Learning Developmental Scaffoldings to Guide Self-Organisation

Codebase for Learning Developmental Scaffoldings to Guide Self-Organisation, which explores the memory-compute trade-off in developmental systems through pre-patterns and neural cellular automata.

Models live in src/, training entrypoints in scripts/, experiment runners in bin/, and figure notebooks in notebooks/.

Setup

Requires Python ≥ 3.13 and uv.

uv sync

JAX is installed with CUDA on Linux and CPU/Metal on macOS automatically.

Running experiments

The bin/ scripts wrap the training entrypoints with the hyperparameter sweeps used in the paper. Edit CUDA_VISIBLE_DEVICES at the top of each script before running.

bash bin/run_siren.sh            # SIREN baseline sweep
bash bin/run_ncas.sh             # standard NCA sweep
bash bin/run_pmcas.sh            # SIREN + (invariant) NCA sweep
bash bin/run_growing.sh          # growing PMCA runs
bash bin/run_compression.sh      # capacity / multi-emoji sweep
bash bin/run_noise_robustness.sh # update-probability robustness sweep

Training scripts

To launch a single run directly:

uv run python -m scripts.train_siren --dataset_name emojis --siren_width 32 --siren_depth 4 ...
uv run python -m scripts.train_nca   --dataset_name emojis --hidden_state 8 ...
uv run python -m scripts.train_pmnca --dataset_name emojis --siren_depth 2 --hidden_state 12 ...

Pass --help to any script for the full argument list. Logs and checkpoints are written under --save_folder.

Figures

uv run jupyter lab notebooks/figures.ipynb