Linking macroscale structure and function in brain-like recurrent neural networks

This paper introduces BrainRNN, a recurrent neural network architecture inspired by macroscale human cortical structure, demonstrating that imposing brain-like structural constraints enables the emergence of functional organization—including modules and gradients—that closely mirrors the human brain and allows function to be inferred from structure.

The Gist

Imagine you are trying to understand how a city works. You could look at the traffic patterns (what people are doing), or you could look at the map of the roads and bridges (how the city is built). In neuroscience, scientists have long known that the map (the brain's physical structure) dictates the traffic (how we think and behave). But in the world of Artificial Intelligence (AI), most computer models are built like a giant, tangled ball of yarn where every wire can connect to every other wire without any rules. They work great, but they don't look like a brain, and it's hard to tell why they work the way they do just by looking at their "wiring."

This paper introduces a new kind of AI called BrainRNN. Think of it as building a city for a computer that actually follows the rules of urban planning found in the human brain.

Here is the breakdown of what they did and what they found, using simple analogies:

1. Building the City (The Architecture)

The researchers didn't just throw neurons together randomly. They built their AI network inside a 3D hemisphere, just like the human brain is shaped.

• The Neighborhoods: They divided this digital brain into three specific neighborhoods:
  • The Back (Visual): Where "eyes" are located to see inputs.
  • The Front (Motor): Where "hands" are located to make movements.
  • The Middle (Association): The vast downtown area where thinking, planning, and decision-making happen.
• The Wiring Cost: In real life, building a bridge across a wide river is expensive. In the brain, connecting neurons that are far apart is "expensive" in terms of energy and space. The researchers added a rule to their AI: "Don't build long wires unless you really have to." This forces the AI to be efficient, just like a real brain.

2. The Experiment: Teaching the City to Work

They taught this "Brain City" to perform 22 different cognitive tasks, ranging from simple things (like "look at a dot and move your hand to it") to complex things (like "remember a location, ignore a distraction, and then decide where to move").

3. What They Discovered (The "Aha!" Moments)

A. The "Downtown" Effect
When the AI was allowed to wire itself freely (no rules), it used every single neuron for every task. It was like a city where everyone runs to the grocery store for a cup of milk.
But when they added the "Wiring Cost" rule (making long connections expensive), something magical happened:

• Simple tasks (like looking and moving) only used the "Back" and "Front" neighborhoods.
• Complex tasks (like memory and decision-making) forced the AI to activate the "Downtown" (Association) area.
• The Lesson: Just like in humans, the "middle" part of the brain is reserved for the hard thinking. If you make the brain efficient, it naturally expands its "thinking district" to handle complex problems.

B. The Map Predicts the Traffic
In a normal AI, if you look at the wiring, you can't guess what the computer is thinking. But in this BrainRNN, the structure predicts the function.

• Because the "Back" is close to the "Front," they talk to each other easily.
• Because the "Downtown" is in the middle, it acts as a hub.
• The researchers found that the AI developed functional modules (specialized teams) and gradients (smooth transitions from simple to complex processing) that looked exactly like the maps scientists see in human brains.
• The Analogy: It's like if you built a subway system with specific rules, and without telling the trains where to go, they naturally organized themselves into efficient lines that matched the city's geography.

C. The "Trade-Off"
When they made the "wiring cost" rule very strict (forcing the AI to be super efficient), the AI got worse at complex tasks. It couldn't afford to connect the "Downtown" to the "Front" enough times to do hard thinking.

• The Lesson: This explains why evolution might have expanded the human "association cortex" (the thinking part). We needed that extra space and those extra connections to handle complex thoughts, even though it costs more energy to build and maintain.

The Big Takeaway

This paper is a bridge between Biology and AI.

• Before: AI was like a magic box. We knew it worked, but we couldn't explain its internal logic just by looking at its structure.
• Now: By building AI with "brain-like" rules (physical space, wiring costs, and specific input/output zones), the AI naturally organizes itself in a way that looks like a human brain.

In short: If you build a computer with the same physical constraints as a brain, it will naturally learn to think like a brain. This helps scientists understand why our brains are built the way they are, and it gives AI researchers a new way to build smarter, more interpretable machines that we can actually understand.

Technical Summary

1. Problem Statement

The relationship between neural structure and function is a central challenge in both neuroscience and artificial intelligence (AI).

• In Neuroscience: Macroscale brain organization (parcellations, modules, hierarchies) is known to be inferred from intrinsic structural architecture (topography and topology) without explicit task activation. However, the causal mechanisms linking these structural constraints to functional organization remain poorly understood.
• In AI: While artificial neural networks (ANNs) successfully model human cognition at the representational level, their internal architectures are typically freely optimized with goal-driven training. They lack the biologically grounded structural constraints (e.g., spatial embedding, wiring economy) found in the brain. Consequently, it is unclear if principles linking brain anatomy to function can be extended to ANNs, or if imposing such constraints can induce comparable functional organization.

