🧠 neuroscience

Geometric constraints and cognitive inputs jointly shape emergent brain dynamics and topology

By training recurrent neural networks with varying degrees of spatial constraints on a working-memory task, this study demonstrates that only networks embedding the brain's physical geometry can successfully predict empirical fMRI activity and develop brain-like topological features, revealing that physical geometry and cognitive inputs play distinct, complementary roles in shaping emergent brain dynamics.

Original authors: Beyh, A., Kim, J. Z., Bajwa, W. U., Parkes, L.

Published 2026-08-08

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Beyh, A., Kim, J. Z., Bajwa, W. U., Parkes, L.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the human brain as a bustling, super-complex city. This city isn't just a random jumble of buildings; it has a specific map, with neighborhoods (regions) connected by roads (wiring) that follow the laws of physics and geography. Some roads are short and cheap to build between nearby houses, while others are long, expensive highways stretching across the city to connect distant districts. Scientists have long wondered: Does this physical map dictate how the city functions? Or is the city's daily life (thinking, feeling, remembering) so flexible that it can ignore the map entirely?

To understand this, we need to look at two main ingredients. First, there's geometry, which is just the physical layout—how far apart different parts of the brain are in space. Second, there's cognitive input, which is the information the brain receives from the world, like seeing a red ball or hearing a song. The big question is: How do these two things work together to create the brain's amazing ability to think and adapt? If you only look at the map, you might think the city is rigid. If you only look at the traffic, you might think the roads don't matter. But the truth is likely a dance between the two. This is the puzzle that a team of researchers set out to solve using a special kind of computer simulation.

The Brain's Blueprint vs. The Brain's Dance

The researchers built three different types of "digital brains" (called Recurrent Neural Networks, or RNNs) to see which one could best mimic how real human brains work. They trained all of them on a tricky memory game: a "Delayed Match-to-Sample" task. Imagine you see a specific pattern of lights, wait for a moment, and then have to pick that exact pattern out of three options while ignoring two fake ones. It requires holding information in your mind, waiting, and making a smart choice.

Here are the three types of digital brains they tested:

The "Vanilla" Brain: A standard computer brain with no rules about where things are. It's like a city where buildings can be placed anywhere, and roads can go anywhere, with no regard for distance.
The "Masked" Brain: This one has a rule about where information enters and leaves. It's like a city where all mail must enter through the "Visual District" and all decisions must be made in the "Thinking District," but the roads inside the city are still random.
The "Bio" Brain (bioRNN): This is the star of the show. It has the same mail rules as the Masked brain, plus a rule that mimics real brain geography. The connections between its "buildings" are shaped by the actual distances between brain regions in a human head. Short connections are easier to make; long ones are harder.

The Surprise: Geometry is a Shortcut, Not a Cage

When they started training these brains, something cool happened. The "Vanilla" brain learned the game the fastest, which makes sense because it had no rules holding it back. However, the "Bio" brain learned significantly faster than the "Masked" brain, even though it had the most rules to follow. It turns out that giving the digital brain a realistic physical map actually helped it learn the game more efficiently. The physical layout acted like a helpful guide, narrowing down the possibilities so the brain didn't have to waste time guessing where to build its roads.

But the real magic happened when they tested these brains against real human data. The researchers took brain scans (fMRI) from 100 real people doing a similar memory game and asked: "Which digital brain's activity looks most like the real human brain?"

The result was clear: Only the "Bio" brain succeeded. The Vanilla and Masked brains were completely off-base; their internal activity looked nothing like real human brains. The Bio brain, however, organized its activity in a smooth, flowing pattern that recapitulated the real brain's layout. It even recovered the brain's "main highway"—a specific axis that runs from the sensory parts of the brain (feeling and seeing) to the association parts (thinking and planning)—to a moderately close degree compared to the real brain scans themselves.

The Twist: The Map Changes as You Learn

Here is where the story gets even more interesting. The researchers watched the Bio brain learn over time and found a strange, looping pattern called a "hysteresis loop."

Stage 1 (The Setup): At the very beginning, the Bio brain's internal connections looked very much like the physical map of the brain. It was rigid and stuck to the geometry. But at this stage, it couldn't predict how real humans think at all. It had the right map, but no idea how to drive.
Stage 2 (The Breakthrough): As it started to learn the game, its activity suddenly became very similar to real human brain activity. It found the "dance" it needed to do.
Stage 3 (The Departure): Once it mastered the game, something surprising happened. The brain's internal connections started to drift away from the strict physical map. It didn't stay perfectly glued to the geometry anymore. It built new, long-distance shortcuts that were expensive to build but necessary for the task.

This suggests that the physical geometry of the brain is like a scaffold or a stage. You need the stage to build the play, but once the play is running, the actors move freely across it. The brain uses its physical shape to get started, but then it tweaks its connections to do the actual work.

The Secret Sauce: Hubs and Long Roads

As the Bio brain got better at the task, it started developing a specific "topology" (a fancy word for the shape of its connections) that looks exactly like a real human brain. It built hubs—super-connected central stations that link different parts of the network together. It also started building long-range connections that crossed the whole network, even though these were "expensive" to build according to its geometry rules.

This is a big deal because it shows that the brain isn't just a simple grid. It's a complex web where local neighborhoods talk to each other, but a few special "super-highways" connect the whole city. The simulation showed that these complex features emerged naturally as the brain learned, without anyone explicitly telling it to build them.

What This Means

The study suggests that the brain is a master of compromise. It uses its physical shape (geometry) to limit the infinite possibilities of how it could connect, making learning easier. But once it starts learning, it uses cognitive inputs (the tasks and information) to decide which specific connections to strengthen.

