Changing static properties
Below we show the evolution of the lizard and bug Persistent NCA used in the static properties experiment
Video 2: left: a persistent lizard, right: a persistent bug.
Below, the left panel shows the evolution of the organism,
while the right panel shows where the cells are located, orange being the original cells and blue the adversaries.
Video 3: red lizard experiment.
Video 4: no tail lizard experiment.
Video 5: bug to butterfly experiment.
We found that adversaries trained to overtake the entire organism from a small 2x2 seed, oftentimes are robust to other
cell injection patterns, in the following we show examples of this behavior.
Cutting the tail of the lizard can also be done when starting from a spray pattern, where
1% of the original cells are randomly replaced by adversaries, or even starting from a single cell.
Video 6: the adversaries start from a spray pattern in which 1% of the original cells are substituted.
Video 7: the adversaries start from a single cell.
We can turn a bug into a butterfly even if we start from spray patterns of different densities, like, 1% and 10%.
Video 8: the adversaries start from a spray pattern with 1% of cells substituted.
Video 9: the adversaries start from a spray pattern with 10% of cells substituted.
While we said that the red lizard was the simplest one to train, it is also true that it is the adversary that is the least
robust of the three, we think this is due to the training, since it took so few number of
gradient updates to train, the adversaries didn't had the time to see a lot of different patterns.
Changing dynamic properties
Below we show the evolution of the three Growing NCA used in the dynamic properties experiment,
as you can see the NCA quickly decay after reaching the target state.
Video 10: left: a growing lizard, center: a growing butterfly, right: a growing firework.
Below, the left panel shows the evolution of the organism,
while the right panel shows where the cells are located, orange being the original cells and blue the adversaries.
Video 11: lizard growing to persistent, the adversaries are injected in a 3x3 square seed.
Video 12: butterfly growing to persistent, 3% of the cells are substituted in a spray pattern.
Video 13: firework growing to persistent, 50% of cells are substituted in a spray pattern
Also dynamic properties can show robustness to other injection patterns.
First, we have the lizard, which is able to turn into a persistent NCA even when we start from a spray pattern.
Video 14: lizard growing to persistent, 1% of the cells are substituted in a spray pattern.
The butterfly, even though it was trained with 3% of cells substituted, works perfectly fine even with 1% of cells substituted.
Video 15: butterfly growing to persistent, 1% of the cells are substituted in a spray pattern.
These extra experiments, show how much the adversaries can adapt to out of distribution patterns,
this property is essential in biological situations, where changing the behaviour of individual cells cannot be done with perfect accuracy.
Having adversaries that work even for misplaced or random locations solves this problem at the root.