2. Methodology: BrainRNN Architecture

The authors introduce BrainRNN, a recurrent neural network (RNN) designed to mimic macroscale human cortical structure through specific structural priors:

• Topographic Embedding:
  • Spatial Manifold: 128 recurrent units are distributed quasi-uniformly across a 3D hemispheric manifold using a Fibonacci lattice.
  • Areal Segregation: Units are assigned fixed coordinates along Left-Right (L-R), Posterior-Anterior (P-A), and Dorsal-Ventral (D-V) axes.
  • Functional Interfaces:
    • Visual Input: Restricted to 32 units in the posterior region (simulating primary visual cortex).
    • Motor Output: Read out from 16 units in the anterior region (simulating primary motor cortex).
    • Association Units: The remaining 80 units located between visual and motor regions.
• Topological Constraints (Wiring Economy):
  • The network is trained with a wiring-cost regularization term ($\lambda$) in the loss function.
  • The penalty is calculated as the L1 norm of recurrent weights multiplied by the Euclidean distance between connected units ($L_{cost} = \lambda |W_{rec} \odot D|$).
  • This enforces a trade-off between functional integration and the physical cost of long-range connections, mimicking biological "economical connectivity."
• Training Paradigm:
  • Tasks: The network is trained on 22 cognitive tasks across four families: Visuomotor (VM), Working Memory (WM), Match–Nonmatch (MNM), and Decision-Making (DM).
  • Objective: Minimize task performance loss (Mean Squared Error) while minimizing wiring costs.
  • Comparison Baselines: The authors compare BrainRNNs against:
    • CRNNs: Conventional RNNs with no structural constraints ($\lambda=0$, no spatial embedding).
    • SpRNNs: Spatially embedded RNNs without areal interfaces (randomized input/output locations).

3. Key Contributions

1. Framework Development: Creation of BrainRNN, the first model to jointly integrate hemispheric topographic embedding, area-specific input/output segregation, and wiring-cost constraints.
2. Causal Link Demonstration: Providing evidence that structural constraints (specifically wiring economy and spatial interfaces) are sufficient to induce functional organization (modules and gradients) in ANNs, allowing function to be inferred from structure.
3. Mechanistic Insight: Revealing that the expansion of association regions for higher-order cognition is a direct consequence of structural constraints and the segregation of sensory/motor interfaces.

4. Key Results

A. Structural Constraints Shape Activation and Cognitive Capacity

• Activation Patterns: As the wiring-cost constraint ($\lambda$) increases, the total number of activated units decreases. Crucially, association units (the "middle" region) are selectively suppressed, while visual and motor units remain active.
• Connectivity Shift: Higher $\lambda$ shifts connectivity from association-heavy to visual/motor-heavy.
• Cognitive Hierarchy:
  • Simple visuomotor tasks remain robust even under strict constraints (relying on visual/motor units).
  • Complex tasks (Working Memory, Decision-Making, Match-Nonmatch) suffer significant performance drops as $\lambda$ increases, correlating with the loss of activated association units.
  • Conclusion: Higher-order cognition requires the recruitment of association regions, which is structurally constrained by wiring costs.

B. Structure–Function Coupling

• Validation against Human Data: The authors compared BrainRNNs to Human Connectome Project (HCP) MRI data.
• Distance vs. Function: Both human brains and BrainRNNs exhibit a negative correlation between spatial distance and functional connectivity (distant regions are less functionally correlated).
• Connectivity vs. Function: Both exhibit a positive correlation between the cosine similarity of structural connectivity and functional connectivity.
• Superiority of BrainRNN: BrainRNNs showed significantly stronger structure–function coupling than CRNNs or SpRNNs, closely mirroring empirical human data.

C. Emergence of Functional Modules

• Module Discovery: Functional modules were identified by clustering task variance.
• Spatial Distribution: Modules in BrainRNNs were neither purely localized nor randomly distributed; they occupied an intermediate state similar to human cortical modules (e.g., Yeo 7/17 networks).
• Alignment: Functional modules significantly overlapped with:
  1. Predefined visual/motor/association territories.
  2. Network modules derived from structural connectivity (Louvain algorithm).
• Role of Constraints: As $\lambda$ increased, network modularity increased, and the alignment between structural and functional modules strengthened. This suggests structural constraints drive the emergence of functional modularity.

D. Discovery of Functional Gradients

• Gradient Extraction: Functional gradients were derived via Principal Component Analysis (PCA) on the cosine similarity of recurrent connectivity.
• Axis Alignment: The first two gradients in BrainRNNs mirrored human brain gradients:
  1. Sensorimotor-Association Axis: Units transitioned from visual $\to$ delay/association $\to$ motor.
  2. Sensory-Motor Axis.
• Topographic Embedding: Gradient values were significantly negatively correlated with Euclidean distance, indicating that functional transitions are anchored to the physical topography.
• Constraint Effect: Stronger wiring constraints reduced the number of significant functional gradients, linking structural economy to reduced functional diversity.

5. Significance and Implications

• Bridging Neuroscience and AI: The study demonstrates that the principles used in systems neuroscience to infer function from structure (e.g., using structural connectivity to predict functional networks) are applicable to artificial networks, provided the networks are structurally grounded.
• Evolutionary Insight: The results offer a mechanistic explanation for the evolutionary expansion of the association cortex: it is a necessary trade-off to support higher-order cognition under the constraints of wiring economy and spatial segregation of sensory/motor interfaces.
• Interpretability: By embedding structural priors, BrainRNNs offer a more interpretable framework where functional roles can be predicted based on anatomical position and connectivity, moving beyond post-hoc attribution methods common in standard deep learning.
• Clinical Relevance: The framework provides a potential "in silico" platform to study how structural disruptions (e.g., in psychiatric disorders) might lead to specific functional deficits, aiding in the discovery of biomarkers.

In summary, the paper establishes that macroscale structural constraints are not merely biological artifacts but are fundamental drivers of functional organization, enabling artificial networks to replicate the complex structure-function relationships observed in the human brain.