The physical map sets the stage, but the play is written by experience. The brain's amazing flexibility comes from its ability to start with a geometrically constrained structure and then gently bend it to meet the demands of the world. The researchers found that you can't have one without the other: the map gives the brain a head start, but the mental work gives it the freedom to be smart.

In short, the brain is like a city that starts with a strict zoning law but eventually builds a few illegal, super-fast tunnels to get the job done. And it turns out, those tunnels are exactly what make the city work.

Technical Summary: Geometric Constraints and Cognitive Inputs Jointly Shape Emergent Brain Dynamics and Topology

Problem Statement
A central question in neuroscience is how the brain's fixed physical geometry gives rise to its flexible functional repertoire. While brain geometry, connectivity, and function are tightly coupled, the brain must selectively uncouple function from structure to adapt to complex, changing environments. Previous modeling efforts have often addressed structure or function separately, or in the absence of behavioral learning goals. This study seeks to understand the relationship between spatial embedding (geometry), connectivity, and function within a single end-to-end model during task learning, specifically asking how physical constraints and cognitive inputs jointly shape emergent neural dynamics.

Methodology
The authors employed Recurrent Neural Networks (RNNs) as in silico testbeds to investigate these relationships. They trained three classes of RNNs on a delayed match-to-sample with distractors (DMS-D) working memory task, creating a graded hierarchy of spatial constraints:

Vanilla RNNs: Standard RNNs with no spatial priors; inputs and outputs are unrestricted.
Masked RNNs: RNNs with projection constraints where inputs enter only through a subset of nodes (visual cortex) and outputs are read only from a separate subset (association cortex), but internal connectivity is unconstrained by geometry.
Biophysical RNNs (bioRNNs): RNNs incorporating both projection constraints and a spatial embedding. The hidden-to-hidden recurrent weights are regularized to align with the Euclidean distance matrix derived from empirical human brain data (Schaefer 200-region atlas), prioritizing connectivity patterns that mirror the brain's geometry.

Each class consisted of 100 randomly initialized networks (100 nodes representing left-hemisphere regions). The networks were trained using backpropagation through time (Adam optimizer) for up to 60,000 epochs. The study assessed performance using:

Task Accuracy: Performance on the DMS-D task.
fMRI Prediction: The ability of the trained networks' hidden-state dynamics (summarized via Principal Component Analysis) to predict empirical fMRI time series from the Human Connectome Project (N-back task), without the networks ever seeing this data.
Topological Analysis: Graph-theoretic metrics (degree, rich-club coefficient, clustering, participation coefficient) of the intermediate hidden-layer nodes.
Trajectory Analysis: Tracking the evolution of structural similarity (to the geometric kernel) versus functional alignment (fMRI prediction) over training.

Key Results

Learning Advantage of Spatial Embedding: While Vanilla RNNs achieved the highest final accuracy and fastest learning, bioRNNs learned significantly faster than Masked RNNs, despite the additional burden of the spatial penalty. This suggests that biologically grounded physical constraints act as useful inductive biases that enhance task learning.
Prediction of Empirical Brain Dynamics: Only bioRNNs successfully predicted empirical fMRI activity above chance levels ( $z_{sub} \approx 5.54$ ). Vanilla and Masked RNNs performed at chance. Crucially, this predictive power depended on the veridical anatomical arrangement of nodes, not merely on spatial autocorrelation; geometry-null models (permuted embeddings) performed significantly worse than the true bioRNNs.
Recovery of Cortical Hierarchy: bioRNNs were the only networks to recapitulate the brain's principal functional hierarchy, the sensorimotor–association (S–A) axis, with an adjusted $R^2$ of 0.49 (approaching the ceiling set by empirical fMRI data, $R^2 = 0.68$ ). This hierarchy emerged from the interaction of geometry and task training, not from geometry alone.
Hysteresis Trajectory: The evolution of bioRNNs followed a distinct three-stage hysteresis loop:
- Stage 1: Structural similarity to the brain's geometry increased rapidly while fMRI prediction remained at chance.
- Stage 2: fMRI prediction rose sharply, largely preceding gains in behavioral accuracy, while structural similarity remained high.
- Stage 3: Structural similarity decreased (trading back) to an intermediate level, while fMRI prediction and task accuracy remained high.
  This indicates that brain-like dynamics are not a fixed function of geometric similarity but are acquired along a path where geometry is laid down first and then partially relaxed as the task is mastered.
Emergence of Topology: During the third training phase, bioRNNs spontaneously developed brain-like topological features, including a rich-club core of highly connected hubs, increased local clustering, and selective long-range connections. These topological features emerged before task mastery and stabilized thereafter, mirroring the structure-before-function ordering seen in the hysteresis trajectory.
Role of Inputs: Variance decomposition revealed that while the shared dynamics between task-driven and noise-driven regimes (driven by the embedded architecture) were the strongest predictor of fMRI activity, structured task inputs provided unique predictive information that noise inputs could not.

Significance and Claims
The paper claims that physical geometry and cognitive inputs play distinct, complementary roles in the emergence of brain organization. Geometry constrains the space of possible brain dynamics, while cognitive inputs determine which specific dynamics are expressed. The study situates topology as the intermediate scaffold through which a physically embedded system reconciles the competing demands of wiring costs and computational demands.

The authors argue that bioRNNs serve as a versatile in silico complement to human neuroimaging, allowing for the causal probing of how anatomy and experience jointly give rise to neural dynamics. By demonstrating that a spatial geometry prior, once shaped by task demands, can reconstruct the brain's principal functional hierarchy and topology without explicit optimization, the work supports the view that the brain's structural scaffold is necessary but not sufficient for function, requiring a flexible departure from fixed geometry to achieve complex cognition. The findings suggest that the brain's hub-and-rich-club scaffold may be laid down early in development to serve future functions, rather than arising solely as a consequence of them.