🧬 biology

Learning Developmental Scaffoldings to Guide Self-Organisation

This paper introduces a joint learning framework combining Neural Cellular Automata with learned coordinate-based pre-patterns to demonstrate that offloading developmental information into initial conditions enhances the robustness, encoding capacity, and symmetry-breaking capabilities of self-organizing systems compared to purely self-organizing alternatives.

Original authors: Milton L. Montero, Elias Najarro, Jakob Schauser, Sebastian Risi

Published 2026-05-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Milton L. Montero, Elias Najarro, Jakob Schauser, Sebastian Risi

Original paper licensed under CC BY 4.0 (http://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

The Big Idea: The "Blueprint" vs. The "Recipe"

Imagine you want to build a complex sandcastle. You have two ways to do it:

The Pure Recipe (Self-Organization): You give a group of builders a single rule: "If you see a neighbor to your left, move right." You start with a flat pile of sand. The builders follow the rule, and eventually, a castle emerges. The problem? If the wind blows (noise) or one builder makes a mistake, the whole castle might collapse. Also, if you want a different castle (like a tower instead of a wall), you have to invent a completely new set of rules.
The Scaffolded Approach (This Paper): You give the builders the same simple rule, but before they start, you place a faint, temporary outline of the castle in the sand. This outline isn't the castle itself; it's just a guide. The builders follow their simple rule, but because they start on this guide, they naturally build the castle correctly, even if the wind blows or a builder slips.

This paper is about the second approach. The researchers wanted to understand how nature builds complex things (like a human embryo) by combining a "guide" (initial conditions) with a simple set of local rules (self-organization).

The Experiment: Teaching a Robot to Draw

The researchers built a computer simulation to test this. They used two main tools:

The "Builder" (NCA): A Neural Cellular Automaton. Think of this as a grid of tiny pixels that talk to their neighbors. They have a simple brain that says, "Based on what my neighbors are doing, I will change my color." By themselves, these pixels are blind; they don't know where they are or what the final picture should look like.
The "Guide" (SIREN): A special AI function that looks at the coordinates (x, y) of the grid and paints a starting picture. This is the "pre-pattern" or "scaffold."

The Goal: They trained both the "Builder" and the "Guide" at the same time to create 20 different emojis (like a heart, a flower, or a butterfly).

What They Discovered

The paper found three major benefits to using a "Guide" (pre-pattern) instead of relying on the "Builder" alone:

1. The Guide Doesn't Need to Be Perfect (It's a Compass, Not a Map)

You might think the "Guide" would try to draw the perfect emoji right away, so the Builder just has to copy it.

The Surprise: The Guide didn't draw the perfect emoji. Instead, it drew a blurry, rough sketch that looked like a "compass."
The Analogy: Imagine trying to find a hidden treasure. A perfect map shows you exactly where to dig. A compass just points North. The Guide acted like a compass. It didn't tell the Builder exactly where the flower petals go; it just tilted the sand so that the Builder's simple rules naturally led to the petals forming in the right place.
Why it matters: The Guide sets the "trajectory." It biases the process so the Builder doesn't have to figure out everything from scratch.

2. It's Much Tougher Against Mistakes (Robustness)

The researchers tested what happens if the "Builders" (the pixels) start making mistakes or ignoring their neighbors (simulating "noise" or unreliable cells).

The Result: The system with the Guide kept working almost perfectly, even when the builders were unreliable. The system without the Guide (pure self-organization) fell apart immediately.
The Analogy: If you are walking in a foggy forest with a compass (the Guide), you can still find your way even if you stumble. If you are walking in the same fog with no compass, relying only on your memory of the path, one stumble sends you off course forever. The Guide provides a "safety net" that the Builder can rely on.

3. You Can Store More Ideas in Less Space (Capacity)

They tried to teach the system to draw 1, 2, 4, 8, and finally 16 different emojis.

The Result: The system with the Guide could learn all 16 emojis with high quality. The system without the Guide struggled and produced blurry, bad images as the number of emojis increased.
The Analogy: Think of the "Builder's" brain as a small backpack.
- Without a Guide: To draw a flower, the backpack must carry the instructions for "how to make petals." To draw a butterfly, it must carry "how to make wings." It gets heavy and full quickly.
- With a Guide: The Guide holds the specific instructions for "this is a flower" or "this is a butterfly" in the starting position. The Builder's backpack only needs to carry the general rule of "how to build." This is a trade-off: you use "memory" (the Guide's starting image) to save "computation" (the Builder's brain space). This allows the system to handle many more complex shapes.

The "Growing" Bonus

The researchers also tested what happens if the system starts as a tiny dot and grows into a larger shape (like an egg hatching).

The Finding: The Guide naturally learned to draw a wavy, striped pattern (like a sine wave) across the starting dot.
The Connection: This looks exactly like how real biological embryos work. In nature, mother organisms leave behind chemical gradients (like stripes of light and dark) that tell the embryo which end is the head and which is the tail. The computer figured this out on its own! It realized that to break symmetry (decide "left" vs. "right") in a growing system, a wavy gradient is the most efficient way to start.

Summary

The paper proves that complexity doesn't require complex rules.

Nature (and this AI) builds amazing structures not by having a master plan that controls every single cell, but by having a simple set of rules plus a smart starting point. The starting point (the pre-pattern) does the heavy lifting of "setting the stage," allowing the simple rules to do the rest. This makes the system stronger against errors, able to learn more things, and capable of breaking symmetry to create distinct shapes.

In short: Don't try to program every detail. Just set the stage correctly, and let the actors (the self-organizing cells) do the rest.

