💻 computer science

MorphoNAS: Embryogenic Neural Architecture Search Through Morphogen-Guided Development

This paper introduces MorphoNAS, a neural architecture search method inspired by biological morphogenesis and the Free Energy Principle, which uses simple morphogen-based developmental rules to evolve compact genomes into complex, high-performing neural networks for tasks like structural targeting and CartPole control.

Original authors: Mykola Glybovets, Sergii Medvid

Published 2026-03-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mykola Glybovets, Sergii Medvid

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a complex city.

The Old Way (Traditional AI):
Currently, when engineers design Artificial Intelligence (AI), they act like master architects. They sit down with a blueprint and manually draw every single street, building, and power line. They decide exactly how many neurons (brain cells) the AI has and how they connect. This is slow, tedious, and requires a human to know exactly what they want before they start. If they want a different city, they have to redraw the whole blueprint.

The New Way (MorphoNAS):
The paper introduces MorphoNAS, which is like hiring a "biological city planner" instead of an architect. Instead of drawing the final city, you write a tiny, simple set of rules for how the city grows. You start with a single seed (one cell) and let it grow into a complex metropolis based on chemical signals, just like a human embryo grows into a baby.

Here is how it works, broken down into simple concepts:

1. The "Chemical Weather" (Morphogens)

Imagine the development area is a field. In this field, cells release invisible "chemical scents" called morphogens.

Some scents say, "Grow bigger!"
Others say, "Stop growing and become a brain cell."
Others say, "Don't grow here, it's too crowded."

Just like how a flower knows which way is up because of the sun, or how a bird knows where to fly because of the wind, the cells in this system sense these chemical scents. They don't have a master plan; they just react to what is happening right next to them.

2. The "DNA Recipe" (The Genome)

In this system, the "genome" (the DNA code) isn't a blueprint of the final brain. It's more like a recipe for a cake.

The recipe doesn't say, "Make a cake with 500 sprinkles."
It says, "Mix flour and sugar. If the batter is hot, add eggs. If it's cold, add milk."
The final cake (the neural network) emerges naturally from following these simple rules.

In MorphoNAS, the "recipe" tells the cells:

How fast to release the chemical scents.
How far those scents travel.
When a cell should split into two.
When a cell should grow a "wire" (axon) to connect to a neighbor.

3. The "Evolutionary Gardener"

Since we don't know the perfect recipe to build a specific brain, the researchers use Evolution.

They grow 2,000 different "cities" (neural networks) using random recipes.
They test them: "Did this city solve the puzzle?"
The ones that fail are thrown away. The ones that succeed get to "reproduce."
The recipes of the winners are mixed and tweaked (mutated) to make the next generation even better.

Over time, the system "learns" the perfect set of simple rules to grow a brain that can do a specific job.

What Did They Prove?

The researchers tested this in two ways:

Test 1: The "Lego Challenge" (Structural Targeting)
They asked the system: "Grow a brain that looks exactly like this specific random shape with 31 nodes and 87 connections."

Result: The system successfully grew the exact shape 100% of the time, even though the "recipe" was tiny and simple. It proved that complex shapes can emerge from simple rules.

Test 2: The "Pole Balancer" (Functional Control)
They gave the system a classic video game task: CartPole. The AI has to balance a pole on a moving cart.

Result: The system grew a brain that could balance the pole perfectly.
The Cool Part: When they told the system, "Make the brain as small as possible," it didn't just shrink randomly. It evolved a tiny, efficient brain with only 6 or 7 neurons that could still balance the pole perfectly. This is incredibly efficient compared to traditional AI, which often uses thousands of neurons for simple tasks.

Why Does This Matter?

Efficiency: It creates small, efficient brains instead of bloated ones.
Adaptability: It mimics how nature works. Instead of hard-coding a solution, it builds systems that can grow and adapt.
Simplicity: It suggests that we don't need to be genius architects to build complex AI. We just need to write the right "rules of growth," and the complexity will take care of itself.

In a nutshell: MorphoNAS is a way to grow AI brains from a single seed using simple chemical rules, rather than building them brick-by-brick by hand. It's the difference between drawing a tree on a piece of paper and planting a seed that grows into a tree.

1. Problem Statement

Modern Artificial Neural Architecture Search (NAS) methods typically rely on explicit, manual, or routine search strategies that define network structures directly. This contrasts with biological neural systems, which develop from compact genomes using simple, local rules to produce complex, adaptable architectures.

The Gap: Current NAS lacks the efficiency and adaptability of biological development. It often requires massive computational resources and labeled data, failing to leverage the "genomic bottleneck" hypothesis (where a small genome encodes complex developmental rules rather than a full blueprint).
The Goal: To develop a system that grows neural networks from a single progenitor cell using morphogenetic self-organization, inspired by the Free Energy Principle, reaction-diffusion systems, and Gene Regulatory Networks (GRNs), thereby automating architecture search through biological principles.

2. Methodology: The MorphoNAS Framework

MorphoNAS is an evolutionary computational system where neural networks are not explicitly defined but are phenotypes grown from compact genomes.

