A new study suggests that the brain’s neural activity may be governed by a surprisingly simple principle: strong, highly localized feedback connections can sharply limit the number of dimensions available to neural populations, even when those populations contain thousands or millions of neurons. The finding, reported by David Dahmen, Stefano Recanatesi, X. Jia and colleagues in Nature Neuroscience, offers a fresh explanation for why brain activity often occupies a compact, structured space rather than wandering freely through every possible combination of neuronal states. The work could help explain how different brain areas remain flexible enough to process complex information while still maintaining stable, recognizable patterns of activity.
At first glance, the brain appears to be an almost impossibly high-dimensional system. Every neuron can change its firing rate independently, at least in principle, creating a vast mathematical space in which the collective activity of a neural population could evolve. If a network contains thousands of neurons, the number of possible activity patterns becomes astronomically large. Yet experiments repeatedly show that real neural activity is far more constrained. When researchers record many neurons simultaneously, they often find that the activity can be described using a relatively small number of coordinated patterns, known as neural dimensions. These dimensions do not represent individual neurons; instead, they capture collective modes in which groups of cells rise, fall or interact together.
The new research focuses on recurrence, the process by which neural signals feed back into the same network or return to a nearby circuit after passing through other neurons. Recurrence is a defining feature of biological brains. Unlike a simple one-way chain of information processing, the brain is filled with loops. A signal can influence a local population, alter its future state, and then be fed back into the circuit milliseconds later. Such feedback can amplify activity, stabilize it, or push the network into a new configuration. The study argues that the strength and spatial organization of these recurrent connections are crucial in determining how many independent patterns of activity a brain area can support.
The central result is that strong recurrence does not necessarily make a network more complex in the sense of increasing its effective dimensionality. Instead, when recurrent interactions are concentrated among nearby or functionally related neurons, they can compress the network’s activity into a smaller set of dominant modes. In mathematical terms, the network’s activity becomes confined to a lower-dimensional manifold within the full space defined by all individual neurons. A manifold can be imagined as a curved surface embedded in a much larger space: although countless coordinates are available, the system’s actual states remain close to a restricted structure. This compression may be one of the ways the brain turns enormous biological complexity into manageable computation.
The researchers distinguish between the number of neurons in a region and the number of dimensions that are actually used by its activity. These quantities are not equivalent. A population may contain a large number of cells but behave collectively as if it were controlled by only a few variables. For example, many neurons may vary their activity in highly correlated ways, meaning that their signals carry overlapping rather than independent information. Strong local recurrence can generate precisely this type of coordination. Instead of allowing every neuron to fluctuate separately, feedback links can synchronize or organize subsets of cells, reducing the effective degrees of freedom while preserving meaningful dynamics.
This mechanism may also explain why dimensionality differs across brain areas. Regions involved in fast sensory encoding may require a broad repertoire of activity patterns to represent rapidly changing features of the outside world. Other areas, including circuits involved in memory, decision-making or motor planning, may benefit from more constrained dynamics that can stabilize internal states and guide behavior over time. According to the study’s framework, these differences do not require every region to follow a completely separate design. They can emerge from variations in the strength, range and localization of recurrent connectivity. A small change in how strongly nearby neurons influence one another could alter the geometry of the entire population’s activity.
The result has implications for how scientists interpret neural recordings. A common approach is to calculate the dimensionality of a population by examining the covariance or correlation structure of its activity. If many neurons fluctuate together, the data can be summarized by a small number of principal components. But correlations alone do not reveal why those patterns exist. The new work connects the observed dimensionality to the underlying architecture of the network, showing how local feedback can shape the spectrum of collective activity. In such a network, a few modes may become especially dominant, while other potential patterns are suppressed because recurrent interactions pull the system back toward preferred configurations.
The findings may help bridge two seemingly conflicting views of the brain. One view emphasizes the immense richness and flexibility of neural computation; the other highlights the strong regularities and constraints visible in large-scale recordings. Local recurrent circuits could provide both. Their feedback may reduce unnecessary variation, making neural states more robust against noise, while still allowing the network to switch between multiple stable or metastable patterns. In this picture, low dimensionality is not a sign that the brain is performing a simple computation. Rather, it may indicate that the computation has been organized efficiently, with a limited set of coordinated variables carrying the information most relevant to the task.
The study also raises questions about how neural dimensionality changes with learning, development and disease. Learning may modify recurrent weights, strengthening some local loops and weakening others, thereby reshaping the activity manifold without requiring large-scale anatomical rewiring. Development could progressively tune these circuits so that different brain areas acquire distinct computational roles. Conversely, disorders that disrupt excitation, inhibition or the spatial structure of connectivity might cause neural activity to become either excessively constrained or abnormally diffuse. Conditions associated with altered network stability, including epilepsy, schizophrenia or neurodegenerative disease, could therefore involve changes not only in how strongly neurons fire but also in the dimensionality of the collective dynamics.
For artificial intelligence, the work offers an appealing design principle. Artificial neural networks often become difficult to control when feedback is added, because recurrent loops can produce unstable or chaotic activity. The brain’s strategy suggests that carefully localized recurrence may provide a way to obtain stable, low-dimensional dynamics without sacrificing the ability to represent complex sequences. Engineers could use this principle to build more efficient recurrent systems in which strong local interactions create reliable computational states, while longer-range connections preserve flexibility and communication between modules. The broader message is that intelligence may not require every unit in a network to remain independently expressive; it may depend on organizing many units into a small number of powerful, coordinated dynamical modes.
The authors’ conclusion places recurrence at the center of a major question in neuroscience: how does anatomy become computation? The answer emerging from this work is that the spatial arrangement of feedback may be just as important as the number of neurons or the strength of their individual responses. When recurrence is strong and localized, it can act like an invisible sculptor, shaping the high-dimensional activity of a neural population into a smaller and more functional form. Brain areas may therefore differ in their computational capacities not simply because they contain different cell types, but because their recurrent architecture channels activity through different geometric spaces. By linking microscopic connectivity with macroscopic neural dynamics, the study provides a potential framework for understanding how the brain remains both extraordinarily complex and remarkably organized.
Subject of Research: The influence of strong, localized recurrent connectivity on the dimensionality and organization of neural activity across brain areas.
Article Title: Strong and localized recurrence controls the dimensionality of neural activity across brain areas.
Article References: Dahmen, D., Recanatesi, S., Jia, X. et al. Strong and localized recurrence controls the dimensionality of neural activity across brain areas. Nature Neuroscience (2026). https://doi.org/10.1038/s41593-026-02395-w
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41593-026-02395-w
Keywords: neural activity, recurrent networks, brain connectivity, neural dimensionality, population dynamics, low-dimensional manifolds, computational neuroscience, brain areas, localized recurrence, neural computation
Tags: brain information processingbrain region connectivitycortical and subcortical neural interactionsdimensionality reduction in neurosciencefeedback mechanisms in neural circuitshigh-dimensional brain activitylocalized feedback connections in neural networksneural activity dimensionalityneural activity structureneural network stabilityneural population dynamicsstable neural pattern formation
