Conditional Morphogenesis: Emergent Generation of Structural Digits via Neural Cellular Automata

AI summary

Overview

Research area: Neural Cellular Automata (NCA), self-organizing generative models, and class-conditional image generation (cs.NE).

Technical level: Intermediate — the paper is written accessibly and explains its architecture step by step, but assumes familiarity with convolutional networks, backpropagation through time, and generative model evaluation metrics.

Scope: The paper presents a Conditional Neural Cellular Automata (c-NCA) that grows any of the ten MNIST digit classes from a single generic seed pixel, using one fixed set of local rules steered by a spatially broadcast one-hot class vector.

What This Paper Is About

Most Neural Cellular Automata are "specialists": one trained model grows one shape or one infinite texture. This paper asks whether a single set of local, translation-equivariant rules can instead encode the "DNA" for ten different handwritten digits and selectively express whichever one it is told to grow. The goal is a biologically plausible, lightweight alternative to global-receptive-field generative models such as GANs and VAEs, where a single decentralized system grows distinct structural targets from a one-pixel seed.

Key Contributions

1. A Conditional NCA architecture (c-NCA) that injects a one-hot class vector of size 10 directly into the cellular perception field at every time step — persistent guidance rather than a one-time instruction encoded only in the seed state (as in the earlier StampCA work).
2. Demonstration that one fixed set of local rules can produce ten distinct geometric attractors on full MNIST, with all digits emerging from the same generic seed at grid center.
3. A rigorous self-repair study across all ten digit classes, showing recovery after instantaneously dropping 50% of active pixels and evolving for 48 recovery steps, without any training on damaged data.
4. An inference-time ablation on stochasticity, showing that setting the update probability to p = 1.0 (deterministic) produces "ghosting" artifacts and blurred edges compared with the standard p = 0.5.

Main Findings

• Generation accuracy: A standard LeNet-5 classifier trained on original MNIST, used as an external judge on 1,000 digits generated by the c-NCA from random seeds, achieved a recognition accuracy of 96.30%.
• Semantic confidence: The same classifier assigned a mean softmax probability of 94.76% to the generated digits.
• Homeostatic stability: Running the simulation to t = 128 and comparing with t = 64, the MSE was 0.0297 — a low residual indicating a stable dynamic equilibrium that "breathes" slightly due to stochasticity while maintaining global topology.
• Structural similarity: Average SSIM between generated digits and random test samples was 0.4826, which the paper attributes to the model learning a canonical "ideal" shape rather than memorizing specific handwriting variations.
• Parameter efficiency: The entire model has 10,048 trainable parameters, contrasted with roughly 3 × 10^6 typically required by DCGAN-style models.
• Systematic failure mode: The dominant confusion occurs where the target digit '1' is misclassified as '8', affecting approximately 30% of '1' samples. The paper attributes this to "structural over-growth" — stochastic fluctuations allowing lateral artifacts to form that the classifier reads as the closed loops of an '8'.
• Developmental timeline: All digits share a common ancestry from t = 0 to 8 as a single active pixel, undergo symmetry breaking from t = 8 to 24 guided by the condition vector, and refine from t = 32 to 64 as boundaries sharpen into a stable attractor.
• Stochasticity matters: The p = 0.5 stochastic update produced sharp boundaries, while the p = 1.0 deterministic variant showed blurred edges and synchronization artifacts, confirming stochasticity acts as a regularizer.

Methodology in Plain English

The researchers treat image generation as a growth process on a 28 × 28 grid. Each cell carries 16 channels: four visible RGBA values and twelve hidden channels that act as a latent "chemical" memory for long-range communication via local propagation. The grid starts as all zeros except for a single seed activation at the center (14, 14) with alpha = 1.

At each step, every cell looks only at its immediate 3 × 3 Moore neighborhood through a learnable depthwise convolution, keeping the system translation-equivariant. The one-hot class label (10 values) is broadcast across the whole grid so that every cell sees the same global instruction, and it is concatenated with the local perception vector. A small update network — two 1 × 1 convolutional layers with 128 hidden units and a ReLU — maps that combined input to a proposed state change. A stochastic mask sampled from a Bernoulli distribution with p = 0.5 decides which cells actually update, mimicking asynchronous biological behavior and preventing oscillatory patterns.

