Harnessing cortical geometry, wiring, and function as inductive biases for recurrent neural networks

This paper demonstrates that leveraging the MICrONS program's co-registered anatomical and functional data to initialize and constrain recurrent neural networks significantly enhances their performance on cognitive tasks while promoting biologically plausible organizational principles.

The Gist

Imagine you are trying to teach a robot how to think. Usually, when scientists build these "thinking" machines (called Recurrent Neural Networks, or RNNs), they start with a blank slate. They give the robot a brain full of random connections, like a tangled ball of yarn, and hope it figures out how to untangle itself through trial and error.

This paper asks a different question: What if we didn't start with a tangled mess? What if we started with a blueprint?

The researchers decided to build their robot's brain using the actual "blueprint" of a mouse's brain. They didn't just guess how the wires should be connected; they used real data from a massive project called MICrONS, which mapped out the exact location, wiring, and activity of nearly 12,000 neurons in a mouse's visual cortex.

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

The Three Ingredients of the "Mouse Blueprint"

The researchers took three specific things from the mouse brain and used them as rules for their robot:

1. The Wiring Diagram (Function): They looked at how the mouse neurons actually talked to each other when the mouse was seeing things. They used this "conversation history" to decide which robot connections should be strong and which should be weak right from the start.
   • Analogy: Instead of giving the robot random friends, they introduced it to the exact people it needs to know based on who actually hung out together in the mouse brain.
2. The Map (Geometry): They placed the robot's "neurons" in 3D space, using the exact coordinates where the real mouse neurons live. They also added a rule: "It costs more energy to talk to someone far away, so try to keep your friends close."
   • Analogy: Imagine a city where you have to pay extra for long phone calls. The robot learns to organize its neighborhood so that people who talk a lot live next door to each other.
3. The Traffic Rules (Communication): They added a rule to encourage the network to be efficient, ensuring information can flow quickly without getting stuck in traffic jams.

The Experiment: Three Different Games

To test if this "mouse blueprint" helped, they taught the robots three different cognitive games:

• The "One-Choice" Game: Remember a goal, wait, then pick the right direction to get there.
• The "Perception" Game: Look at a blurry, noisy picture and decide which way the dots are moving.
• The "Go/No-Go" Game: See a signal and decide whether to act or stay still.

The Results: The Blueprint Wins

The robots built with the mouse blueprint learned much faster and better than the robots with random brains.

• The Secret Sauce: The biggest boost came from the Wiring Diagram (knowing who talks to whom). This was like giving the robot a head start on the hardest part of the puzzle.
• The Map Helped Too: Using the real 3D Map of the neurons gave an extra boost, making the robot even more efficient.
• The "Positive Only" Test: In a super-hard test where the robot was only allowed to use "exciting" signals (no "stop" signals allowed), the random robots completely failed. But the robots with the mouse blueprint kept working perfectly. It was as if the blueprint gave them a sturdy foundation that didn't collapse under pressure.

What the Robot's Brain Looked Like

When the researchers looked inside the brains of the successful robots, they saw something fascinating. The robots didn't just learn the tasks; they organized themselves in a way that looked like a real brain:

• Low Entropy: Instead of a chaotic mess, the connections were highly organized and structured.
• Small-World: They created "neighborhoods" (clusters) that were tightly knit but also had short, fast highways connecting to other neighborhoods. This is exactly how real brains work to be both specialized and efficient.
• Modularity: They formed distinct groups of neurons that handled specific jobs, just like departments in a company.

The Bottom Line

The paper concludes that the physical shape, the wiring, and the activity patterns of a real brain are not just biological details; they are powerful shortcuts for learning.

By copying the "machinery" of the cortex (its geometry, wiring, and function), the researchers built artificial brains that learned more effectively and organized themselves into smarter, more efficient structures. They proved that if you want to build a better thinking machine, you don't just need more data or bigger models; you need to start with a better map.

Technical Summary

Technical Summary: Neuronal Constraints Drive Superior Learning in RNNs

Problem Statement
A central challenge at the intersection of neuroscience and machine learning is understanding how the specific geometry, wiring, and functional organization of the cerebral cortex shape recurrent computation. While Recurrent Neural Networks (RNNs) are powerful tools for modeling temporal dynamics, they are typically initialized with random weights and unconstrained connectivity. This contrasts with biological circuits, where neurons occupy physical space, adhere to wiring economy principles, and participate in structured anatomical and functional relationships. Previous work has suggested that imposing biological constraints (e.g., spatial embedding) can improve learning, but most studies rely on stylized or randomized embeddings rather than measured cortical organization. The unresolved question is whether the specific, measured geometry, wiring, and functional structure of real cortical circuits can serve as powerful inductive biases that drive superior learning compared to generic biological inspiration.

Methodology
The authors leverage the Machine Intelligence from Cortical Networks (MICrONS) dataset, a multimodal functional connectomics resource from the mouse visual cortex. This dataset uniquely co-registers dense two-photon calcium imaging (functional activity) with high-resolution electron microscopy (anatomical connectivity and spatial coordinates) for the same neuronal population (~12,000 excitatory neurons).

