Neural Cellular Automata

Neural Cellular Automata (NCA) are differentiable, self-organizing systems: they use local neural networks to learn decentralized rules for tasks like image regeneration and computational morphogenesis.

Neural Cellular Automata merge classical cellular automata with deep learning: they replace fixed, local rules with a trainable neural network (often a small Recurrent Convolutional Neural Network). Each cell updates its state based only on its immediate 8 neighbors, yet these local interactions collectively achieve complex global goals. For example, the 'Growing NCA' model (Mordvintsev et al., 2020) demonstrates robust self-repair: a partially destroyed image can regenerate itself pixel-by-pixel without centralized control. This architecture is a powerful framework for bio-inspired AI, distributed robotics, and modeling biological self-organization.

https://neuralca.org