Epiplexity of Neural Cellular Automata

Interactive demo for exploring Epiplexity, an observer-dependent notion of structural complexity. More powerful observers can find structure where weaker ones see only noise. Generate random Neural Cellular Automata, estimate their epiplexity by training a predictor (transformer or convolutional) to forecast future states, and use a genetic algorithm to find the most complex patterns detectable by that observer. Or directly train a continuous NCA to maximize a differentiable reservoir estimate. [GitHub]

0.05

Neural Cellular Automaton (NCA)

1
4

Simulation

10

Epiplexity Estimation

Epiplexity: --
Step: 0 / 1500 Loss: --

Data

50
3

Observer

3
24

Training

1500
100
0.8
10%

Epiplexity captures the amount and complexity of structure, relative to an observer. Low epiplexity with low final loss means a pattern is trivial; low epiplexity with high final loss means no learnable patterns were found. High epiplexity indicates the presence of complex but learnable structure. The Epiplexity Zoo below lets you explore the world from the selected observer's perspective—the highest-scoring NCAs contain the most intricate patterns that model can detect.

Epiplexity Zoo

50%
Measure an NCA's epiplexity and save it to start your collection