An "alive masking" rule resets cells whose alpha is not above 0.1 and that have no mature neighbor, preventing background noise. The loss is plain Mean Squared Error between the visible channels at T = 64 steps and the target 28 × 28 × 4 RGBA image, with gradients flowing back through time. Targets were built from MNIST by replicating grayscale intensity into RGB and binarizing alpha (1.0 for the digit, 0.0 for background), so the model must learn both texture and silhouette.

Training used PyTorch on a single NVIDIA GPU, Adam with a learning rate of 1 × 10^-3, batch size 64, 25 epochs, gradient clipping at a norm of 1.0, and a de novo growth strategy — every iteration starts from blank seeds run through all 64 steps, with no replay pool of persistent states.

Why This Matters

Research impact: The work pushes NCAs from texture synthesis and single-target recovery toward structural, class-conditional morphogenesis, showing that strictly local rules can satisfy global topological constraints across multiple classes. It also provides an explicit, class-wide self-repair result that earlier conditional NCA formulations (such as StampCA) did not investigate, and it frames conditional generation as a morphogenesis problem rather than a global generator problem.

Potential real-world applications (not evaluated in the paper, but implied by the approach):

• Fault-tolerant image or asset generation, where damaged or partially lost outputs can regenerate from surviving local context.
• Distributed or on-device generative systems, since the model has only 10,048 parameters.
• Shape and topology design tasks where a target silhouette is constrained but local decoration is flexible.
• Educational or research testbeds for studying self-organization, symmetry breaking, and emergent repair in learned dynamical systems.

Industry relevance: The extreme parameter efficiency suggests a route to generative models that run under tight memory and compute budgets, and the decentralized update rule is conceptually aligned with systems where no central controller exists. The paper itself frames this as a proof-of-concept rather than a deployed system.

Future Directions

1. Scaling to richer data: The paper explicitly names extending the architecture to RGB datasets such as CIFAR-10 as future work.
2. Fixing the '1' → '8' over-growth failure mode: Addressing the specific structural over-growth artifact where roughly 30% of '1' samples gain lateral features is an open problem, likely requiring better inhibition or longer-range coordination.
3. Raising generative fidelity: The SSIM of 0.4826 reflects a canonical "ideal" digit rather than sample-level diversity, so improving variation while retaining stability remains unresolved.
4. Comparing conditional control mechanisms directly: The paper positions its persistent condition injection against seed-encoded conditioning (StampCA), latent-space hypernetworks (VNCA), dynamic perception kernels (AdaNCA), and goal maps (Goal-Guided NCA), but does not report head-to-head experiments — that comparison is an open question.

Target Audience

This paper is most useful to researchers and graduate students working on Neural Cellular Automata, self-organizing and decentralized generative models, or biologically inspired AI. It also suits practitioners interested in ultra-lightweight conditional generation and robustness through self-repair, and readers looking for a concrete, reproducible entry point (code and pretrained models are available at https://github.com/alisakour/Conditional-NCA-Digits) into conditional morphogenesis on MNIST.

Authors’ abstract

Biological systems exhibit remarkable morphogenetic plasticity, where a single genome can encode various specialized cellular structures triggered by local chemical signals. In the domain of Deep Learning, Differentiable Neural Cellular Automata (NCA) have emerged as a paradigm to mimic this self-organization. However, existing NCA research has predominantly focused on continuous texture synthesis or single-target object recovery, leaving the challenge of class-conditional structural generation largely unexplored. In this work, we propose a novel Conditional Neural Cellular Automata (c-NCA) architecture capable of growing distinct topological structures - specifically MNIST digits - from a single generic seed, guided solely by a spatially broadcasted class vector. Unlike traditional generative models (e.g., GANs, VAEs) that rely on global reception fields, our model enforces strict locality and translation equivariance. We demonstrate that by injecting a one-hot condition into the cellular perception field, a single set of local rules can learn to break symmetry and self-assemble into ten distinct geometric attractors. Experimental results show that our c-NCA achieves stable convergence, correctly forming digit topologies from a single pixel, and exhibits robustness characteristic of biological systems. This work bridges the gap between texture-based NCAs and structural pattern formation, offering a lightweight, biologically plausible alternative for conditional generation.