Stability and Geometry of Attractors in Neural Cellular Automata
New analysis shows NCAs exhibit complex oscillatory behavior, challenging a core assumption in the field.
A new paper by researchers Mia-Katrin Kvalsund and James Stovold applies rigorous mathematical analysis from dynamical systems theory to Neural Cellular Automata (NCAs), a class of AI models that evolve patterns over a grid. Using the famous 'growing gecko' NCA from Mordvintsev et al. (2020) as a case study, they visualized the system's attractor dynamics for the first time and analyzed them using Lyapunov and Fourier spectra. Their findings challenge a fundamental assumption in the field: instead of converging to simple, static 'fixed point' attractors, the NCA exhibited complex oscillatory, periodic, and quasi-periodic behaviors that emerge early in training.
This work introduces a new analytical toolkit for NCA researchers, moving beyond visual inspection to quantify system stability and robustness. The study also demonstrated that large perturbations can push the NCA into a completely different behavioral mode, highlighting potential fragility. By proving NCAs are more dynamic than previously thought, this research provides a foundation for building more predictable and reliable self-organizing AI systems, which are used in graphics, materials design, and biological modeling.
- Analysis of the 'growing gecko' NCA revealed complex oscillatory and periodic behaviors, not simple fixed-point attractors.
- The study used Lyapunov and Fourier spectra from dynamical systems theory to quantify NCA stability for the first time.
- Large perturbations were shown to throw the system into a secondary behavioral mode, revealing potential instability.
Why It Matters
Provides a mathematical foundation for analyzing and building more robust, predictable self-organizing AI systems.