It must remember specific places—a favorite café, a childhood home—while simultaneously understanding the abstract rules of how space works, allowing us to apply that knowledge to a new city we have never visited. This ability relies on two distinct but connected systems in the brain. One system, centered in the hippocampus, acts like a high-capacity filing cabinet for specific events and locations, storing the unique details of what we have seen. The other system, involving the entorhinal cortex, provides a universal grid, a mental coordinate system that measures distance and direction regardless of what is actually being looked at. While scientists have long known these systems exist, it has remained a mystery how they work together to let us learn quickly, generalize to new situations, and remember everything without overwriting old memories.

A team of researchers at MIT and elsewhere has now peeled back the layers of this mystery by testing how humans, monkeys, and computer models handle a specific navigation challenge. They designed a task where subjects had to move through a sequence of images, like stepping stones, from a starting picture to a target picture. The researchers tested three levels of difficulty. First, they asked subjects to navigate with visual clues, then without them, forcing the brain to simulate the journey internally. Second, they tested if subjects could navigate between picture pairs they had never practiced before. Third, they introduced entirely new sets of images to see if the brain could transfer its navigation skills to a completely new environment. The results were stark: while humans and monkeys mastered all three challenges with ease, standard recurrent neural network models failed completely. However, when the researchers built a new kind of computer model that mimicked the brain's specific architecture—separating the "map" from the "memory"—it succeeded where the others failed.

The experiment began with a simple setup. Human volunteers and two monkeys sat before a screen displaying a fixed line of images, such as a series of distinct objects. They were given a starting image and a target image and had to use a joystick to move through the sequence until they reached the goal. In the first phase, the images scrolled by, providing constant visual feedback. Once the subjects learned the order of the images, the researchers removed the visual clues. The subjects now had to move the joystick while the screen went dark, relying entirely on their internal sense of how far they had traveled. This is known as mental navigation. The subjects performed this effortlessly, showing they had built a mental map of the sequence. Next, the researchers tested transitive generalization by asking subjects to navigate between start and end points they had never practiced together. Finally, they introduced a third environment with a completely different set of images to test structural generalization.

The performance of the biological subjects was remarkable. Humans and monkeys not only navigated the new routes and new environments with high accuracy, but they also retained their ability to navigate the original environment without forgetting it. This lack of "catastrophic forgetting" is a critical feature of biological learning that artificial systems struggle to replicate. When the researchers tried to replicate this behavior using standard recurrent neural networks, these models failed at every test. They could learn the visual task with feedback, but the moment the visual clues were removed, or the route changed, or the environment was new, the models collapsed. They could not generalize, nor could they remember old lessons while learning new ones.

The breakthrough came when the researchers constructed a model that mirrored the brain's proposed structure. This model, which they call the Vector-HaSH augmented Action policy, consists of two main parts working in tandem. The first part is a structured memory scaffold based on grid cells. Think of this as a rigid, content-independent ruler that measures distance and direction. It does not care what the images are; it only cares about the space between them. The second part is a policy network that learns how to move based on the distance measured by the ruler. Because the ruler is fixed and universal, the policy learned in one environment can be applied immediately to a new one. The model binds new images to this ruler using a fast, one-time learning process, much like the brain's hippocampus.

When this new model was tested, it reproduced the full range of human and monkey behavior. It navigated without visual feedback, took novel routes, and adapted to new environments instantly, all while remembering the old ones. The researchers then looked inside the model to see how it worked and compared its internal activity to the actual electrical signals recorded from the brains of the monkeys performing the task. They found a precise match. The model's "distance" module, which tracked how far the subject was from the goal, behaved exactly like neurons in the monkey's entorhinal cortex. The model's "action" module, which decided when to move and when to stop, matched the activity in the posterior parietal cortex. This confirmed that the brain likely splits the task of navigation into two distinct computations: one that measures the distance and another that decides the action.

One final discovery revealed how the brain calibrates this system. The researchers found that for the model to work perfectly, it had to learn a specific scaling factor to align its internal ruler with the spacing of the images. If the ruler's ticks were too far apart or too close together relative to the images, the model struggled. However, when the model was allowed to adjust this scale factor during learning, it quickly found the perfect alignment, just as the brain does. This suggests that the brain actively tunes its internal sense of speed and distance to match the structure of the environment it is exploring.

The study concludes that the brain's ability to learn, generalize, and remember is not a result of having more data or more processing power, but of a specific architectural design. By separating the invariant structure of space from the specific content of experience, the brain creates a system that is both flexible and stable. The entorhinal-hippocampal circuit provides a universal metric that allows the brain to apply past lessons to new situations without erasing the past. This factorization of memory and action explains how we can walk into a new city, find our way around, and still remember the way home, a feat that standard recurrent neural network models have yet to achieve.