Reproduce “Variational Neural Cellular Automata” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rasmus Berg Palm, Miguel González-Duque, Shyam Sudhakaran, Sebastian Risi
Researchers from the IT University of Copenhagen developed Variational Neural Cellular Automata (VNCA), a probabilistic generative model integrating Neural Cellular Automata (NCA) within a Variational Autoencoder framework. The VNCA demonstrates self-organizing generation of diverse outputs like MNIST digits and CelebA-like images, while a specialized variant exhibits robust recovery from significant damage through continued self-organization.
Key insights
- Self-organizing systems based on local communication and iterative updates can effectively learn to generate diverse data, acting as a viable probabilistic generative model.
- The NCA decoder can encourage a more structured and disentangled latent representation, as evidenced by clearer clustering of digit types in the MNIST latent space compared to a standard VAE.
- While the doubling VNCA showed some inherent resilience, a non-doubling VNCA, when specifically trained for it, demonstrated a strong capacity for robust self-repair and recovery from substantial damage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.