Technical Summary: Learning Developmental Scaffoldings to Guide Self-Organisation

Problem Statement

Self-organisation describes how global structure emerges from local interactions without a central blueprint. However, a fundamental question remains: where does the information driving these processes originate? In natural systems, particularly biological development, a significant portion of this information is not generated by the self-organising process itself but is "offloaded" to the system's initial conditions. These initial conditions take the form of spatially structured pre-patterns (e.g., morphogen gradients, maternal mRNAs) that scaffold the subsequent self-organisation.

While biological and theoretical studies have characterized these phenomena, there is a lack of controlled, quantitative models that allow researchers to systematically vary the balance between pre-patterned information and self-organisation to measure downstream consequences. Existing approaches either focus on specific biological mechanisms (genes/proteins) or abstract theoretical models, neither of which permits the joint learning and manipulation of both components under a unified objective.

Methodology

The authors propose a computational framework that jointly learns a pre-pattern generator and a self-organising system to generate a set of target patterns.

Model Architecture

The system consists of two primary components trained simultaneously:

Self-Organising Component (Neural Cellular Automata - NCA): An isotropic NCA is used as the self-organising engine. In its standard form, the NCA has no pre-specified spatial structure in its initial state, making it a clean baseline for studying the effects of pre-patterning. It updates cell states based on local interactions using learnable update rules (MLPs).
Pre-Pattern Generator (SIREN): To generate the initial conditions (the scaffold), the authors employ a SIREN (Sinusoidal Representation Network). Unlike traditional CPPNs which are often trained via neuroevolution, SIRENs use periodic activation functions (sine) in a Multi-Layer Perceptron (MLP) and are amenable to gradient-based optimization.
- Modulation: To handle multiple targets with a single set of dynamics, the SIREN is modulated. An embedding model generates per-target embeddings, which are used to create scaling factors for the SIREN's hidden units. This allows the generation of distinct spatial pre-patterns for different targets while sharing the underlying function parameters.
- Initialization: The SIREN outputs the initial values for the NCA's visible channels (RGBA), which are concatenated with zero-valued hidden channels to form the initial state $s_0$ .

Training and Evaluation

Objective: The system is trained end-to-end using Stochastic Gradient Descent (SGD) with Nesterov momentum to minimize the Mean-Squared Error (MSE) between the final NCA output and the target pattern.
Dataset: The model is evaluated on a dataset of 20 hand-crafted emojis featuring various symmetries (radial and axial) and structural similarities.
Metrics: The study employs information-theoretic measures (Normalized Mutual Information, Linear Probe $R^2$ , Structured Similarity Index Measure - SSIM) to analyze the distribution of information between the pre-pattern and the final reconstruction.
Comparisons: The proposed model is compared against a purely self-organising alternative, the Goal-Directed NCA (GoalNCA), which conditions its update function on target embeddings but lacks a pre-pattern scaffold.

Key Contributions and Results

1. Information Offloading and Trajectory Bias

Information-theoretic analysis reveals that effective pre-patterns do not simply approximate the final target. Even when the SIREN has high capacity, the information content of the pre-pattern does not substantially increase, nor does it become a direct reconstruction of the target. Instead, the pre-pattern learns to bias the developmental dynamics. It sets the NCA on a trajectory that is easier to traverse in developmental space, facilitating convergence without needing to encode the full target structure upfront.

2. Robustness to Cellular Noise

The study demonstrates that offloading information to initial conditions significantly improves robustness.

Experiment: The reliability of cell updates was reduced (simulating noise) during testing.
Result: The pre-patterned NCA maintained high performance even with low update probabilities, whereas the GoalNCA (pure self-organisation) diverged rapidly.
Mechanism: The pre-pattern initializes cells along a path requiring less information propagation via self-organisation. In contrast, the GoalNCA must propagate boundary condition information entirely through self-organisation, a process that fails when inter-cell communication becomes unreliable.

3. Encoding Capacity (Memory-Compute Trade-off)

The authors show that the combination of pre-patterns and self-organisation improves encoding capacity, analogous to a memory-compute trade-off.

Experiment: Models were tasked with encoding an increasing number of patterns (1 to 16) with a fixed parameter budget.
Result: The pre-patterned model degraded much more gracefully than the GoalNCA as the number of patterns increased.
Interpretation: By storing per-target positional information in the initial state (memory), the system spares the self-organising process (compute) from having to encode complex propagation rules to disambiguate every target. This allows the same parameter budget to encode a larger set of patterns.

4. Growth and Symmetry Breaking

The framework was extended to growing systems where the substrate expands from an initial ellipsoidal shape.

Observation: In these simulations, the learned pre-patterns consistently exhibited sinusoidal-like structures along the anterior-posterior axis, despite the targets not possessing this shape.
Significance: This mirrors the spatial arrangement of homeobox gene expression in early biological development (e.g., Drosophila). The sinusoidal pattern appears to be an optimal mechanism for symmetry breaking in growing systems, enabling the propagation of developmental signals downstream.

Significance and Claims

The paper claims to provide the first computational model capable of jointly learning top-down pre-patterns alongside a shared self-organising developmental program. By directly comparing a scaffolded system against a non-scaffolded alternative, the work offers an information-theoretic account for why pre-patterns are prevalent in complex developmental systems.

The authors argue that the benefits of scaffolding—robustness, increased encoding capacity, and effective symmetry breaking—are different manifestations of the same memory-compute trade-off. Information transferred into the initial state (memory) reduces the burden on the self-organising process (compute), which no longer needs to generate this information from scratch.

The study concludes that while self-organisation can theoretically achieve similar tasks, it is less efficient, less robust, and less expressive without the guidance of initial conditions. The findings suggest that the structure of initial conditions and the dynamics of self-organisation have a non-trivial relationship, where pre-patterns do not merely "pre-compute" the answer but rather guide the system toward a solution space that is easier to navigate.