Growing Neural Cellular Automata

3 days ago

Morphogenesis is a self-organizing process where cells communicate to form complex, robust biological structures, and some species can regenerate damaged parts.
The central puzzle is how cell collectives know what to build and when to stop; genomics and stem cell biology are insufficient to explain the algorithm for regeneration.
Differentiable cellular automata (CA) models, using continuous cell states and gradient-based optimization, can simulate morphogenesis and learn to grow predefined patterns from a single seed cell.
Key components include perception via Sobel filters for gradients, a learned update rule using neural networks, stochastic updates to avoid global synchronization, and masking to enforce living cells.
Training strategies like sample pools help stabilize patterns as attractors, enabling regenerative capabilities without explicit damage training, and damage exposure during training improves robustness.
Model extensions show rotation invariance and potential for physical implementations as decentralized systems, with applications in biomedicine, bioengineering, and reliable self-organizing agents.

Hasty Briefsbeta