Granule cells steer cortical activity along distinct paths for different contexts

A new study in mice has revealed how the brain may solve one of learning’s most difficult problems: recognizing that two situations are different while still applying knowledge that they share. The research identifies a division of labor between the cerebral cortex and the cerebellum, suggesting that the cortex preserves broad, reusable patterns of activity while cerebellar granule cells transform those patterns into context-specific signals. The finding challenges a long-standing expectation that the cerebellum separates experiences by dramatically expanding neural representations into a high-dimensional code. Instead, the study reports that granule-cell activity remains comparatively compact, but its low-dimensional trajectories rotate into different orientations depending on the task. This geometric reorganization could allow animals to generalize familiar strategies without confusing the circumstances in which each strategy is appropriate.

Learning almost always involves this tension between generalization and separation. An animal that encounters a new situation must identify similarities with previous experiences; otherwise, every event would require learning from scratch. At the same time, treating related situations as identical can produce costly mistakes. A movement that is successful in one environment may fail in another, and the same sequence of actions may require a different outcome when the surrounding cues or goals change. Neuroscientists have proposed that the neocortex supports generalization by organizing activity on low-dimensional neural manifolds. Although thousands or millions of neurons may be active, their collective activity can be constrained to a smaller number of task-relevant dimensions. This compression makes learning more efficient because the brain can reuse dynamic patterns instead of constructing a completely new representation for every context.

The cerebellum has often been associated with the opposite computational strategy. Its granule cells, the most numerous neurons in the brain, receive combinations of signals from mossy fibers carrying information about movement, sensation and internal state. Classic theories proposed that the enormous number of granule cells and their sparse, diverse inputs create an expansion layer. In this arrangement, relatively similar input patterns are distributed across a much larger population of neurons, projecting them into a high-dimensional feature space. Such expansion could make overlapping contexts easier to distinguish, particularly when the cerebellum must select among competing motor commands. The new study directly examined whether this presumed expansion is how cerebellar granule cells separate contexts, or whether they use a subtler mechanism that preserves some of the brain’s underlying structure.

Garcia-Garcia and colleagues simultaneously imaged activity in two key components of the cortico-cerebellar pathway while mice learned two distinct skills. The tasks shared a temporal structure, meaning that the animals had to organize behavior according to a similar sequence or timing pattern, but the skills represented different contexts. This design allowed the researchers to test whether the brain would preserve a common dynamic framework while modifying its meaning according to the task. They focused on layer 5 pyramidal tract neurons in the premotor cortex and cerebellar granule cells. Layer 5 pyramidal tract neurons are major cortical output cells that can convey movement-related commands and other signals toward subcortical targets, including pathways that influence the cerebellum. By observing both populations at once, the researchers could compare the cortical patterns entering the broader circuit with the cerebellar patterns generated downstream.

The measurements produced a result that was surprising in its precision. Granule cells did not abandon low-rank, low-dimensional encoding when the mice switched between tasks. Instead of spreading activity into an entirely expanded and unrelated set of dimensions, the granule-cell population retained a compact structure. The key difference appeared in how that structure was organized over time. Cortical activity patterns generalized across the two contexts: the trajectories traced by the neural population retained similar geometry, reflecting shared temporal dynamics and common elements of the learned skills. Granule-cell activity, by contrast, underwent a temporal remapping. The same broad sequence of population states was reoriented so that the trajectory associated with one task pointed in a different direction from the trajectory associated with the other.

The researchers describe this effect as a rotation of low-dimensional neural trajectories. A neural trajectory is a way of representing how the combined activity of a population changes over time. Each moment in a behavior corresponds to a point in a multidimensional space, and the succession of points forms a path. If two tasks generate similarly shaped paths, the population may be preserving a shared computational pattern. If those paths are oriented differently, the system may be distinguishing the tasks without discarding their common structure. In the study, granule-cell trajectories separated between contexts through coherent reorientation rather than random scrambling. This distinction matters: random remapping would make each context unique but could also destroy the temporal relationships needed for smooth control. A coordinated rotation instead preserves the internal geometry of the cortical pattern while changing how that pattern is read by the cerebellar circuitry.

The finding also bears on the relationship between cortical and cerebellar activity. Despite the remapping in granule cells, cortico-cerebellar coupling remained stable. In other words, the communication link between the cortical and cerebellar nodes did not appear to break down when the mice changed tasks. The cortex continued to provide an invariant dynamic primitive—a reusable pattern that captures the timing or organization of the behavior—while the cerebellar representation adjusted its orientation. This arrangement could allow the cerebellum to act as a context-sensitive transformation layer. It would not need to relearn the entire temporal structure of a skill. Instead, it could preserve the cortical dynamics and apply a context-dependent reconfiguration that helps generate an appropriate output.

The degree of separation increased as the mice became more expert. Granule-cell trajectories diverged most strongly in animals that had acquired greater proficiency, suggesting that context separation is not merely an early consequence of learning or an unavoidable response to novelty. It may be refined with practice. As the animals gain experience, the cerebellar system could progressively sharpen the angular difference between task-specific trajectories, making it easier to select the correct output while retaining the shared features that support efficient performance. This pattern offers a possible neural explanation for why expertise often produces both flexibility and precision. An experienced animal can recognize the common structure of related tasks, yet rapidly distinguish which learned response should be deployed in the present setting.

The study’s proposed architectural division of labor could help reconcile two influential views of brain computation. One view emphasizes cortical representations that are stable enough to support abstraction, transfer and generalization. The other emphasizes cerebellar mechanisms that differentiate similar inputs and calibrate highly specific motor outputs. The results suggest that these functions need not be implemented by entirely separate, incompatible codes. A low-dimensional cortical trajectory can remain stable across contexts, while a downstream cerebellar population reorients that trajectory into distinct task-dependent channels. The cerebellum therefore may separate contexts not by exploding every input into an enormous unstructured feature space, but by preserving a compact representation and changing its orientation in a controlled way. Such a mechanism could be computationally efficient because it combines reuse with selectivity.

Although the experiments were conducted in mice and focused on learned skills with shared temporal structure, the principle may extend beyond the particular behaviors tested. Human learning also depends on recognizing recurring patterns while adapting them to changing environments, and the cerebellum is increasingly understood as contributing to cognition as well as movement. The work raises questions about how trajectory rotations are generated, which synaptic mechanisms stabilize them, and how they influence downstream cerebellar nuclei and behavior. It also suggests that disorders involving motor adaptation or context-dependent control might arise not only from disrupted connectivity, but from faulty geometric transformations within population activity. By showing that neural representations can separate experiences through coherent reorientation rather than wholesale expansion, the study offers a new way to think about how brains remain both economical and remarkably flexible.