Continuous NCA · 2D
Exported experiment
Download JSONTrainable NCA
One neighborhood convolution, then pointwise layers. A residual update keeps two state channels normalized at each cell.
Color shows a cell’s continuous state. Training updates the rule; Play explores its evolution on a larger 48 × 48 grid.
Optimize Epiplexity
● Training batches● Independent evaluation grids
Evaluation: —
Completed steps: 0
Prediction task
Frozen CNN observer
Gradient ascent
Ready
The observer predicts each cell’s next several states from its current neighborhood. Gradient ascent changes the NCA rule to increase the fitted readout’s spectral description length. Each batch uses two fresh grids, 32 burn-in steps, and a small perturbation. Evaluation uses independent initial grids with the same frozen observer.
Experimental 2D adaptation of Zhang & Levin’s learnable novelty estimator. This score depends on the observer and settings; it is separate from the learning-curve estimates in Measure & Evolve. Changing prediction settings starts a new experiment.