A New Kind of Network? Review and Reference Implementation of Neural Cellular Automata

Researchers Martin Spitznagel and Janis Keuper have published a comprehensive review of Neural Cellular Automata (NCA) on arXiv, alongside a reference implementation called NCAtorch. The paper revisits Stephen Wolfram's 2003 thesis that cellular automata (CA) could replace differential equations for modeling complex systems. Two decades later, NCA merge Wolfram's ideas with learnable artificial neural networks, enabling models to learn update rules directly from data rather than being hand-crafted.

NCAtorch offers a modular PyTorch framework and unified notation for building and training NCA. This allows researchers to experiment with self-organizing generative systems that can model phenomena like pattern formation, morphogenesis, and distributed computing. The paper aims to lower the barrier to entry for NCA research, providing tools to explore how these networks can tackle problems in computer vision and pattern recognition where traditional mathematical formalizations fall short.