Start with a neighbourhood
A cellular automaton is a collection of cells that repeatedly update their state from nearby information. In a classical automaton, someone writes the update rule by hand. In a neural cellular automaton, a neural network supplies part or all of that local update.
The interesting idea is not that every cell has a separate large brain. Many cells can reuse the same small network while carrying different internal states. A complex global behaviour can emerge from repeated local interactions.
The anatomy of a cell
A typical neural cellular automaton gives each cell a vector of numbers. Some channels may correspond to visible colour or activity. Others are hidden variables that help the cells coordinate. A perception operation gathers local state; a shared neural rule computes a proposed change; the cell updates and the process repeats.
The geometry matters. A two-dimensional grid is natural for images and growth. A ring can support explicit communication experiments. A causal line can model a byte sequence without allowing information to arrive from the future.
What makes the rule neural?
The update contains learned parameters. During training, an objective measures how the resulting dynamics perform, and gradients can adjust the shared rule across an unrolled sequence of updates. After training, the same rule can run repeatedly without changing its parameters.
This distinction matters: a changing picture or a moving state does not prove that a model is training while it runs. State evolution, adaptive memory and durable parameter learning are different mechanisms.
Growth and regeneration are one research direction
The Growing Neural Cellular Automata work by Mordvintsev and colleagues demonstrates learning local rules for pattern growth and explores robustness through techniques such as state pools and damage during training. It is a useful introduction to the field, not a result produced by this F2i project.
Our current public checkpoints investigate associative memory and causal byte generation. They are not image-growth or self-repair checkpoints, and the ring illustration on this website is a conceptual diagram rather than a trained organism.
Our experiments ask different questions
Cellular memory asks whether locally observed associations can be stored in private adaptive state and recovered through a neighbour-to-neighbour query. Causal language asks whether a shared local neural update can predict bytes without attention.
These are two separate models. Success on a small memory task does not make the language model an online knowledge learner, and fluent completion of a short template does not establish general language understanding.
See the boundary between a rule and a learned rule
The Game of Life control runs a fixed, non-neural local rule through ZIPP. Compare it with the trained memory checkpoint. Both have cells and local dynamics; only the latter uses learned key and value representations.