Reproduce “Generative Adversarial 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
Maximilian Otte, Quentin Delfosse, Johannes Czech, Kristian Kersting
Researchers at TU Darmstadt introduced Generative Adversarial Neural Cellular Automata (GANCA), a model combining Neural Cellular Automata with adversarial training to enable a single NCA to generate multiple distinct outputs from varying inputs. GANCA demonstrated a drastic increase in performance on out-of-distribution data, effectively handling novel inputs with a generator utilizing only approximately 20,000 parameters.
Key insights
- Neural Cellular Automata inherently possess a degree of generalization from their local, iterative update mechanism, even when overfitted to training data.
- Integrating NCAs with adversarial training drastically enhances their generalization capabilities, particularly for out-of-distribution inputs.
- Stable adversarial training for NCAs benefits significantly from established GAN techniques like noise injection, label smoothing, and Wasserstein loss.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.