MetaNCA teaches neural networks to self-organize without backpropagation

Researchers in July 2026 published a paper on arXiv about MetaNCA — a Meta Neural Cellular Automata framework that trains local rules to self-organize the weights of artificial neural networks. After meta-training, the system generates weights for MLP, CNN, and ResNet without a single step of backpropagation and generalizes to architectures it has never seen during training.

Where the idea came from

Self-organization is a fundamental property of living systems. Biological neurons exchange signals through synapses and throughout an organism's lifetime adapt connections based solely on local information. There is no global error signal — only local interactions, from which intelligence emerges.

Neural Cellular Automata (NCA) have already demonstrated their ability to model morphogenesis through simple local update rules: they are robust to multiple iterations and random perturbations. The authors of MetaNCA took this paradigm and applied it not to the development of biological form, but to the generation of weights in artificial neural networks.

How MetaNCA works

The framework consists of two components. The first is a rule network, which iteratively updates the parameters of the second component — the task network. Interaction occurs only locally: each weight is updated based on information from neighboring weights and hidden states, without access to the global network structure.

To implement this idea, the authors developed a new architecture called Weight Transformer. It applies linear attention to aggregate signals from neighbors in the computational graph. After completing meta-training, the rule network is capable of generating weights for completely new architectures — without additional gradient descent.

Key facts:

• Rule network architecture: Weight Transformer with linear attention
• Test architectures: MLP, CNN, ResNet
• Datasets: MNIST and CIFAR-100
• Maximum scale: networks up to 2 million parameters
• Generalization: MetaNCA works on architectures not included in the meta-training set

What can the system do after training?

After completing meta-training, the rule network does not require backpropagation to generate weights for new networks. This is a fundamental difference from the standard approach: typically, each new architecture requires a full training cycle with gradient descent. MetaNCA replaces this process with iterative application of local rules.

"Self-organization is a property of life, produced by the collective

behavior of individual components acting on the basis of local information," — the authors formulate the key motivation.

Particularly noteworthy is the result on generalization: MetaNCA, trained on a set of different architectures, successfully generates weights for configurations that were not in the training set. The authors also show that diversity of architectures in the meta-training phase directly enhances this transfer capability.

What this means

MetaNCA offers an alternative path to training neural networks — through local self-organization instead of global gradients. While the system has been tested on relatively small networks and standard datasets, the principle of architectural generalization without retraining is potentially an important step toward more flexible and adaptive neural networks.

*Meta is recognized as an extremist organization and is banned in the RF.