The automata just completely destroys the image if I draw too much over the stabilized image with the brush. 5 horizontal swipes are enough to destroy the kitty, is that to be expected?
The NeuralCA both generates and maintains the pattern. Because the NCA was not exposed to damage or erasure during training, its regeneration capability is a purely emergent phenomenon. However, this ability remains somewhat brittle, particularly when the central regions of the pattern are erased.
With the old model (and I suspect this one too) it's trained to generate from a single 'seed' pixel in the center of the image. If you erase the center of the image, that's when it completely collapses.
It must be more general than that, otherwise the cells wouldn’t be able to repair their area if the damage came from the wrong direction (repair is not center-out).
The model generally learns to generate each pixel from its surroundings, even if the surroundings are partially missing.
There's hidden state in the model which presumably it uses to communicate position, ie there's the 3 colors but then a bunch of other channels that the model can use how it wants.
Have you actually tried that? If you specifically erase the center, the image does change a lot at first, but rebuilds itself eventually (albeit to a slightly different final state). It's uncanny how "biological" is feels!
I have yes.. You need to erase a larger amount of the center, but it almost always results in a collapse wheras erasing around the center typically regrows.
At a glance it looks like it could be just iterative texture sampling.
The difference is when creating each pixel, there’s no coordinate to look up, instead it’s using only a set of rules like Conway’s game of life.
But the rules come from a neural network trained on the image, so… it’s kind of memorizing enough information to effectively do the same thing as texture sampling, but using only local information.
I’m sure I’m missing something about how it works or what makes it interesting…
To me, it is intriguing as a toy model for how cells are able to grow into complex tissue and organisms based only on local information, and how they are able to repair and recover harmed tissue.
Of course, this is as close to cells, as neurons from neural networks are to real neurons. And I have no idea what it could be applied to (inpainting/outpainting?), but it’s interesting as exploratory research.
If you're familiar with CAs (e.g. Conway's Game of Life), you can think of a NeuralCA as a CA where the update rule is given by a neural network. Here we optimize the neural net weights so that it behaves a certain way (e.g. grow a lizard from a single seed).
The input to the NN is just the 3x3 neighborhood around a cell.
We can overlap two NNs on the same grid (through interpolation). Checkout https://meshnca.github.io
to see the effect. When the brush is in graft mode, it basically allows you to paint some regions that will follow a different NN rule.
The abstract implies that strictly local updates are a hinderance to high res, however i would have thought there would be an interesting way to get speed up gains from neighbor-only traffic on GPUs CAM-style. am i making that up?
I think performance is not the only issue for scaling to larger grids. CUDA Convolution implementation already utilizes coalescing to improve performance.
The main bottleneck is that in larger grids, cells are further apart, and it takes more steps for them to be able to communicate.
Really interesting demo, nicely done :) Would be fun if switching the "Target Image" when using the second brush mode in the Growing Demo didn't erase/reset the existing canvas, so we could "stamp" new things on top of other images. Small thing perhaps but I got sad when it disappeared when I wanted to merge a kitten on top of the chameleon but couldn't :(
IIRC training starts with the initial state and the end state, and the end state is always oriented the same way. It would be interesting to see what would happen if the end state was rotated randomly though I suspect it wouldn't work so well.
EDIT: video here: https://imgur.com/a/ItZGd5X
The model generally learns to generate each pixel from its surroundings, even if the surroundings are partially missing.
The difference is when creating each pixel, there’s no coordinate to look up, instead it’s using only a set of rules like Conway’s game of life.
But the rules come from a neural network trained on the image, so… it’s kind of memorizing enough information to effectively do the same thing as texture sampling, but using only local information.
I’m sure I’m missing something about how it works or what makes it interesting…
Of course, this is as close to cells, as neurons from neural networks are to real neurons. And I have no idea what it could be applied to (inpainting/outpainting?), but it’s interesting as exploratory research.
Well that sounds like black magic. Nice. Thanks for the reply.