Mass conservation boosts NCA reservoir computing by 1.27x
A simple conservation rule makes neural cellular automata both faster and more critical.
Self-organized criticality (SOC) is a dynamical state where systems operate at the edge of chaos, maximizing computational capacity for tasks like reservoir computing. Neural cellular automata (NCA) can be evolved to exhibit SOC, but achieving robust criticality often requires careful tuning. In a new preprint on arXiv, Tong Zhang and colleagues propose using mass conservation—a simple local redistribution rule that preserves total lattice activity—as an inductive bias to automatically push NCA toward critical dynamics.
Across multiple runs, mass-conserving NCA consistently achieved better power-law scaling of avalanche sizes and durations, a signature of SOC. They also evolved 1.27x faster than standard NCA. Downstream performance on sequential memory (5-bit), digit classification (MNIST), and temporal control (CartPole) was comparable or better—the most critical reservoir scored highest on CartPole. This shows that conservation can strengthen criticality without trade-offs, offering a practical, plug-and-play improvement for NCA-based reservoir computing systems.
- Mass conservation biases NCA toward self-organized criticality with perfect power-law fits.
- Evolved 1.27x faster than standard NCA across multiple independent runs.
- Achieved comparable or better performance on 5-bit memory, MNIST, and CartPole tasks.
Why It Matters
A simple conservation rule makes neural cellular automata reservoirs faster, more robust, and more computationally powerful for sequential tasks.