How much can Growing NCAs do?
Growing Neural Cellular Automata was a Distill Paper the came out of nowhere. On their website, the Google Researchers showed that you could use a Neural Network to learn a function for a cellular automaton that enabled morphogenesis. I.e. they could learn a function tha only used local data, but eventually produced complex images on a grid.
The animations are beautiful and the work is clear and impressive.
But the authors learn a single network for each new image. So I wondered: As this is essentially a Generative Model, can we tell the CA what kind of image to produce? The answer is yes - kind of.
Instead of learning a network for a single image, I tried to get the network to produce one of several flower icons - the specific one depended on the input first pixel chosen.
-> Explanation Image TODO
This was reasonably successfull. After training, I could get the CA to produce different flowers, solely depending on the input chosen. This means that such CAs can not only produce morphogenesis, but also have enough computational power to specify subtypes of images that we want to produce.
One fun experiment was “crossing” different flowers. By adding the input codes for the flowers together we could mutate them. THe resulting flowers were not as pretty, but depending on the flowers added, we could see some “mutations” happen.
This complete image was generated using a single cellular automata!