Summary: New research shows that the brain contains neurons that track our position within a sequence of actions — not just our physical position in space. In mice, scientists discovered “goal-progress cells” in the cortex that fire according to how far an animal has progressed through a behavioral task. These cells enable animals to predict what comes next, even in situations they have never encountered before.
Rather than mapping physical environments, these neurons form internal maps of behavior, supporting flexible, generalized problem-solving. The finding sheds light on how brains support everyday inference and may inspire more adaptable approaches in artificial intelligence.
Key Facts:
- Goal-Progress Cells: Cortical neurons encode progress through behavioral sequences instead of physical location.
- True Generalization: Mice inferred task structure and predicted next steps without prior exposure to the exact configuration.
- Biological-AI Connection: These results help bridge how brains represent structure with how AI could gain flexible, transferable skills.
Source: The Conversation
For decades, neuroscientists have developed mathematical frameworks to explain how patterns of brain activity support predictable, repetitive behaviors, such as performing games or routines.
Those models have described neural activity with impressive precision and helped build artificial intelligence that excels at narrowly defined tasks like playing Atari games or Go.
However, these models struggle to capture a key feature of human and animal intelligence: our ability to generalize, infer structure, and adapt to novel situations. Our recent study, published in Nature, investigates how brain cells in mice support this kind of flexible behavior.
Humans and animals can solve new problems by generalizing from prior experience. We try new recipes, meet unfamiliar people, and find alternate routes — often predicting consequences of choices we have not previously faced.
The idea of internal representations that organize experience goes back to psychologist Edward Tolman’s concept of “cognitive maps.” From the 1970s onward, researchers found specialized neurons in the hippocampus and entorhinal cortex that form literal spatial maps: place cells fire at particular locations and grid cells tile space in a regular pattern.
Those spatial maps explain how animals navigate physical environments. More recently, attention has shifted toward how similar neural systems might support higher-level cognition — generalization, imagination, social reasoning and flexible memory — beyond pure navigation.
Cells for generalizing?
We asked whether neurons exist that organize knowledge about behavior itself rather than the external world, and if so, how they do it. What algorithms underlie the activity of brain cells when we generalize from past episodes to new ones?
Our experiments revealed such neurons. These cells indicate “where” an animal is within a behavioral sequence.
To identify them, we trained mice on a task made up of a repeating sequence of actions. The animals moved among four reward locations labeled A, B, C and D arranged in loops. Crucially, the precise physical locations of these goals were changed across sessions.
When goal positions shifted, mice immediately inferred what would come next in the sequence even when the specific configuration was novel. For example, upon reaching goal D placed in a new location, mice reliably returned to goal A on the very first trial — a behavior that could not be explained by simple recall of past locations. Instead, it reflects an understanding of the task’s abstract structure and the animal’s position within it.
We recorded neural activity with electrodes implanted in the cortex during task performance. We found that many cortical neurons tiled the progression toward goals: individual cells fired at particular fractions of the distance or steps to a goal, independent of the goal’s absolute position.
Some neurons signaled progress toward short-term subgoals, analogous to intermediate steps in a recipe, while others tracked progress toward the overall objective, like completing the meal. For example, a neuron might become active whenever the animal was roughly 70% of the way toward a given goal, regardless of how far away it was or where it lay.
Together these “goal-progress cells” formed an internal coordinate system for behavioral space. Importantly, the representation was flexible: neurons stretched or compressed their tuning to accommodate different goal distances, enabling instant generalization to new task layouts.
Why generalize?
Why would the brain develop generalized structural representations instead of building a separate memory for every situation? Because the behaviors we produce to reach goals are highly structured and repetitive. Generalization allows knowledge to transfer across tasks that share common patterns.
For example, experience making one tomato-based sauce helps when preparing another: many steps overlap, such as sautéing aromatics early and adding fresh herbs near the end. By encoding abstract relationships between events, actions and outcomes, goal-progress cells provide a compact framework that supports such transfer.
We propose these cortical cells act as the building blocks of behavioral schemata — internal frameworks that organize abstract sequences of steps. Although our experiments were performed in mice, the underlying principle likely applies across mammals, including humans.
Documenting these cellular networks and the algorithms they implement narrows the gap between animal and human neuroscience and helps translate biological insight into more flexible artificial intelligence approaches.
About this cognition and neuroscience research news
Author: Mohamady El-Gaby
Source: The Conversation
Contact: Mohamady El-Gaby – The Conversation
Image: The image is credited to Neuroscience News
Original Research: Open access.
“A cellular basis for mapping behavioural structure” by Mohamady El-Gaby et al. Nature
Abstract
A cellular basis for mapping behavioural structure
To adapt flexibly to new situations, brains must capture regularities both in the external world and within our own behavioral routines. While algorithms for mapping physical environments have become better understood, the biological mechanisms that map structured, goal-directed behaviors remained unclear.
This work reveals a neuronal implementation of an algorithm that encodes abstract behavioral structure and allows that structure to be transferred to new scenarios. Mice trained on multiple tasks that shared a common sequential organization but differed in goal locations discovered the underlying structure and made correct inferences on the first trial of novel tasks.
Most neurons recorded in the medial frontal cortex tiled progress toward goals, much like place cells tile physical space. These “goal-progress cells” generalized their tuning by stretching or compressing to fit different goal distances. A subset of cells was specifically tuned to fire with fixed lags relative to particular behavioral steps, acting as task-structured memory buffers.
Collectively, these dynamics instantaneously encoded the entire sequence of upcoming behavioral steps and computed appropriate actions at each moment. The same patterns reappeared during offline sleep, indicating that the neural dynamics mirrored the task’s abstract structure both on-task and off-line. These findings support the idea that complex behavioral schemata can arise by shaping progress-to-goal tuning into buffers that represent individual steps of a task.