mass-conservation-nca-reservoir-criticality

name	mass-conservation-nca-reservoir-criticality
description	Mass conservation as inductive bias for self-organized criticality in neural cellular automata reservoirs. Demonstrates 1.27× faster evolution with comparable downstream performance. Activation: self-organized criticality, neural cellular automata, reservoir computing, mass conservation, criticality
tags	["neuroscience","reservoir computing","neural cellular automata","self-organized criticality","inductive bias"]

Overview

Investigates whether mass conservation—a local redistribution rule that preserves total lattice mass—serves as an inductive bias toward self-organized criticality (SOC) in evolved neural cellular automata (NCA) reservoirs.

Core Methodology

Neural Cellular Automata Reservoirs

• Evolve NCA toward critical avalanche dynamics
• Use as reservoirs for memory and classification tasks
• Compare mass-conserving vs standard NCA variants

Mass Conservation Mechanism

• Local redistribution rule preserving total lattice mass
• Acts as inductive bias toward SOC
• Promotes robust criticality without sacrificing performance

Key Findings

Criticality Enhancement

• Mass-conserving NCA consistently exhibit stronger criticality
• More runs achieve perfect power-law fits across avalanche distributions
• 1.27× faster during evolution

Downstream Performance

• Both variants achieve comparable performance across three benchmarks:
  • 5-bit sequential memory
  • MNIST digit classification
  • CartPole-v1 temporal control
• Conservation does not impair downstream utility
• Reservoir with perfect criticality achieves highest temporal control score

Applications

• Reservoir computing for memory tasks
• Temporal control systems
• Classification tasks requiring critical dynamics
• Neuromorphic computing with biological plausibility

Implementation Notes

• Mass conservation is simple to implement
• Effective mechanism for promoting robust criticality
• No trade-off between criticality quality and task performance

Pitfalls

• Criticality alone does not guarantee optimal performance
• Need to balance criticality with task-specific requirements
• Evolution process still requires careful fitness function design

References

• arXiv:2606.23115 (June 2026)
• Authors: Tong Zhang, Etienne Guichard, Sidney Pontes-Filho, Stefano Nichele
• 8 pages, 5 figures