2/n To study this question we paired an NCA with a learned spatial pre-pattern generator in the form of an implicit representation model, specifically a SIREN.
The pre-pattern acts as a scaffolding which sets initial conditions of the NCA but cannot reproduce the target directly
5/n Using a goal-directed NCA formulation, we show that for similar parameter counts, using pre-patterns leads to systems that can encode more patterns:
3/n Our results show 3 things:
1. Pre-patterns improve robustness,
2. they increase model capacity,
3. and, perhaps surprisingly, are not used to directly approximate the target (or some approximation of it)