Think of the roar of a city as heard from a high floor: sirens, traffic, construction, voices, wind. The brain faces a similar flood of information every waking moment, but it must do something far harder than just hearing the noise. It must keep signals straight. A flash of light and the sensation of your own feet hitting the ground arrive in the same cortical region at the same time. How does the brain keep what you see distinct from the fact that you moved?
A team of neuroscientists at Albert Einstein College of Medicine, McGill University, and the Icahn School of Medicine at Mount Sinai has found a surprisingly elegant answer. In research published July 21 in Nature Communications, the group, led by co-first authors Eliezyer Fermino de Oliveira and Soyoun Kim alongside senior authors Lucas Sjulson, Renata Batista-Brito, and Adrien Peyrache, showed that the primary visual cortex (V1) uses an invisible geometry built into its own circuitry to both mix and separate movement and visual information as needed. The key to seeing this hidden structure was something unexpected: letting the mice sleep.
The hidden shape of neural activity
Neuroscientists have long known that the brain does not use all of its possible activity patterns equally. Out of the astronomically large number of ways a population of neurons could fire, only a tiny fraction actually occurs. The allowed patterns form a low-dimensional space called the intrinsic manifold. Think of a marble rolling on a curved surface: gravity and the shape of the terrain constrain where it can go, even though the marble could theoretically move in any direction. The manifold is that terrain for neural activity.
What the team realized is that during non-REM sleep, when the mouse is not seeing or moving, the brain keeps generating its intrinsic dynamics. Those dynamics trace out the manifold like a topographer walking the land. By recording from deep layers of V1 during NREM sleep and comparing those patterns with activity during wakefulness, the researchers could map the terrain and then ask where different kinds of information land on it.
They identified three kinds of spaces within the activity. The “on-manifold” subspace contains patterns that the brain readily generates during sleep. The “off-manifold” subspace consists of patterns that the intrinsic dynamics actively suppress, states the brain avoids during sleep. A third “unstructured” subspace holds dimensions not systematically related to the intrinsic dynamics.
This three-way split turned out to be the key to the mystery.
Where movement and vision meet, and where they part
When the mice ran, whisked their faces, or navigated their environment, the resulting neural activity landed squarely on the manifold. Movement representations are on-manifold signals. They follow the grain of the brain’s intrinsic dynamics. Visual stimuli, natural images shown to head-fixed mice, also landed on the manifold, where they mixed with movement signals in a straightforward, additive way. This is the integration side of the story. On the manifold, vision and movement coexist in the same space, their signals adding constructively.
But that raised a problem. If movement and vision both live on the same low-dimensional manifold, how does the brain keep movement noise from corrupting visual perception? Imagine two conversations happening in the same room on the same frequency. They will interfere.
The answer lies in the off-manifold subspace. The team discovered that stimulus representations also reach into the off-manifold, dimensions that the intrinsic dynamics suppress during sleep. Movement signals, crucially, do not. This means that visual information has a second, protected channel where movement-related activity cannot follow. The brain does not need to choose between integration and segregation. It achieves both by using different subspaces for different purposes.
Chorister neurons and sparse coding
The off-manifold coding turned out to rely on a specific class of cells. Neurons in V1 can be sorted along a spectrum from “soloists” to “choristers.” Soloist neurons fire in patterns that are relatively uncorrelated with the rest of the population, each singing its own tune. Chorister neurons, by contrast, are strongly modulated by the low-dimensional dynamics; their activity rises and falls in synchrony with the manifold.
It is the chorister neurons that carry off-manifold stimulus information. When a visual stimulus appears, it drives sparse, population-wide activity in these chorister cells. The finding is counterintuitive. Chorister neurons are the ones most strongly influenced by movement, yet they are also the ones that encode pure visual information in a way that avoids movement interference.
The paradox resolves through the geometry of the manifold. Because chorister neurons are so tightly constrained by the intrinsic dynamics, their spontaneous activity during sleep stays within a narrow region of state space. This leaves large swaths of unused territory, the off-manifold dimensions, that the brain can reach with sparse, stimulus-driven firing. Soloist neurons, by contrast, roam more freely across state space, so there is no unused protected region for stimulus information to hide in.
The result is a system where the very cells most coupled to movement create a clean, quiet space for vision precisely because of their tight constraints. The constraints that seem limiting are actually enabling.
Why it matters
This study resolves a long-standing tension in systems neuroscience. On one hand, brain-wide recordings have revealed low-dimensional representations of movement and internal states that seem to dominate neural activity everywhere, including sensory areas. On the other hand, sensory cortices are known to encode high-dimensional information about the world. These facts seem contradictory. How can V1 be both low-dimensional and high-dimensional at the same time?
The answer is that it uses different subspaces for different tasks. The low-dimensional manifold carries movement and integrated visual-movement signals. The off-manifold dimensions, actively suppressed during rest, provide a high-dimensional protected workspace for pure sensory coding. The brain does not flatten itself into a single mode of representation. It layers modes on top of each other, separated by the geometry of its own intrinsic dynamics.
The work also illustrates the value of sleep as a scientific tool. By recording during NREM, the researchers could isolate the brain’s intrinsic dynamics without the confounding influence of sensory input or behavior. The manifold that sleep reveals is the same structure that organizes waking activity, but sleep lets researchers see it clearly for the first time.
Limits
The study was conducted in mice, and while the fundamental principles of V1 organization are broadly conserved across mammals, the specifics may differ in primates, including humans. The recordings targeted deep layers of V1, and it remains to be seen whether superficial layers follow the same rules. The off-manifold coding was identified using natural images, and other visual features or cognitive tasks may recruit different subspaces. The work is also correlational in part: while sleep reveals the manifold structure, causality between the intrinsic dynamics and the organization of sensory representations has not been directly demonstrated through perturbation experiments.
The bottom line
The primary visual cortex uses low-dimensional intrinsic dynamics, visible during NREM sleep, to carve out a functional geometry that simultaneously integrates movement and vision on the manifold and protects pure visual coding in the off-manifold. The system relies on chorister neurons, whose tight coupling to the manifold paradoxically creates a clean, interference-free subspace for sparse stimulus representation. The brain’s hidden geometry, glimpsed during sleep, may be the key to understanding how it manages the constant tension between mixing and separating the streams of experience.
Source: de Oliveira, E.F., Kim, S., Qiu, T.S., et al. “Sleep reveals dynamics integrating and segregating movement and stimulus representations in V1.” Nature Communications (2026). DOI: 10.1038/s41467-026-75347-4.