The study constructs a family of 11 biologically grounded RNN variants to systematically disentangle the contributions of three specific cortical priors:

1. Functional Weight Initialization ($W^*$ or $W!$): Recurrent weights are initialized using functional relationships derived from neuronal activity (Pearson correlation and Spike Time Tiling Coefficient, or STTC). A control variant ($W!$) permutes these weights to preserve statistical distribution while disrupting the specific neuron-to-weight mapping.
2. Real Spatial Embedding ($D^*$): Recurrent units are assigned positions based on measured neuronal soma coordinates from MICrONS, as opposed to an artificial grid ($D$).
3. Communicability-Based Regularization ($C$ or $C^*$): During learning, a regularization term penalizes long-range communication or enforces similarity between the model's communicability distribution and the empirical distribution derived from anatomical connectivity.

These models are trained on three cognitive decision-making tasks: One-Choice Inference, Perceptual Decision-Making, and Go/NoGo. Performance is evaluated not only by task accuracy but also by emergent graph-theoretic properties of the learned weight matrices, including entropy, modularity, assortativity, and small-worldness. Robustness is further tested by restricting recurrent weights to be strictly positive (excitatory-only regime).

Key Contributions

• Empirical Validation of Cortical Priors: The study demonstrates that grounding RNNs in measured cortical geometry, wiring, and function yields significantly better learning outcomes than unconstrained or partially constrained baselines.
• Hierarchical Contribution of Priors: Through ablation studies, the authors establish a hierarchy of inductive biases:
  • Functional Initialization provides the largest and most general performance gain.
  • Real Spatial Embedding offers a robust, secondary improvement.
  • Communicability Regularization provides selective, context-dependent refinements.
• Structural-Functional Synergy: The research shows that combining functional initialization with real spatial coordinates and communicability constraints leads to networks that are not only more accurate but also converge toward low-entropy, modular, and small-world topologies characteristic of biological networks.
• Robustness under Sign Constraints: The authors show that functional initialization stabilizes learning even when the recurrent architecture is constrained to positive-only weights (mimicking an excitatory-only regime), a condition under which randomly initialized models fail.

Results

• Task Performance: Fully constrained models (e.g., $W^*D^*C$ and $W^*D^*C^*$) consistently outperformed baselines and partially constrained variants across all three tasks. For instance, on Task 1, the fully constrained model achieved ~0.985 accuracy compared to ~0.625 for the baseline.
• Ablation Insights:
  • Removing functional initialization caused a dramatic drop in performance, particularly in Tasks 2 and 3.
  • Replacing real coordinates with a grid ($D$) significantly reduced performance, confirming that actual cortical geometry acts as a meaningful computational prior.
  • The permuted weight variant ($W!$) performed comparably to the structured variant ($W^*$), indicating that the statistical structure of the biological weight distribution is the primary driver, rather than a precise one-to-neuron mapping.
• Emergent Topology:
  • Entropy: Biologically constrained models converged to low-entropy (highly structured) weight distributions, whereas baselines remained high-entropy (random-like).
  • Modularity & Small-Worldness: Models with direct communicability regularization and real coordinates developed strong modularity and small-world structures ($\sigma \approx 1.5–2.5$). Notably, some variants achieved low entropy and high performance without strong modularity, suggesting these are distinct but related outcomes.
  • Assortativity: Functional initialization tended to drive networks toward disassortative (hub-periphery) structures, while spatial-only constraints sometimes led to assortative (hub-rich) structures.
• Positive-Only Regime: Under the constraint of positive-only weights, randomly initialized models collapsed to chance performance, while functionally initialized models retained high accuracy, demonstrating that biological priors can stabilize optimization in difficult parameter spaces.

Significance and Claims
The paper claims that the "machinery of cortex"—specifically its geometry, wiring, and functional structure—can be harnessed as a powerful inductive basis for building recurrent networks. The authors argue that these priors do more than make models more "biologically realistic"; they fundamentally improve learning efficiency and robustness.

Key takeaways emphasized by the authors include:

1. Inductive Bias over Complexity: Better recurrent computation can be achieved not by increasing model size or architectural complexity, but by reshaping the optimization landscape through biologically grounded priors.
2. Structure-Function Interdependence: While anatomical structure provides a scaffold, functional information (activity-derived relationships) is the decisive constraint for specifying recurrent dynamics. Structure alone does not uniquely determine dynamics; functional priors are required to narrow the solution space to effective regimes.
3. Design Principles for AI: The findings suggest that cortical organization principles (spatial embedding, functional initialization, wiring economy) are computationally consequential and can serve as reusable design principles for artificial systems, particularly in domains requiring data efficiency and robust internal organization.

The authors maintain a modest scope, acknowledging limitations such as the focus on excitatory neurons, the specific experimental regime of the MICrONS data, and the use of statistical rather than causal functional measures. However, they assert that the results provide direct empirical support for the hypothesis that cortical microcircuits are organized around efficiency principles that can be effectively transferred to artificial recurrent networks.