A. Core Theoretical Foundations

Free Energy Principle (FEP): Cells minimize "surprise" (variational free energy) by reacting to local morphogen concentrations. This drives decisions on cell division, differentiation, and axon growth to achieve homeostasis.
Reaction-Diffusion Systems: Morphogens (chemical signals) are secreted by cells, diffuse across a spatial field, and inhibit/activate other morphogens. This creates concentration gradients (Turing patterns) that provide positional information.
Gene Regulatory Networks (GRNs): The genome encodes parameters for morphogen secretion rates, diffusion profiles, inhibition interactions, and cellular response thresholds.

B. The Developmental Model

The system operates on a 2D toroidal developmental field ( $\Lambda$ ) containing cells and morphogen concentrations.

Components:
- Progenitors: Undifferentiated cells that can divide or differentiate.
- Neurons: Differentiated cells that grow axons.
- Morphogens: Chemical fields guiding growth.
Process Dynamics (Iterative Steps):
1. Secretion: Cells release morphogens based on genome-defined rates.
2. Diffusion: Morphogens spread via convolution/diffusion matrices.
3. Inhibition: Cross-inhibition between morphogens creates complex patterns.
4. Fate Decisions: Based on local concentration thresholds:
  - Division: Progenitor splits.
  - Differentiation: Progenitor becomes a neuron.
  - Axon Growth: Neurons extend axons toward higher morphogen gradients.
5. Connection & Weighting: Axons form synapses if they reach a target neuron and exceed a threshold. Weights are assigned based on local morphogen concentration and distance, with competitive scaling applied later.

C. Evolutionary Search

An Evolutionary Algorithm (EA) evolves the genomes (parameters for the developmental rules) rather than the network structure itself.

Genome Representation: Encodes field dimensions, morphogen count, secretion rates, diffusion tensors, inhibition matrices, and fate thresholds.
Fitness Functions:
- Structural Targeting: Measures how closely the grown graph matches target node counts, edge counts, and source node counts.
- Functional Performance: Evaluates the resulting Recurrent Neural Network (RNN) on control tasks (e.g., CartPole). Includes penalties for excessive network size to encourage compactness.

3. Key Contributions

MorphoNAS System: A novel framework that generates neural architectures via bottom-up, self-organizing developmental processes rather than top-down graph construction.
Compact Genomic Encoding: Demonstrates that complex network topologies can be encoded by a small set of rules (morphogen dynamics and thresholds), mimicking the biological "genomic bottleneck."
Dual-Objective Validation: Successfully validated the system in two distinct domains:
- Structural: Generating graphs with specific topological properties.
- Functional: Evolving controllers for reinforcement learning tasks.
Size-Adaptive Evolution: Showed the ability to evolve both maximal (100 neurons) and minimal (6-7 neurons) functional architectures by adjusting evolutionary pressure (fitness penalties).

4. Experimental Results

A. Structural Targeting (Graph Generation)

Setup: The EA was tasked with evolving genomes to grow graphs matching specific targets (Node count $N$ , Edge count $E$ , Source count $S$ ) across 20 random configurations (ranging from 8 to 31 nodes).
Outcome:
- 100% Success Rate: Valid solutions were found for all 20 target configurations.
- Robustness: While some complex targets (e.g., high edge density) had lower success rates per run, the system consistently found valid solutions across the entire search space.
- Expressiveness: A single compact genome (e.g., for a 31-node, 87-edge graph) successfully guided the self-organization of the required complex topology.

B. Functional Control (CartPole Task)

Setup: Evolved RNNs were tested on the Gymnasium CartPole environment (balancing a pole on a cart).
Experiment 1 (Standard):
- Result: 94% of runs achieved the ideal fitness score (500) in the initial random population (0 generations).
- Network Size: Most successful networks were large (89% had $\ge$ 90 neurons), indicating the developmental process naturally favors larger, viable structures when not constrained.
Experiment 2 (Size Penalized):
- Modification: Added a penalty to the fitness function for networks exceeding a specific connection count.
- Result: The evolutionary pressure successfully shifted the distribution toward compact networks.
- Outcome: 72% of successful runs evolved controllers with only 6 to 7 neurons. This proves the system can adapt architecture size based on selective pressure without losing functional capability.

5. Significance and Implications

Biological Plausibility: MorphoNAS offers a pathway to creating AI that learns and adapts more like biological organisms, potentially reducing the need for massive labeled datasets and explicit architectural design.
Efficiency in NAS: By using developmental rules, the search space is constrained to "growable" architectures, potentially making NAS more efficient than brute-force or gradient-based search methods.
Scalability: The system demonstrated the ability to handle varying structural constraints and functional tasks. The use of local interactions suggests the approach could scale to larger, more complex systems where global coordination is difficult.
Future Directions: The authors propose extending this to hierarchical developmental processes, continuous physical control tasks, and integrating plasticity mechanisms to allow post-development adaptation, bridging the gap between current AI and biological neural systems.

In summary, MorphoNAS validates the hypothesis that complex, functional neural architectures can emerge from simple, biologically inspired developmental rules, offering a promising alternative to traditional Neural Architecture Search.